{ "version": 2, "title": "结构化数据论文库", "updatedAt": "2026-10-19", "description": "围绕表格基础模型与结构化决策，按研究领域、子方向和具体议题整理模型、方法及应用证据。", "categories": [ { "id": 1, "name": "结构化数据模型的业界进展", "fullName": "结构化数据模型的业界进展", "description": "跟踪表格基础模型与结构化决策模型的能力、产品形态和应用证据。通过技术报告、独立评测和行业案例，比较模型路线、数据条件与实际使用成本。", "questions": [ "模型提供哪些结构化数据能力？", "评测设置是否接近真实数据？", "行业应用中的收益来自哪里？" ], "children": [ { "id": "industry-models", "name": "模型与产品动态", "description": "整理已有模型的能力范围、输入输出和推理方式，比较单表、关系表等路线，以及开源、微调和部署形态。", "children": [ { "id": "tabular-releases", "name": "通用表格模型", "description": "收录单表预测及通用结构化数据模型的技术报告，关注表征方式、上下文规模与能力扩展。", "code": "01.01.01" }, { "id": "relational-releases", "name": "关系数据模型", "description": "关注多张关联表上的预测能力，包括表间连接、时间约束和任务信息的组织方式。", "code": "01.01.02" } ], "code": "01.01" }, { "id": "industry-applications", "name": "结构化数据应用", "description": "按业务场景观察模型的实际表现，记录数据规模、切分方式、对比方法，以及预测或决策环节中的适用条件。", "children": [ { "id": "credit-risk", "name": "金融风控与信用预测", "description": "研究违约概率、损失率和不平衡数据，比较上下文采样、模型选择与传统方法。", "code": "01.02.01" }, { "id": "engineering-applications", "name": "工业与工程预测", "description": "关注道路、汽车等真实业务数据中的预测任务，观察缺失、少样本和弱语义字段的影响。", "code": "01.02.02" }, { "id": "decision-evaluation", "name": "文本判断与结构化决策", "description": "比较将文本转成标签、量表和固定判断的实际效果，关注准确率、概率校准及成本。", "code": "01.02.03" }, { "id": "biomedical-applications", "code": "01.02.04", "name": "生命科学与医疗预测", "description": "将基因表达、临床字段与删失结果转为结构化预测，关注领域模型与通用表格模型的适用差异。" } ], "code": "01.02" }, { "id": "industry-benchmarks", "name": "技术路线与生态", "description": "通过综述和评测理解不同路线的比较基础，关注数据覆盖、实验协议、计算预算与结论的可迁移性。", "children": [ { "id": "tabular-surveys", "name": "路线综述", "description": "梳理表格基础模型的方法谱系、共同模块与发展方向，为跨模型比较提供背景。", "code": "01.03.01" }, { "id": "general-benchmarks", "name": "通用评测与分布泛化", "description": "关注随机、时间和分组切分，以及大表、高维和预算变化下的模型比较。", "code": "01.03.02" }, { "id": "semantic-relational-benchmarks", "name": "语义与关系数据评测", "description": "评估字符串、表格嵌入和关系数据库任务，比较不同粒度的表示与预测能力。", "code": "01.03.03" } ], "code": "01.03" } ], "code": "01", "shortName": "业界与评测" }, { "id": 2, "name": "表格预测与决策优化", "fullName": "表格预测与决策优化", "description": "研究真实表格任务中的预测质量，以及如何把模型输出用于判断和行动。覆盖缺失与异常、概率和分布预测、置信度校准，以及决策反馈带来的数据变化。", "questions": [ "复杂数据条件下能否稳定预测？", "概率与预测区间是否可信？", "预测结果如何转成具体决策？" ], "children": [ { "id": "prediction-quality", "name": "复杂数据下的预测优化", "description": "针对数据缺失、噪声及异常，研究清洗、补全和检测方法，以及它们与表格预测器的配合方式。", "children": [ { "id": "missing-noisy-data", "name": "缺失补全与噪声清洗", "description": "处理缺失模式和低质量记录，考察补全或清洗是否改善下游预测。", "code": "02.01.01" }, { "id": "anomaly-detection", "name": "表格异常检测", "description": "利用已有表格模型识别异常样本，关注正常样本建模、近邻选择和异常分数。", "code": "02.01.02" } ], "code": "02.01" }, { "id": "prediction-distributions", "name": "预测概率与分布优化", "description": "从单个预测值进一步考察概率、条件分布和区间质量，区分排序准确性、概率校准与不确定性表达。", "children": [ { "id": "probability-calibration", "name": "概率校准与预测一致性", "description": "研究标签比例变化、样本加权、决策阈值、集成及输入变换下的概率偏差和预测一致性。", "code": "02.02.01" }, { "id": "predictive-distributions", "name": "条件分布与不确定性分解", "description": "关注回归分布、概率评分规则，以及模型不确定性和数据噪声的区分。", "code": "02.02.02" }, { "id": "prediction-intervals", "name": "预测区间与覆盖率", "description": "通过分位数和保形预测等方法构造区间，检验不同数据条件下的覆盖率。", "code": "02.02.03" } ], "code": "02.02" }, { "id": "prediction-decisions", "name": "结构化输出与决策优化", "description": "研究评分、选项、概率与显式公式如何支持具体判断，同时考察被预测对象对决策规则的响应。", "children": [ { "id": "scores-routing", "name": "评分、固定选项与路由", "description": "关注受约束的评分或判断任务，以及置信度能否支持复核和模型路由。", "code": "02.03.01" }, { "id": "strategic-feedback", "name": "策略响应与反馈分布", "description": "研究个体根据决策规则改变特征后，模型的预测质量与适配方法。", "code": "02.03.02" }, { "id": "symbolic-regression", "name": "显式表达式与符号回归", "description": "从数值样本学习可执行的数学表达式，比较公式搜索、常数优化及后训练在拟合精度、复杂度与外推上的取舍。", "code": "02.03.03" } ], "code": "02.03" } ], "code": "02", "shortName": "预测与决策" }, { "id": 3, "name": "结构化数据的因果分析与决策", "fullName": "结构化数据的因果分析与决策", "description": "从结构化观测或实验数据中研究变量关系、干预效果和决策策略。分别跟踪因果图发现、效应与分布估计，以及因果知识和表格基础模型的结合方式。", "questions": [ "变量之间的因果结构是什么？", "改变一个变量会带来什么结果？", "因果信息如何支持模型与策略？" ], "children": [ { "id": "causal-discovery", "name": "表格数据的因果发现", "description": "输入观测或干预样本，识别变量之间的因果连接和方向；关注不同机制、数据条件与结构约束下的图恢复。", "children": [ { "id": "causal-graphs", "name": "因果图结构恢复", "description": "研究从多个数据集学习图恢复能力，以直接推断或摊销推断替代逐表搜索。", "code": "03.01.01" }, { "id": "causal-mechanisms", "name": "异质机制与结构推理", "description": "研究不同因果机制下的结构恢复，以及节点排序、父节点选择和无环约束。", "code": "03.01.02" } ], "code": "03.01" }, { "id": "causal-estimation", "name": "基于结构化数据的因果推断", "description": "在明确因果识别条件的前提下，估计平均或个体处理效应、干预分布与策略价值，覆盖二元和连续处理。", "children": [ { "id": "treatment-effects", "name": "处理效应与干预分布", "description": "关注平均处理效应、异质效应及干预结果分布，以及支持这些估计的识别条件。", "code": "03.02.01" }, { "id": "continuous-iv", "name": "连续处理与工具变量", "description": "研究剂量变化下的响应曲线，以及利用工具变量处理未观测混杂的方法。", "code": "03.02.02" }, { "id": "experimental-regimes", "name": "观测与实验数据融合", "description": "组合观测数据和不同干预实验，研究跨实验条件的信息利用与因果查询。", "code": "03.02.03" }, { "id": "causal-policy", "name": "干预分配与策略学习", "description": "把因果效应估计转为干预分配规则，在预算或规则复杂度等约束下评估策略价值。", "code": "03.02.04" } ], "code": "03.02" }, { "id": "causal-models", "name": "因果与表格模型结合", "description": "研究统一因果模型、已有知识注入和目标任务适配，并通过去偏及误差诊断检查估计的可靠性。", "children": [ { "id": "causal-unified", "name": "统一因果模型与路线", "description": "收录覆盖结构、结果和反事实查询的统一模型，以及因果基础模型的路线综述。", "code": "03.03.01" }, { "id": "causal-knowledge", "name": "因果知识与生成约束", "description": "把已知或部分已知的因果关系用于条件化模型、约束数据生成和减少伪相关。", "code": "03.03.02" }, { "id": "causal-adaptation", "name": "因果适配器与目标微调", "description": "通过冻结模型的适配器、干预任务微调或因果数据增强，迁移已有表格模型能力。", "code": "03.03.03" }, { "id": "causal-reliability", "name": "效应去偏与可靠性", "description": "研究因果估计偏差、区间覆盖和场景敏感性，包含去偏校准与真实条件模拟评测。", "code": "03.03.04" } ], "code": "03.03" } ], "code": "03", "shortName": "因果分析" }, { "id": 4, "name": "语义结构化与表格特征构建", "fullName": "语义结构化与表格特征构建", "description": "研究从文本、文档和业务知识到结构化数据的转换过程。覆盖字段与特征生成、跨表信息整合，以及语义表示与数值、类别特征的联合建模。", "questions": [ "语义信息怎样变成可用字段？", "不同来源的数据怎样对齐？", "文本语义怎样进入表格预测？" ], "children": [ { "id": "semantic-features", "name": "语义结构化与任务驱动特征生成", "description": "将非结构化内容提取为记录、标签或特征，并依据任务反馈优化字段定义和特征程序。", "children": [ { "id": "text-structure", "name": "文本到字段与记录", "description": "从长文本中抽取表格，或定义有明确取值范围的语义字段并逐条填充。", "code": "04.01.01" }, { "id": "knowledge-features", "name": "知识驱动特征构建", "description": "利用业务规范、专家材料和字段元数据，生成有来源、可解释的预测特征。", "code": "04.01.02" }, { "id": "feature-search", "name": "任务反馈与特征搜索", "description": "利用验证集反馈迭代生成、筛选和演化特征程序，检查其对预测的实际贡献。", "code": "04.01.03" } ], "code": "04.01" }, { "id": "semantic-integration", "name": "字段、实体对齐与跨表整合", "description": "对齐不同来源的字段含义与实体记录，处理表示差异和冲突，将多个表组织为一致的数据结构。", "children": [ { "id": "schema-matching", "name": "字段匹配与模式整合", "description": "研究列含义匹配、统一字段体系，以及异构表的关系结构整理。", "code": "04.02.01" }, { "id": "entity-matching", "name": "实体匹配与记录对齐", "description": "判断不同记录是否对应同一实体，并权衡候选检索、模型调用和匹配成本。", "code": "04.02.02" }, { "id": "data-integration", "name": "端到端数据整合", "description": "联合考察字段匹配、值归一、实体对齐与冲突融合后的最终数据质量。", "code": "04.02.03" } ], "code": "04.02" }, { "id": "semantic-fusion", "name": "语义与表格特征融合", "description": "让文本语义与数值、类别及其他模态一起参与预测，比较外部嵌入和表格模型原生融合方式。", "children": [ { "id": "semantic-embeddings", "name": "语义嵌入与特征拼接", "description": "将带任务或上下文信息的文本表示作为特征，接入已有表格预测器。", "code": "04.03.01" }, { "id": "text-tabular-models", "name": "文本与表格联合建模", "description": "在表格模型内部结合列名、文本单元格及统计关系，形成面向任务的表示。", "code": "04.03.02" }, { "id": "multimodal-schema", "name": "跨字段体系与多模态融合", "description": "研究不同字段体系之间的表示迁移，以及表格、文本和图像信息的结合。", "code": "04.03.03" } ], "code": "04.03" } ], "code": "04", "shortName": "语义与特征" }, { "id": 5, "name": "表格模型的上下文学习与推理优化", "fullName": "表格模型的上下文学习与推理优化", "description": "研究已有表格基础模型如何利用上下文样本完成预测。围绕样本选择、推理策略和计算效率，比较有限上下文与有限预算下的性能收益。", "questions": [ "哪些样本值得放入上下文？", "额外推理计算应该用在哪里？", "如何处理大样本和多特征？" ], "children": [ { "id": "context-selection", "name": "上下文样本优化", "description": "在模型可处理的上下文预算内选择、组织和更新样本，研究局部相关性、代表性和数据分布的作用。", "children": [ { "id": "retrieval-shots", "name": "检索与近邻选择", "description": "为查询选择相似样本与示例数量，考察局部上下文、检索空间及微调的配合。", "code": "05.01.01" }, { "id": "representative-context", "name": "代表性选择与上下文优化", "description": "通过原型、样本优化或分布匹配批次，保留有限上下文中的关键信息。", "code": "05.01.02" }, { "id": "stream-context", "name": "流式更新与记忆管理", "description": "在持续到达的数据上更新有限上下文，关注记忆预算与新旧样本的利用。", "code": "05.01.03" } ], "code": "05.01" }, { "id": "inference-strategies", "name": "表格推理策略与机制", "description": "考察推理时处理、集成和流程选择，并通过机制分析理解模型利用特征与上下文的方式。", "children": [ { "id": "test-time-compute", "name": "推理时计算与组合策略", "description": "比较适配、集成和上下文构造等额外计算方式，以及不同策略的组合收益。", "code": "05.02.01" }, { "id": "pipeline-search", "name": "自动管线与策略搜索", "description": "依据任务验证反馈选择或改进预处理、推理和适配管线。", "code": "05.02.02" }, { "id": "inference-mechanisms", "name": "模型机制与行为分析", "description": "研究表格模型的内部表征和预测行为，为推理优化提供机制层面的线索。", "code": "05.02.03" } ], "code": "05.02" }, { "id": "inference-efficiency", "name": "大表推理与部署效率", "description": "减少上下文、注意力和中间状态的开销，研究准确率、延迟、内存与可处理数据规模之间的取舍。", "children": [ { "id": "context-compression", "name": "上下文压缩与信息迁移", "description": "将完整上下文的信息压缩或迁移到短上下文，考察可恢复的预测能力。", "code": "05.03.01" }, { "id": "memory-quantization", "name": "内存优化与注意力量化", "description": "研究注意力计算、中间状态和数值精度的优化，以降低大表推理资源需求。", "code": "05.03.02" } ], "code": "05.03" } ], "code": "05", "shortName": "上下文与推理" }, { "id": 6, "name": "表格基础模型的微调与蒸馏", "fullName": "表格基础模型的微调与蒸馏", "description": "关注已有表格基础模型在目标数据上的进一步优化。比较参数更新、输入适配、分布鲁棒性和知识蒸馏，考察预测收益、过拟合风险及部署成本。", "questions": [ "目标数据适合更新哪些部分？", "分布变化时怎样保持稳定？", "能力能否迁移到轻量预测器？" ], "children": [ { "id": "adaptation-target", "name": "目标表格数据微调与轻量适配", "description": "对已有基础模型进行目标任务适配，比较更新参数与修改输入两类方式，以及样本量和验证策略的影响。", "children": [ { "id": "parameter-finetuning", "name": "参数微调与训练策略", "description": "比较监督式、元学习式和参数高效微调，并检查小样本与不平衡条件下的收益。", "code": "06.01.01" }, { "id": "input-adapters", "name": "输入空间与冻结模型适配", "description": "冻结基础模型，通过输入端的小型修正模块完成适配，并控制无收益时的退化。", "code": "06.01.02" } ], "code": "06.01" }, { "id": "adaptation-distribution", "name": "数据分布适配", "description": "研究目标数据的人群组成或统计分布变化，关注整体表现与弱势子群表现的平衡。", "children": [ { "id": "subpopulation-shift", "name": "子群变化与分布鲁棒性", "description": "通过少量参数更新和鲁棒目标，应对人群比例改变造成的性能下降。", "code": "06.02.01" } ], "code": "06.02" }, { "id": "adaptation-distillation", "name": "表格预测能力蒸馏", "description": "用基础模型生成的预测或软标签训练轻量学生，比较教师组合、查询数据和学生结构。", "children": [ { "id": "distillation-recipes", "name": "教师信号与学生比较", "description": "研究监督信号、合成查询及不同学生模型，比较蒸馏效果和推理成本。", "code": "06.03.01" }, { "id": "tree-distillation", "name": "树模型蒸馏与 CPU 部署", "description": "将基础模型预测迁移到梯度提升树等轻量模型，关注折外监督与部署效率。", "code": "06.03.02" } ], "code": "06.03" } ], "code": "06", "shortName": "微调与蒸馏" } ], "papers": [ { "id": "arxiv-2610.20342", "title": "When Does Task-Aware Imputation Matter? An Analysis of Robustness Under Train-Test Mismatch", "date": "2026-10-14", "updated": "2026-10-14", "category": 2, "section": "prediction-quality", "topicId": "missing-noisy-data", "related": [ 6 ], "status": "preprint", "venue": "预印本", "summary": "把缺失值补全与分类损失联合优化，并区分部署时仍缺失和仅训练集缺失两种情况。在 98 个表格数据集上与 TabImpute 等比较，其中 36 个用于缺失偏移测试；收益集中于困难偏移，同分布总体未显著优于简单补全。下游只测试小型 MLP，尚不能直接推断对 TFM 预测同样有效。", "url": "https://arxiv.org/abs/2610.20342", "pdf": "https://arxiv.org/pdf/2610.20342", "code": null, "keywords": [ "Joint-AmortizedImpute", "缺失补全与噪声清洗", "结构化数据" ], "aliases": [ "Joint-AmortizedImpute", "Joint-LatentImpute", "Task-aware imputation" ], "sources": [ "https://arxiv.org/abs/2610.20342", "https://arxiv.org/html/2610.20342" ], "topic": "缺失补全与噪声清洗", "addedAt": "2026-10-16" }, { "id": "arxiv-2610.19536", "title": "Typed Decision Models on Tabular Data", "date": "2026-10-14", "updated": "2026-10-14", "category": 2, "section": "prediction-decisions", "topicId": "scores-routing", "related": [ 1, 4, 5 ], "status": "preprint", "venue": "预印本", "summary": "比较 Jev 等四种决策模型与三种 LLM，在行分类、表格事实核验和异常检测中测试 21 类保持信息不变的改写。作者发现整体准确率可能掩盖单条决策翻转，置信度也更接近稳定性而非正确性。可用于设计格式测试与人工复核；成本比较受批量大小和 API 快照影响。", "url": "https://arxiv.org/abs/2610.19536", "pdf": "https://arxiv.org/pdf/2610.19536", "code": null, "keywords": [ "Typed Decision Models", "评分、固定选项与路由", "结构化数据" ], "aliases": [ "Typed Decision Models", "Jev tabular decisions" ], "sources": [ "https://arxiv.org/abs/2610.19536", "https://arxiv.org/html/2610.19536" ], "topic": "评分、固定选项与路由", "addedAt": "2026-10-16" }, { "id": "arxiv-2610.20418", "title": "OrionMoE-v1.0: A Mixture-of-Experts Foundation Model for Tabular In-Context Learning", "date": "2026-10-14", "updated": "2026-10-14", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 5 ], "status": "preprint", "venue": "预印本", "summary": "在双轴表格 Transformer 的特征轴引入稀疏专家，由支持集汇总的全表表示参与路由，默认组合八种输入视图。作者在 TabArena 等基准报告了小模型的竞争力，但分组／时间漂移和大数据仍有弱项，尚无等参数稠密模型消融来隔离路由贡献。论文注明采用限制性源码许可，制品下载尚待核实。", "url": "https://arxiv.org/abs/2610.20418", "pdf": "https://arxiv.org/pdf/2610.20418", "code": null, "keywords": [ "OrionMoE", "通用表格模型", "结构化数据" ], "aliases": [ "OrionMoE", "OrionMoE-Small", "OrionMoE v1.0" ], "sources": [ "https://arxiv.org/abs/2610.20418", "https://arxiv.org/html/2610.20418" ], "topic": "通用表格模型", "addedAt": "2026-10-16" }, { "id": "arxiv-2610.20462", "title": "Learning by Examples: Power Grid Steady State Analysis with a Tabular Foundation Model", "date": "2026-10-14", "updated": "2026-10-14", "category": 1, "section": "industry-applications", "topicId": "engineering-applications", "related": [ 2, 4, 5 ], "status": "preprint", "venue": "预印本", "summary": "将电网潮流、故障筛查和电压稳定裕度写成表格任务，以直流潮流残差、功率解耦和拓扑特征辅助 TabPFN。118／2000 节点仿真显示部分任务可与专用网络竞争，但接近失稳的分布外情形仍明显退化，推理也更慢。价值在于用上下文与特征注入物理知识，尚非实时电网部署证据。", "url": "https://arxiv.org/abs/2610.20462", "pdf": "https://arxiv.org/pdf/2610.20462", "code": null, "keywords": [ "TabPFN power grid", "工业与工程预测", "结构化数据" ], "aliases": [ "TabPFN power grid", "TabPFN-R", "TabPFN-FD" ], "sources": [ "https://arxiv.org/abs/2610.20462", "https://arxiv.org/html/2610.20462" ], "topic": "工业与工程预测", "addedAt": "2026-10-16" }, { "id": "arxiv-2610.13724", "title": "Amortized Counterfactual Prediction with Graph Neural Networks", "date": "2026-10-09", "updated": "2026-10-09", "category": 3, "section": "causal-models", "topicId": "causal-unified", "related": [ 5 ], "status": "preprint", "venue": "预印本", "summary": "CFGNP 用观测上下文编码与事实—反事实双图消息传递学习反事实分布，并测试从 10 节点向 30 节点图迁移。合成实验有积极结果，但训练使用基线没有的反事实样本对，优势不能全部归因于架构；观测数据下不可识别的问题仍不可识别。适合作为摊销因果推理的探索路线。", "url": "https://arxiv.org/abs/2610.13724", "pdf": "https://arxiv.org/pdf/2610.13724", "code": "https://github.com/Zelos1/CFGNP", "keywords": [ "CFGNP", "统一因果模型与路线", "结构化数据" ], "aliases": [ "CFGNP", "Counterfactual Graph Neural Process" ], "sources": [ "https://arxiv.org/abs/2610.13724", "https://arxiv.org/html/2610.13724", "https://github.com/Zelos1/CFGNP" ], "topic": "统一因果模型与路线", "addedAt": "2026-10-16" }, { "id": "arxiv-2607.04809", "title": "Context-Constrained Transfer Learning for Tabular Foundation Models via Data Distillation", "date": "2026-07-06", "updated": "2026-07-06", "category": 5, "section": "inference-efficiency", "topicId": "context-compression", "related": [ 2, 6 ], "status": "preprint", "venue": "预印本", "summary": "TL-ANDI 用带后验差异惩罚的最优传输选择源数据锚点，再平滑标签、校准目标残差，在有限上下文内做跨域迁移。TabPFN-2.5 的模拟与房价、糖尿病两个数据集实验提供初步支持；方法使用待预测样本的无标签特征。避免负迁移的理论依赖独立验证集，不能直接覆盖最终全量重拟合。", "url": "https://arxiv.org/abs/2607.04809", "pdf": "https://arxiv.org/pdf/2607.04809", "code": null, "keywords": [ "TL-ANDI", "上下文压缩与信息迁移", "结构化数据" ], "aliases": [ "TL-ANDI" ], "sources": [ "https://arxiv.org/abs/2607.04809", "https://arxiv.org/html/2607.04809" ], "topic": "上下文压缩与信息迁移", "addedAt": "2026-10-16" }, { "id": "arxiv-2607.05103", "title": "Choosing a parallel heterogeneous ensemble method for tabular classification", "date": "2026-07-06", "updated": "2026-07-06", "category": 5, "section": "inference-strategies", "topicId": "test-time-compute", "related": [ 1, 2 ], "status": "preprint", "venue": "预印本", "summary": "在 56 个 OpenML 分类任务上比较投票、Blending、Stacking 等方法，再用不重叠的 28 个 TabArena 任务验证组合规则。作者报告组合优于单一最佳模型，但未显著胜过最强单项集成；另列的 Stacking 试验存在数据泄漏，不能据其排名作公平比较。适合参考集成选择与评测协议。", "url": "https://arxiv.org/abs/2607.05103", "pdf": "https://arxiv.org/pdf/2607.05103", "code": null, "keywords": [ "Parallel heterogeneous ensembles", "推理时计算与组合策略", "结构化数据" ], "aliases": [ "Parallel heterogeneous ensembles", "Robust Soft Voting" ], "sources": [ "https://arxiv.org/abs/2607.05103", "https://arxiv.org/html/2607.05103" ], "topic": "推理时计算与组合策略", "addedAt": "2026-10-16" }, { "id": "arxiv-2610.17867", "title": "RT-J: Large-Scale Pretraining of Relational Transformers for Context-Efficient Predictions", "date": "2026-05-25", "updated": "2026-10-15", "category": 1, "section": "industry-models", "topicId": "relational-releases", "related": [ 4, 5, 6 ], "status": "published", "venue": "NeurIPS 2026", "summary": "RT-J 结合关系随机游走检索、局部邻域扩展与上下文调参／集成，在 RelBench 的 7 个数据库、21 项任务中提高有限上下文的利用率。作者的等标签比较支持回归收益，但完整调参另需最多 4,096 条验证标签，分类调参可能退化；不支持链接预测。连接语义、关系数据、推理和微调路线。", "url": "https://arxiv.org/abs/2610.17867", "pdf": "https://arxiv.org/pdf/2610.17867", "code": "https://github.com/stanford-star/relational-transformer", "keywords": [ "RT-J", "关系数据模型", "结构化数据" ], "aliases": [ "RT-J", "RT-PluRel", "Large-Scale Pretraining unlocks Few-Shot Prediction for Relational Data" ], "sources": [ "https://arxiv.org/abs/2610.17867", "https://arxiv.org/html/2610.17867", "https://star-project.stanford.edu/rt-j/", "https://openreview.net/forum?id=oQINTd9din", "https://openreview.net/forum?id=sMFUopPlwy", "https://neurips.cc/virtual/2026/poster/149169", "https://github.com/stanford-star/relational-transformer" ], "dateNote": "早期同作以 Large-Scale Pretraining unlocks Few-Shot Prediction for Relational Data 公开于 2026-05-25；arXiv 于 2026-10-13 首次提交，2026-10-15 修订。", "topic": "关系数据模型", "addedAt": "2026-10-16" }, { "id": "arxiv-2605.18702", "title": "Distilling Tabular Foundation Models for Structured Health Data", "date": "2026-05-18", "updated": "2026-05-18", "category": 6, "section": "adaptation-distillation", "topicId": "tree-distillation", "related": [ 1, 2, 5 ], "status": "preprint", "venue": "预印本", "summary": "用五折折外教师预测生成软标签，将六类 TFM 蒸馏为树模型或 MLP，在 19 个医疗数据集上检查精度、校准、公平性及成本。收益主要出现在树学生，MLP 可能退化，多分类实现也有缺口。提供避免上下文标签泄漏的部署思路；数据集含同一队列的不同版本，耗时比较采用不同硬件。", "url": "https://arxiv.org/abs/2605.18702", "pdf": "https://arxiv.org/pdf/2605.18702", "code": null, "keywords": [ "Healthcare TFM distillation", "树模型蒸馏与 CPU 部署", "结构化数据" ], "aliases": [ "Healthcare TFM distillation", "OOF healthcare distillation" ], "sources": [ "https://arxiv.org/abs/2605.18702", "https://arxiv.org/html/2605.18702" ], "topic": "树模型蒸馏与 CPU 部署", "addedAt": "2026-10-16" }, { "id": "arxiv-2605.09424", "title": "TabFORGE: A Tabular Foundation Model for Generative Modelling", "date": "2026-05-10", "updated": "2026-10-15", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 3, 6 ], "status": "published", "venue": "NeurIPS 2026", "summary": "冻结 PFN 中间层编码，以潜空间扩散和去噪对齐解码器生成表格；目标数据需短步微调。在来自 TabArena／TabStruct 的 45 个数据集上比较 22 种方法，分开检查生成效用和过拟合。全局效用支持结构保真价值，但不等于因果识别或隐私保证，失败基线的指标填补也影响排名解释。", "url": "https://arxiv.org/abs/2605.09424", "pdf": "https://arxiv.org/pdf/2605.09424", "code": "https://github.com/SilenceX12138/TabFORGE", "keywords": [ "TabFORGE", "通用表格模型", "结构化数据" ], "aliases": [ "TabFORGE", "Tabular Foundation Model for Generative Modelling" ], "sources": [ "https://arxiv.org/abs/2605.09424", "https://arxiv.org/html/2605.09424", "https://neurips.cc/virtual/2026/poster/149211", "https://github.com/SilenceX12138/TabFORGE" ], "topic": "通用表格模型", "addedAt": "2026-10-16" }, { "id": "arxiv-2610.10321", "title": "Estimating Uncoded Crash Factors with Tabular Foundation and System One Models: Kumo Tabular and Jev", "date": "2026-10-07", "updated": "2026-10-07", "category": 1, "section": "industry-applications", "topicId": "engineering-applications", "related": [ 2, 4 ], "topic": "工业与工程预测", "status": "preprint", "venue": "预印本", "summary": "把事故表格预测、叙述文本识别和人工抽样校准结合，估计编码字段未记录的因素。原文限定于得州 2017–2025 年有坐标和有效叙述的约 502 万事故；Kumo 采用五折交叉拟合，Jev 输出经人工样本再校准。叙述中出现的因素不等于事故真实成因；不同输入字段的基线差异也不能全部归因于模型。", "url": "https://arxiv.org/abs/2610.10321", "pdf": "https://arxiv.org/pdf/2610.10321", "code": "https://github.com/pozapas/kumo-jev-crash-records", "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "Kumo Tabular Jev", "JEV crash records" ], "sources": [ "https://arxiv.org/abs/2610.10321" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.10317", "title": "Thinking in Depth: Retrospective Inference for Tabular Foundation Models", "date": "2026-10-07", "updated": "2026-10-07", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "Retro 用注意力残差重读早期表示，并用查询条件门控调节上下文更新。在 TabArena、TALENT 与 RelArena 比较预测和成本，并通过层间读出、门控干预检查机制。这是配套训练的新模型；冻结模型上的干预不能独立证明各组件训练时的因果贡献，排行榜位置也依赖候选配置。", "url": "https://arxiv.org/abs/2610.10317", "pdf": "https://arxiv.org/pdf/2610.10317", "code": null, "keywords": [ "TFM", "表格基础模型", "Retro" ], "aliases": [ "Retro", "Retrospective Inference" ], "sources": [ "https://arxiv.org/abs/2610.10317" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.08694", "title": "GeneICL: A Tabular Foundation Model for Bulk Transcriptomics", "date": "2026-10-06", "updated": "2026-10-06", "category": 1, "section": "industry-applications", "topicId": "biomedical-applications", "related": [ 2, 5 ], "topic": "生命科学与医疗预测", "status": "preprint", "venue": "预印本", "summary": "GeneICL 将转录组先验和共享循环块用于分类、回归，并通过 Cox 残差扩展生存预测。评测含 27 个分类、25 个回归和 28 个生存任务；前两类共用五折切分，传统方法在训练折内调参。多个任务共享底层队列，不能视为 80 个独立人群验证；临床可用性仍需外部队列评估。", "url": "https://arxiv.org/abs/2610.08694", "pdf": "https://arxiv.org/pdf/2610.08694", "code": null, "keywords": [ "TFM", "表格基础模型", "GeneICL" ], "aliases": [ "GeneICL" ], "sources": [ "https://arxiv.org/abs/2610.08694" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.07559", "title": "TAFFY: A Task-Adaptive Tabular Foundation Model with In-Context Diversity", "date": "2026-10-06", "updated": "2026-10-06", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "TAFFY 结合相关环境的上下文多样性、共享块循环和支持集统计门控。在六组分类和五组回归基准比较，并对上下文构造、门控和循环深度做控制实验。分类与回归使用分别训练的模型；作者报告的平均排名取决于指定配置，不能概括为所有数据条件下最优。", "url": "https://arxiv.org/abs/2610.07559", "pdf": "https://arxiv.org/pdf/2610.07559", "code": null, "keywords": [ "TFM", "表格基础模型", "TAFFY" ], "aliases": [ "TAFFY", "Taffy-4L" ], "sources": [ "https://arxiv.org/abs/2610.07559" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.06625", "title": "JEV versus LLMs: Accuracy, Cost and Calibration on Seven Political Science Replications", "date": "2026-10-05", "updated": "2026-10-06", "category": 1, "related": [ 2, 4 ], "topic": "结构化决策模型的独立评测与行业验证", "status": "preprint", "venue": "预印本", "summary": "在七组政治科学复现实验中比较 Jev 与 LLM 的文本标签、量表、成本和校准，部分任务报告配对不确定性区间。作者未发现全面的成本或能力优势，主要便利是速度与概率读取；校准优势也依赖对照模型。适合作为结构化决策模型的任务适配参考。", "url": "https://arxiv.org/abs/2610.06625", "pdf": "https://arxiv.org/pdf/2610.06625", "code": null, "keywords": [ "Jev" ], "section": "industry-applications", "topicId": "decision-evaluation", "addedAt": "2026-10-12" }, { "id": "paper-baf74a758c", "title": "Pavement performance prediction via a tabular foundation model", "date": "2026-09", "updated": "2026-09", "category": 1, "related": [ 2 ], "topic": "结构化数据应用", "status": "published", "venue": "Computer-Aided Civil and Infrastructure Engineering · 2026", "summary": "用 TabPFN 预测下一巡检周期的路面状态，主实验为日本国道 6 号的 189 段路面，并与五类固定参数基线比较。国道 4 号补充验证表现明显下降，少数维修类 F1 为 0；巡检减负仅为情景推算。可参考小样本行业应用设计，尚不足以支持跨路网推广。", "url": "https://doi.org/10.1016/j.cacaie.2026.100097", "pdf": null, "code": null, "keywords": [], "dateNote": "期次月份；在线首发日待核", "section": "industry-applications", "topicId": "engineering-applications", "addedAt": "2026-10-12", "sources": [ "https://doi.org/10.1016/j.cacaie.2026.100097", "https://www.sciencedirect.com/science/article/pii/S1093968726030835" ] }, { "id": "report-nvidia-kumo-tabular", "title": "NVIDIA Kumo Tabular Sets a New Accuracy-Efficiency Frontier for Tabular Prediction", "date": "2026-09-29", "updated": "2026-09-29", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 5 ], "topic": "通用表格模型", "status": "report", "venue": "NVIDIA 官方技术说明", "summary": "Kumo Tabular 结合列注意力、行压缩与上下文 KV 缓存，官方说明披露四组基准及指定单卡评测。原生单次前向最多十类，更多类别依赖编码扩展；缓存收益与分布外表现需按任务验证。可参考其推理路径，性能结论属于官方报告。", "url": "https://huggingface.co/blog/nvidia/kumo-tabular", "pdf": null, "code": "https://github.com/NVIDIA/structured-data-models", "keywords": [ "TFM", "表格基础模型", "Kumo Tabular" ], "aliases": [ "Kumo Tabular", "Kumo-Tabular", "NVIDIA Kumo" ], "sources": [ "https://huggingface.co/blog/nvidia/kumo-tabular", "https://huggingface.co/nvidia/Kumo-Tabular" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.36108", "title": "LoopICL: Looping a single transformer block to solve tabular tasks", "date": "2026-09-28", "updated": "2026-09-28", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "LoopICL 重复使用一个 Transformer 块，结合分布编码和退出门调节循环深度。对中间训练检查点做匹配堆叠基线与循环消融，并在 TabArena、TALENT 评估。共享参数不代表计算量同步下降；长上下文训练会缩小部分堆叠与循环模型差距。", "url": "https://arxiv.org/abs/2609.36108", "pdf": "https://arxiv.org/pdf/2609.36108", "code": null, "keywords": [ "TFM", "表格基础模型", "LoopICL" ], "aliases": [ "LoopICL" ], "sources": [ "https://arxiv.org/abs/2609.36108" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.27679", "title": "What Do Tabular Foundation Models Compute In Context? In-Situ Representation Refinement through Attention-Gated Updates", "date": "2026-09-23", "updated": "2026-09-25", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "RefineICL 以支持集标签修正表示，通过注意力门控传递更新，并使用低秩交互与类型化记忆。按保存切分评估多个基准，另有支持状态和门控干预、匹配规模的 FFN 对照与 bootstrap。机制证据针对该设计及实验条件，不能据此把全部 TFM 的前向计算等同于同一优化算法。", "url": "https://arxiv.org/abs/2609.27679", "pdf": "https://arxiv.org/pdf/2609.27679", "code": null, "keywords": [ "TFM", "表格基础模型", "RefineICL" ], "aliases": [ "RefineICL", "RefineICL-L24" ], "sources": [ "https://arxiv.org/abs/2609.27679" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.22866", "title": "Causilo Technical Report", "date": "2026-09-19", "updated": "2026-09-19", "category": 1, "related": [ 2, 5 ], "topic": "模型与产品动态", "status": "report", "venue": "技术报告", "summary": "Causilo 在行压缩前细化单元格信息，以固定摘要 token 交换特征信息。报告 TabArena、BeyondArena、ScoringBench 的预测与耗时，并测量特征数扩展曲线。线性特征侧注意力复杂度不等于全流程线性；官方排行榜与速度取决于基准快照、配置和硬件。", "url": "https://arxiv.org/abs/2609.22866", "pdf": "https://arxiv.org/pdf/2609.22866", "code": "https://github.com/nums-ai/causilo", "keywords": [ "TFM", "Causilo" ], "section": "industry-models", "topicId": "tabular-releases", "aliases": [], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.17488", "title": "LimiX-2: A Contextual Mechanism Network Towards General Structured-Data Intelligence", "date": "2026-09-15", "updated": "2026-09-15", "category": 1, "related": [ 2, 3, 5 ], "topic": "模型与产品动态", "status": "report", "venue": "技术报告", "summary": "LimiX-2 采用单元格级双轴建模和上下文条件掩码任务，扩展固定目标列预测。TabArena 沿用官方评测，TALENT 与 BCCO 使用 15 个随机种子，并单独评估因果骨架恢复。注意力恢复骨架不等于识别因果方向或干预效果；多个任务表现不能直接推出通用业务决策能力。", "url": "https://arxiv.org/abs/2609.17488", "pdf": "https://arxiv.org/pdf/2609.17488", "code": null, "keywords": [ "LimiX", "TFM" ], "section": "industry-models", "topicId": "tabular-releases", "aliases": [ "LimiX-2", "LimiX2", "LimiX 2" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.17895", "title": "TabPFN-3.5: Technical Report", "date": "2026-09-15", "updated": "2026-09-22", "category": 1, "related": [ 2, 4, 5 ], "topic": "模型与产品动态", "status": "report", "venue": "技术报告", "summary": "TabPFN-3.5 改进单元格编码、联合任务架构与推理配置，扩展分组、字符串及多模态表格。报告 TabArena、BeyondArena、STRABLE、MulTaBench 与关系任务；区分 Fast、Plus、Thinking 的前向和缓存耗时。多个配置与任务封装不能视作一个同成本模型；大于 128k 行的样本侧计算仍为二次扩展。", "url": "https://arxiv.org/abs/2609.17895", "pdf": "https://arxiv.org/pdf/2609.17895", "code": "https://github.com/PriorLabs/TabPFN", "keywords": [ "TabPFN", "TFM" ], "section": "industry-models", "topicId": "tabular-releases", "aliases": [ "TabPFN-3.5", "TabPFN3.5" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.04540", "title": "Mitra-v2 Technical Report", "date": "2026-09-03", "updated": "2026-09-03", "category": 1, "related": [ 5, 6 ], "topic": "模型与产品动态", "status": "report", "venue": "技术报告", "summary": "Mitra-v2 采用二维表格骨干与混合先验，在 TabArena 多次切分和排除开发集的 TALENT 上评估。主要配置含目标微调与八折集成，成本明显不同于纯前向模型；领先组的排名区间重叠。可参考微调、上下文容量与集成的收益取舍。", "url": "https://arxiv.org/abs/2609.04540", "pdf": "https://arxiv.org/pdf/2609.04540", "code": null, "keywords": [ "Mitra", "TFM" ], "section": "industry-models", "topicId": "tabular-releases", "aliases": [ "Mitra-v2", "Mitra v2" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.03880", "title": "Xiaomi-TabLDM: A Tabular Foundation Model Technical Report", "date": "2026-09-03", "updated": "2026-09-04", "category": 1, "related": [ 5 ], "topic": "模型与产品动态", "status": "report", "venue": "技术报告", "summary": "Xiaomi-TabLDM 组合双流特征分组、注意力残差与稀疏专家，并增加推理时计算。在 TALENT、TabArena、BCCO、CTR23 比较分类、回归和效率，并提供敏感性分析。强项偏回归，具体排名取决于集合与配置；稀疏激活参数量不能替代端到端延迟或显存实测。", "url": "https://arxiv.org/abs/2609.03880", "pdf": "https://arxiv.org/pdf/2609.03880", "code": null, "keywords": [ "TFM", "Xiaomi-TabLDM" ], "section": "industry-models", "topicId": "tabular-releases", "aliases": [], "addedAt": "2026-10-12" }, { "id": "arxiv-2608.25774", "title": "EXAONE Tabular 1.0 : Technical Report", "date": "2026-08-26", "updated": "2026-08-26", "category": 1, "related": [ 2, 5 ], "topic": "模型与产品动态", "status": "report", "venue": "技术报告", "summary": "EXAONE Tabular 通过样本与特征摘要交互，保留单元格表示而不提前压成行向量。分别报告 TabArena、BCCO、TALENT、ScoringBench；后两类预测评测将支持集限制为每折最多 5 万行。TALENT 排除超过十类任务；多分类扩展增加子问题成本，查询分块也不保证所有配置都可复用支持缓存。", "url": "https://arxiv.org/abs/2608.25774", "pdf": "https://arxiv.org/pdf/2608.25774", "code": null, "keywords": [ "EXAONE", "TFM", "EXAONE Tabular" ], "section": "industry-models", "topicId": "tabular-releases", "aliases": [], "addedAt": "2026-10-12" }, { "id": "arxiv-2608.01400", "title": "TabDPT-Turbo: Efficient In-Context Learning for Tabular Prediction", "date": "2026-07", "updated": "2026-08-25", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5, 2 ], "topic": "通用表格模型", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "TabDPT-Turbo 用长上下文替代逐查询检索，在相对基准去重的训练数据上学习，并于 TabArena-Lite、CC18、CTR23 比较预测和耗时。论文对应 v1.2，主要价值是共享上下文带来的效率；v1.3 的概率回归功能来自后续发布说明，其效果不能直接套用论文结果。", "url": "https://arxiv.org/abs/2608.01400", "pdf": "https://arxiv.org/pdf/2608.01400", "code": "https://github.com/layer6ai-labs/TabDPT-inference", "keywords": [ "TFM", "表格基础模型", "TabDPT" ], "aliases": [ "TabDPT-Turbo", "TabDPT v1.2", "TabDPT 1.2", "TabDPT v1.3", "TabDPT 1.3", "TabDPT-1.3" ], "sources": [ "https://arxiv.org/abs/2608.01400", "https://icml-structured-fm-workshop.github.io/accepted-papers/", "https://github.com/layer6ai-labs/TabDPT-inference", "https://github.com/layer6ai-labs/TabDPT-inference/releases/tag/v1.3.0" ], "dateNote": "2026-07 Workshop 版本；arXiv 首次提交 2026-08-02。", "addedAt": "2026-10-12" }, { "id": "arxiv-2609.37959", "title": "TabFM: A Zero-Shot Foundation Model for Tabular Data", "date": "2026-06-30", "updated": "2026-09-29", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 5 ], "topic": "通用表格模型", "status": "report", "venue": "Google Research 技术报告", "summary": "Google TabFM 以单元格编码、行列交互和上下文预测处理新表格，并提供增强推理方案。报告 51 个 TabArena 数据集，并分别介绍原始模型、TabFM+ 特征处理与集成、TabFM-Auto 的 LLM 搜索。不同方案使用的额外计算与外部 LLM 不同；预训练规模上限后的大表推理依赖长度泛化或采样。", "url": "https://arxiv.org/abs/2609.37959", "pdf": "https://arxiv.org/pdf/2609.37959", "code": "https://github.com/google-research/tabfm", "keywords": [ "TFM", "表格基础模型", "Google TabFM" ], "aliases": [ "Google TabFM", "TabFM", "TabFM+", "TabFM Plus" ], "sources": [ "https://research.google/blog/introducing-tabfm-a-zero-shot-foundation-model-for-tabular-data/", "https://arxiv.org/abs/2609.37959" ], "dateNote": "模型技术说明于 2026-06-30 公开；论文于 2026-09-29 提交 arXiv。", "addedAt": "2026-10-12" }, { "id": "arxiv-2606.30410", "title": "Beyond IID: How General Are Tabular Foundation Models, Really?", "date": "2026-06-29", "updated": "2026-06-29", "category": 1, "related": [ 2, 4, 5, 6 ], "topic": "技术路线与生态", "status": "published", "venue": "NeurIPS 2026 · Evaluations and Datasets Track", "summary": "BeyondArena 与 DataFoundry 扩展到时间／分组切分、大表、宽表和文本等数据条件。142 个数据集、11 类模型，分数据条件报告结果，使用重复实验、调参与评测消融。结论对应当时模型版本和计算预算；部分 TFM 仅评估到十万行，不能把未覆盖区域直接判成模型失败。", "url": "https://arxiv.org/abs/2606.30410", "pdf": "https://arxiv.org/pdf/2606.30410", "code": null, "keywords": [], "section": "industry-benchmarks", "topicId": "general-benchmarks", "sources": [ "https://arxiv.org/abs/2606.30410", "https://neurips.cc/virtual/2026/poster/139490", "https://openreview.net/forum?id=Mk0ohjszwO" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2606.15314", "title": "LLMs on Tabular Data with Limited Semantics: Evidence from Industrial Car Retrofit Prediction", "date": "2026-06-13", "updated": "2026-06-13", "category": 1, "related": [ 4 ], "topic": "结构化数据应用与真实场景验证", "status": "preprint", "venue": "预印本", "summary": "在汽车原型改装数据上比较树模型、LLM 嵌入、直接提示和混合堆叠。车辆级任务用 80/20 分层随机切分，月度预测另设 16 个月留出；混合 LLM 结果仅在 200 行子集评估。200 行结果不能和全量基线直接比较；类别哈希削弱语义，随机切分也不等于跨时间部署验证。", "url": "https://arxiv.org/abs/2606.15314", "pdf": "https://arxiv.org/pdf/2606.15314", "code": null, "keywords": [], "section": "industry-applications", "topicId": "engineering-applications", "addedAt": "2026-10-12" }, { "id": "arxiv-2606.07345", "title": "TabSwift: An Efficient Tabular Foundation Model with Row-Wise Attention", "date": "2026-06-05", "updated": "2026-06-05", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "published", "venue": "ICML 2026", "summary": "TabSwift 用行级注意力、门控和 register token 构建轻量模型，并采用逐查询提前退出。在 TALENT 分类和回归任务评估，提供组件消融、退出阈值实验及 Wilcoxon–Holm 多重比较。宽表需投影或 PCA，可能丢失特征信息；阈值与精度成本取舍需在目标分布验证。", "url": "https://arxiv.org/abs/2606.07345", "pdf": "https://arxiv.org/pdf/2606.07345", "code": null, "keywords": [ "TFM", "表格基础模型", "TabSwift" ], "aliases": [ "TabSwift" ], "sources": [ "https://arxiv.org/abs/2606.07345", "https://proceedings.mlr.press/v306/liu26fm.html" ], "recognition": { "label": "Spotlight", "source": "https://arxiv.org/abs/2606.07345", "verifiedAt": "2026-10-12" }, "addedAt": "2026-10-12" }, { "id": "arxiv-2606.04485", "title": "LimiX-2M: Mitigating Low-Rank Collapse and Attention Bottlenecks in Tabular Foundation Models", "date": "2026-06-03", "updated": "2026-06-04", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "published", "venue": "ICML 2026", "summary": "LimiX-2M 用径向基编码 RaBEL 和重排的双轴注意力缓解数值表示与计算路径瓶颈。包含同为两百万参数的对照、编码与模块消融，以及多套分类和回归基准。部分编码基线改为共享投影，结论须保留该条件；合成图上的注意力现象不构成真实因果识别证明。", "url": "https://arxiv.org/abs/2606.04485", "pdf": "https://arxiv.org/pdf/2606.04485", "code": "https://github.com/limix-ldm-ai/LimiX", "keywords": [ "TFM", "表格基础模型", "LimiX" ], "aliases": [ "LimiX-2M", "LimiX2M", "LimiX 2M", "RaBEL" ], "sources": [ "https://arxiv.org/abs/2606.04485", "https://proceedings.mlr.press/v306/wang26kw.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2605.18635", "title": "Data Presentation Over Architecture: Resampling Strategies for Credit Risk Prediction with Tabular Foundation Models", "date": "2026-05-18", "updated": "2026-05-18", "category": 1, "related": [ 2, 5 ], "topic": "结构化数据应用", "status": "preprint", "venue": "预印本", "summary": "在固定上下文预算下比较信用风险任务的类别平衡、重采样和覆盖性选样。在 Home Credit 与 Lending Club 比较七种构造策略，后者使用时间切分；报告 AUC、MCC 与违约召回。仅两个数据集，传统基线未调分类阈值或类别权重；不能由默认阈值下的零召回推出 TFM 架构普遍更强。", "url": "https://arxiv.org/abs/2605.18635", "pdf": "https://arxiv.org/pdf/2605.18635", "code": null, "keywords": [], "section": "industry-applications", "topicId": "credit-risk", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.18147", "title": "Foundation Models for Credit Risk Prediction: A Game Changer?", "date": "2026-05-18", "updated": "2026-05-18", "category": 1, "related": [ 2 ], "topic": "结构化数据应用", "status": "preprint", "venue": "预印本", "summary": "比较表格基础模型与传统方法在违约概率和违约损失率预测中的适用性。覆盖 14 个 PD、7 个 LGD 数据集，分别比较 29 与 22 种方法，并分析数据量与排名差异。LGD 中仅 TabPFNv2 作为基础模型，部分数据私有；组内结果不能外推为所有 TFM 或真实放贷效果。", "url": "https://arxiv.org/abs/2605.18147", "pdf": "https://arxiv.org/pdf/2605.18147", "code": null, "keywords": [], "section": "industry-applications", "topicId": "credit-risk", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.18383", "title": "TabH2O: A Unified Foundation Model for Tabular Prediction", "date": "2026-05-18", "updated": "2026-08-13", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 5 ], "topic": "通用表格模型", "status": "report", "venue": "H2O.ai 技术报告", "summary": "TabH2O 在 TabICL 架构上用统一模型和双输出头处理分类、回归，并加入稳定化与噪声建模。在 TALENT 300 任务比较，v1.1 另提供 TabArena；两个检查点各自对相同五个基线排名。两个独立排名池的名次不能直接相减；统一模型、训练预算与先验变化未完全解耦。", "url": "https://arxiv.org/abs/2605.18383", "pdf": "https://arxiv.org/pdf/2605.18383", "code": null, "keywords": [ "TFM", "表格基础模型", "TabH2O" ], "aliases": [ "TabH2O", "TabH2O v1.1" ], "sources": [ "https://arxiv.org/abs/2605.18383", "https://tabh2o.h2oai.com/" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2605.13986", "title": "TabPFN-3: Technical Report", "date": "2026-05-12", "updated": "2026-05-28", "category": 1, "related": [ 2, 4, 5, 6 ], "topic": "模型与产品动态", "status": "report", "venue": "技术报告", "summary": "TabPFN-3 改用行压缩后的上下文学习，并提供行分块、小 KV 缓存与多种推理配置。公开基准覆盖 TabArena、TALENT 与文本表格，另有大数据、高维、多类及分位数回归的内部基准。公开与内部基准证据分开；单次推理、Plus、Thinking 和蒸馏配置的性能与开销不能混用。", "url": "https://arxiv.org/abs/2605.13986", "pdf": "https://arxiv.org/pdf/2605.13986", "code": null, "keywords": [ "TabPFN", "TFM" ], "section": "industry-models", "topicId": "tabular-releases", "aliases": [ "TabPFN-3", "TabPFN v3", "TabPFN3" ], "sources": [ "https://arxiv.org/abs/2605.13986", "https://priorlabs.ai/technical-reports/tabpfn-3" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2605.12292", "title": "STRABLE: Benchmarking Tabular Machine Learning with Strings", "date": "2026-05-12", "updated": "2026-05-12", "category": 1, "related": [ 4, 5, 6 ], "topic": "语义表格评测与技术路线比较", "status": "published", "venue": "NeurIPS 2026 · Evaluations and Datasets Track", "summary": "STRABLE 保留原始字符串，比较编码、降维与表格学习器的模块化方案。108 个任务来自 33 个来源和八个领域，剔除明显目标泄漏，检查跨领域与预处理变化下的排名稳定性。基准采用 IID 快照，至少两列字符串且限制行数；排名稳定性不证明时间漂移下的可迁移性。", "url": "https://arxiv.org/abs/2605.12292", "pdf": "https://arxiv.org/pdf/2605.12292", "code": "https://github.com/soda-inria/strable", "keywords": [], "section": "industry-benchmarks", "topicId": "semantic-relational-benchmarks", "sources": [ "https://arxiv.org/abs/2605.12292", "https://neurips.cc/virtual/2026/poster/139430", "https://openreview.net/forum?id=RIVxz1UYR5" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2604.21696", "title": "Towards Universal Tabular Embeddings: A Benchmark Across Data Tasks", "date": "2026-04-23", "updated": "2026-09-02", "category": 1, "related": [ 4 ], "topic": "语义表格评测与技术路线比较", "status": "preprint", "venue": "预印本", "summary": "TEmBed 统一评估单元格、行、列和整表嵌入，覆盖检索、匹配与预测。包含 11 类任务、87 个公开数据集，使用确定性测试及统一接口，并按任务等权聚合名次。只纳入无需任务微调且支持多任务的嵌入方法；总名次会掩盖效果幅度和不同粒度的覆盖差异。", "url": "https://arxiv.org/abs/2604.21696", "pdf": "https://arxiv.org/pdf/2604.21696", "code": "https://github.com/IBM/table-representation-evals", "keywords": [], "section": "industry-benchmarks", "topicId": "semantic-relational-benchmarks", "addedAt": "2026-10-12" }, { "id": "arxiv-2604.12596", "title": "KumoRFM-2: Scaling Foundation Models for Relational Learning", "date": "2026-04-14", "updated": "2026-04-14", "category": 1, "related": [ 4, 5, 6 ], "topic": "模型与产品动态", "status": "report", "venue": "技术报告", "summary": "KumoRFM-2 用任务条件化的表内注意力、主外键图注意力与跨样本交互建模关系表。在 RelBenchV1/V2、SALT、4DBInfer 遵循时间约束，默认上下文最多一万例，另评估微调与稀疏性。生产库总行数不等于单次上下文容量；默认使用训练与验证样本作上下文，比较时须核对数据使用口径。", "url": "https://arxiv.org/abs/2604.12596", "pdf": "https://arxiv.org/pdf/2604.12596", "code": null, "keywords": [ "Kumo", "TFM", "KumoRFM" ], "section": "industry-models", "topicId": "relational-releases", "aliases": [ "KumoRFM-2", "Kumo RFM 2" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2602.11139", "title": "TabICLv2: A better, faster, scalable, and open tabular foundation model", "date": "2026-02-11", "updated": "2026-09-16", "category": 1, "related": [ 5, 6 ], "topic": "模型与产品动态", "status": "published", "venue": "ICML 2026", "summary": "TabICLv2 改进特征分组、目标编码和长上下文注意力尺度，并分别支持分类、回归。在 TabArena 与 TALENT 比较，并提供模块消融和训练配置；两套基准采用不同集成规模。TALENT 速度图混用 H100 实测和既有耗时，不能当统一硬件加速比；论文的分布回归验证主要限于玩具数据。", "url": "https://arxiv.org/abs/2602.11139", "pdf": "https://arxiv.org/pdf/2602.11139", "code": "https://github.com/soda-inria/tabicl", "keywords": [ "TabICL", "TFM" ], "section": "industry-models", "topicId": "tabular-releases", "sources": [ "https://arxiv.org/abs/2602.11139", "https://proceedings.mlr.press/v306/qu26a.html" ], "aliases": [ "TabICLv2", "TabICL v2", "TabICL2" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2511.08667", "title": "TabPFN-2.5: Advancing the State of the Art in Tabular Foundation Models", "date": "2025-11-06", "updated": "2026-02-05", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5, 6 ], "topic": "通用表格模型", "status": "report", "venue": "Prior Labs 技术报告", "summary": "TabPFN-2.5 扩展模型深度、特征分组与计算 token，并提供 MLP／树模型蒸馏路径。公共实验主要采用 51 表的 TabArena-Lite 单测试折，另有内部基准与蒸馏耗时分析。Lite 与完整多折 TabArena 不等价；原版、真实数据微调版及调参版的成本与效果应分别看待。", "url": "https://arxiv.org/abs/2511.08667", "pdf": "https://arxiv.org/pdf/2511.08667", "code": "https://github.com/PriorLabs/TabPFN", "keywords": [ "TFM", "表格基础模型", "TabPFN" ], "aliases": [ "TabPFN-2.5", "TabPFN2.5", "TabPFN v2.5" ], "sources": [ "https://arxiv.org/abs/2511.08667", "https://priorlabs.ai/technical-reports/tabpfn-2-5-model-report", "https://storage.googleapis.com/prior-labs-tabpfn-public/reports/TabPFN_2_5_tech_report.pdf?date=2025-11-06" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2511.02818", "title": "Orion-MSP: Multi-Scale Sparse Attention for Tabular In-Context Learning", "date": "2025-11-04", "updated": "2025-11-07", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "Orion-MSP 组合多尺度特征分组、块稀疏注意力和 Perceiver 式跨组件记忆。在 TALENT、CC18、TabZilla 的共同可运行子集比较，分别保留 154、63、27 个任务。因 OOM 或运行错误取模型交集会排除困难任务；结果不能代表完整基准或据此证明普遍扩展优势。", "url": "https://arxiv.org/abs/2511.02818", "pdf": "https://arxiv.org/pdf/2511.02818", "code": "https://github.com/Lexsi-Labs/Orion-MSP", "keywords": [ "TFM", "表格基础模型", "Orion-MSP" ], "aliases": [ "Orion-MSP", "Orion MSP" ], "sources": [ "https://arxiv.org/abs/2511.02818" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2510.21204", "title": "Mitra: Mixed Synthetic Priors for Enhancing Tabular Foundation Models", "date": "2025-10-24", "updated": "2025-10-24", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "published", "venue": "NeurIPS 2025", "summary": "Mitra 研究多类合成先验的性能、差异性与互补性，并在一维和二维注意力骨干上验证混合先验。用 TabRepo、TabZilla、AMLB 等分类／回归任务、十折评测及先验消融，分别比较集成和微调。不同特征规模采用不同任务子集；混合先验有效不代表对任意骨干或分布都有效。", "url": "https://arxiv.org/abs/2510.21204", "pdf": "https://arxiv.org/pdf/2510.21204", "code": null, "keywords": [ "TFM", "表格基础模型", "Mitra" ], "aliases": [ "Mitra", "Mitra v1" ], "sources": [ "https://arxiv.org/abs/2510.21204" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2509.03505", "title": "LimiX: Unleashing Structured-Data Modeling Capability for Generalist Intelligence", "date": "2025-09-03", "updated": "2025-11-07", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 5 ], "topic": "通用表格模型", "status": "report", "venue": "LimiX 技术报告", "summary": "LimiX 原始报告以单元格建模和随机掩码统一分类、回归、补全与条件生成。报告多组公开基准、BCCO 与各任务实验；分类评测筛除了超出样本、特征和类别上限的数据集。联合建模目标不保证所有真实分布上的条件一致性；有限尺度的幂律拟合也不能保证继续扩展后的收益。", "url": "https://arxiv.org/abs/2509.03505", "pdf": "https://arxiv.org/pdf/2509.03505", "code": "https://github.com/limix-ldm-ai/LimiX", "keywords": [ "TFM", "表格基础模型", "LimiX" ], "aliases": [ "LimiX", "LimiX-16M", "LimiX 16M", "LDM" ], "sources": [ "https://arxiv.org/abs/2509.03505" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2506.16791", "title": "TabArena: A Living Benchmark for Machine Learning on Tabular Data", "date": "2025-06-20", "updated": "2025-06-20", "category": 1, "related": [ 2, 5, 6 ], "topic": "技术路线与生态", "status": "published", "venue": "NeurIPS 2025 Datasets & Benchmarks", "summary": "TabArena 将模型、调参、集成与计算预算放进持续维护的表格评测协议。初版比较 16 类模型和 51 个精选数据集，按数据量执行重复三折交叉验证，并开放预测及元数据。CPU 与 GPU 模型使用不同设备，部分榜单会填补缺失结果；Lite 单折和完整评测须区分，固定快照才能复现排名。", "url": "https://arxiv.org/abs/2506.16791", "pdf": "https://arxiv.org/pdf/2506.16791", "code": "https://tabarena.ai", "keywords": [], "section": "industry-benchmarks", "topicId": "general-benchmarks", "recognition": { "label": "Spotlight", "source": "https://arxiv.org/abs/2506.16791", "verifiedAt": "2026-10-12" }, "addedAt": "2026-10-12" }, { "id": "arxiv-2506.05584", "title": "TabFlex: Scaling Tabular Learning to Millions with Linear Attention", "date": "2025-06-05", "updated": "2025-06-05", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "published", "venue": "ICML 2025", "summary": "TabFlex 用非因果掩码的线性注意力替代标准注意力，扩展表格上下文规模。在 115 个 OpenML 数据集及困难 TabZilla 子集比较性能和速度，并研究降维与采样。这里的非因果指注意力掩码，不是因果推断；可处理更大输入不等于所有大表任务都保持精度。", "url": "https://arxiv.org/abs/2506.05584", "pdf": "https://arxiv.org/pdf/2506.05584", "code": "https://github.com/microsoft/ticl", "keywords": [ "TFM", "表格基础模型", "TabFlex" ], "aliases": [ "TabFlex" ], "sources": [ "https://arxiv.org/abs/2506.05584", "https://proceedings.mlr.press/v267/zeng25b.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2502.05564", "title": "TabICL: A Tabular Foundation Model for In-Context Learning on Large Data", "date": "2025-02-08", "updated": "2025-05-24", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "published", "venue": "ICML 2025", "summary": "原版 TabICL 通过列编码和行压缩构造固定维度样本表示，再进行跨样本上下文分类。主要评测 171 个至多十类的 TALENT 任务，排除开发集；TabICL 与 TabPFNv2 各使用 32 次集成。其他深度基线未统一集成，TabPFNv2 支持集被限为三万行；时间和精度优势要保留这些协议条件。", "url": "https://arxiv.org/abs/2502.05564", "pdf": "https://arxiv.org/pdf/2502.05564", "code": "https://github.com/soda-inria/tabicl", "keywords": [ "TFM", "表格基础模型", "TabICL" ], "aliases": [ "TabICL", "TabICLv1", "TabICL v1" ], "sources": [ "https://arxiv.org/abs/2502.05564", "https://proceedings.mlr.press/v267/qu25d.html" ], "addedAt": "2026-10-12" }, { "id": "paper-612809da03", "title": "Accurate predictions on small data with a tabular foundation model", "date": "2025-01-08", "updated": "2025-01-08", "category": 1, "related": [ 2, 5, 6 ], "topic": "模型与产品动态", "status": "published", "venue": "Nature 2025", "summary": "TabPFN-v2 以单元格表示和行列交互，通过带标签上下文完成分类与回归。Nature 论文在至多一万样本、五百特征的数据上比较默认和调参方法，报告区间及逐数据集结果。性能经过逐数据集归一化，不能当原始 AUC 增量；一次适配加预测的耗时也不等于持续服务延迟。", "url": "https://www.nature.com/articles/s41586-024-08328-6", "pdf": "https://www.nature.com/articles/s41586-024-08328-6.pdf", "code": null, "keywords": [ "TabPFN", "TFM" ], "section": "industry-models", "topicId": "tabular-releases", "aliases": [ "TabPFN-v2", "TabPFN v2", "TabPFN2" ], "addedAt": "2026-10-12", "sources": [ "https://www.nature.com/articles/s41586-024-08328-6" ] }, { "id": "arxiv-2410.18164", "title": "TabDPT: Scaling Tabular Foundation Models on Real Data", "date": "2024-10-23", "updated": "2026-01-17", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "published", "venue": "NeurIPS 2025", "summary": "TabDPT 用真实表格自监督学习和近邻上下文检索训练行级 Transformer。比较 CC18、CTR23，提供无验证集变体、消融及推理速度；报告置信区间而非只给排名。逐查询检索增加预测开销；真实训练数据需要逐版本核对与评测集的重复，不能沿用后续 Turbo 的去重结论。", "url": "https://arxiv.org/abs/2410.18164", "pdf": "https://arxiv.org/pdf/2410.18164", "code": "https://github.com/layer6ai-labs/TabDPT-inference", "keywords": [ "TFM", "表格基础模型", "TabDPT" ], "aliases": [ "TabDPT", "TabDPT v1" ], "sources": [ "https://arxiv.org/abs/2410.18164" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2407.20060", "title": "RelBench: A Benchmark for Deep Learning on Relational Databases", "date": "2024-07-29", "updated": "2024-07-29", "category": 1, "related": [ 4, 5 ], "topic": "技术路线与生态", "status": "published", "venue": "NeurIPS 2024", "summary": "RelBench 将主外键关联的数据库定义成带实体、预测时点和目标的关系学习任务。初版含七个数据库、三十个实体分类／回归／链接预测任务，按预测时点屏蔽未来数据。任务间可能共享数据库与实体，不能当三十个独立场景；人工特征工程对照也取决于专家投入。", "url": "https://arxiv.org/abs/2407.20060", "pdf": "https://arxiv.org/pdf/2407.20060", "code": "https://github.com/stanford-star/relbench", "keywords": [], "section": "industry-benchmarks", "topicId": "semantic-relational-benchmarks", "addedAt": "2026-10-12" }, { "id": "arxiv-2406.19380", "title": "TabReD: Analyzing Pitfalls and Filling the Gaps in Tabular Deep Learning Benchmarks", "date": "2024-06-27", "updated": "2024-06-27", "category": 1, "related": [ 2, 6 ], "topic": "技术路线与生态", "status": "published", "venue": "ICLR 2025", "summary": "TabReD 用时间切分和丰富工程特征检查常见表格深度学习结论。在训练／验证／测试时间顺序下调参与早停，重复十五个随机种子，并对随机切分及漂移处理做对照。它补充工业式数据条件而非代表全部行业；部分大数据集经过采样，模型版本也对应论文当时。", "url": "https://arxiv.org/abs/2406.19380", "pdf": "https://arxiv.org/pdf/2406.19380", "code": null, "keywords": [], "section": "industry-benchmarks", "topicId": "general-benchmarks", "addedAt": "2026-10-12" }, { "id": "arxiv-2402.14335", "title": "HyperFast: Instant Classification for Tabular Data", "date": "2024-02-22", "updated": "2024-02-22", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5, 6 ], "topic": "通用表格模型", "status": "published", "venue": "AAAI 2024", "summary": "HyperFast 让超网络读取支持集并生成任务专用分类器，再用生成的小网络预测。在元测试任务比较一次生成、集成及进一步优化，报告拟合加预测的时间预算曲线。单次生成只用固定支持样本，未必覆盖大数据分布；集成和微调结果不能作为一次生成的效果。", "url": "https://arxiv.org/abs/2402.14335", "pdf": "https://arxiv.org/pdf/2402.14335", "code": "https://github.com/AI-sandbox/HyperFast", "keywords": [ "TFM", "表格基础模型", "HyperFast" ], "aliases": [ "HyperFast" ], "sources": [ "https://arxiv.org/abs/2402.14335" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2207.01848", "title": "TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second", "date": "2022-07-05", "updated": "2023-09-16", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5, 6 ], "topic": "通用表格模型", "status": "published", "venue": "ICLR 2023", "summary": "原版 TabPFN 在合成任务上训练 Transformer，用标签上下文直接完成小型数值表格分类。包含玩具边界与 OpenML 分类比较，另分析类别特征、缺失值和多分类条件。原实验特征和类别范围有限，类别列及缺失值会影响表现；样本侧二次复杂度限制大上下文。", "url": "https://arxiv.org/abs/2207.01848", "pdf": "https://arxiv.org/pdf/2207.01848", "code": "https://github.com/PriorLabs/TabPFN", "keywords": [ "TFM", "表格基础模型", "TabPFN" ], "aliases": [ "TabPFN", "TabPFN v1", "TabPFNv1" ], "sources": [ "https://arxiv.org/abs/2207.01848", "https://openreview.net/forum?id=cp5PvcI6w8_" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.13654", "title": "TabReliBench: Evaluating Tabular Foundation Models Beyond Predictive Accuracy", "date": "2026-10-09", "updated": "2026-10-09", "category": 2, "section": "prediction-distributions", "topicId": "predictive-distributions", "related": [ 1 ], "topic": "条件分布与不确定性分解", "status": "preprint", "venue": "预印本", "summary": "TabReliBench 把点预测、概率质量、边际／子群校准和区间效率分开评估。固定切分进行集成规模实验，保形后处理另划辅助、校准与测试集，观察不同指标和数据集的分歧。改善总体覆盖仍会留下子群缺口；跨数据集元特征关联不能直接解释成模型优劣的因果机制。", "url": "https://arxiv.org/abs/2610.13654", "pdf": "https://arxiv.org/pdf/2610.13654", "code": "https://github.com/Tianqi-Zhao/TabReliBench", "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "TabReliBench" ], "sources": [ "https://arxiv.org/abs/2610.13654" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.06693", "title": "Adapting prior-data fitted networks for tabular anomaly detection", "date": "2026-10-05", "updated": "2026-10-05", "category": 2, "related": [ 5, 6 ], "topic": "复杂数据下的预测优化", "status": "preprint", "venue": "预印本", "summary": "ZEN 用冻结 TabPFN 表征的邻近距离检测异常，FOCUS 进一步做单类微调。在 47 个 ADBench 表格任务区分干净与污染训练集，并修正小数据集复制后再切分的泄漏问题。大数据先限制到一万样本，多折与多特征子集推理仍昂贵；污染鲁棒性不能等同任意未知异常的保证。", "url": "https://arxiv.org/abs/2610.06693", "pdf": "https://arxiv.org/pdf/2610.06693", "code": "https://github.com/Maximilianb1/zen-focus-tabular-anomaly-detection", "keywords": [ "TabPFN" ], "section": "prediction-quality", "topicId": "anomaly-detection", "addedAt": "2026-10-12" }, { "id": "arxiv-2610.07406", "title": "Evaluation of Active Feature Acquisition Policies with Tabular Foundation Models", "date": "2026-10-05", "updated": "2026-10-05", "category": 2, "section": "prediction-decisions", "topicId": "scores-routing", "related": [ 4, 5 ], "topic": "评分、固定选项与路由", "status": "preprint", "venue": "预印本", "summary": "用贝叶斯预测推断分离认知不确定性，降低主动获取特征策略的离线价值估计偏差。合成实验检验估值、区间覆盖和策略排序，另在 OpenML 人工 MCAR 缺失及 Sepsis 回顾数据上评估。临床缺失机制未知，分析依赖指定的倾向模型与可识别性假设；离线改进不能直接视为临床收益。", "url": "https://arxiv.org/abs/2610.07406", "pdf": "https://arxiv.org/pdf/2610.07406", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "Active Feature Acquisition", "AFA" ], "sources": [ "https://arxiv.org/abs/2610.07406" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.06751", "title": "MatrixFormer: A Foundation Model for Matrix Completion", "date": "2026-10-05", "updated": "2026-10-05", "category": 2, "section": "prediction-quality", "topicId": "missing-noisy-data", "related": [ 1, 3 ], "topic": "缺失补全与噪声清洗", "status": "preprint", "venue": "预印本", "summary": "MatrixFormer 用行列交替注意力完成整张矩阵，并按已观测项误差混合通用与推荐专家。覆盖表格补全、评分矩阵、模型评测分数和因果面板，107 倍提速来自固定 10 列、1024 行的单专家与 TabImpute 比较。完整集成需两次前向；混合权重未用独立留出项选择，处理效应区间的有效性也未验证。", "url": "https://arxiv.org/abs/2610.06751", "pdf": "https://arxiv.org/pdf/2610.06751", "code": null, "keywords": [ "TFM", "表格基础模型", "MatrixFormer" ], "aliases": [ "MatrixFormer" ], "sources": [ "https://arxiv.org/abs/2610.06751" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.29769", "title": "Decision Models vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places", "date": "2026-09-24", "updated": "2026-09-28", "category": 2, "related": [ 1, 4 ], "topic": "固定选项与评分决策、概率校准及置信度路由", "status": "preprint", "venue": "预印本", "summary": "比较 Jev、Decider 与 LLM 的评分及置信度级联，发现跨模型错误相关会限制级联增益。九组二元／序数评分面板统一评估准确率、一致性、成本与延迟，并区分标注者一致和分歧样本。部分序数任务表现较弱；两种决策模型不足以代表整类技术，成本取决于当时 API 定价与请求设置。", "url": "https://arxiv.org/abs/2609.29769", "pdf": "https://arxiv.org/pdf/2609.29769", "code": null, "keywords": [ "Jev" ], "section": "prediction-decisions", "topicId": "scores-routing", "addedAt": "2026-10-12", "aliases": [ "JEV vs. LLMs as Rubric Judges: Cheaper, Faster, and Wrong in the Same Places" ] }, { "id": "arxiv-2608.06004", "title": "Do Tabular Foundation Models Agree with Themselves?", "date": "2026-08-06", "updated": "2026-08-06", "category": 2, "related": [ 5 ], "topic": "预测概率与分布优化", "status": "preprint", "venue": "预印本", "summary": "用边缘化一致性与分解顺序一致性检验 TFM 多目标预测能否对应同一联合分布。对多种 TFM 和 OpenML 成对目标比较总变差距离，并分析一致性与预测精度的关系。回归一致性依赖离散化近似，样本删除缺失行；观察到的不一致不等于所有单目标预测失效。", "url": "https://arxiv.org/abs/2608.06004", "pdf": "https://arxiv.org/pdf/2608.06004", "code": null, "keywords": [], "section": "prediction-distributions", "topicId": "probability-calibration", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.31272", "title": "Algorithmic Recourse of In-Context Learning for Tabular Data", "date": "2026-05-29", "updated": "2026-07-15", "category": 2, "section": "prediction-decisions", "topicId": "strategic-feedback", "related": [ 5 ], "topic": "策略响应与反馈分布", "status": "published", "venue": "ICML 2026", "summary": "ASR-ICL 用黑盒零阶查询与自适应特征子空间搜索稀疏反事实修改。在五个分类数据集、TabPFN／TabICL 与 LLM 上考察有效率、修改成本及上下文重采样稳定性。收敛理论针对线性 ICL；部分比较使用不同预测模型，修改特征使预测翻转并不证明现实因果可行性。", "url": "https://arxiv.org/abs/2605.31272", "pdf": "https://arxiv.org/pdf/2605.31272", "code": null, "keywords": [], "aliases": [ "ASR-ICL", "Algorithmic Recourse" ], "sources": [ "https://arxiv.org/abs/2605.31272", "https://proceedings.mlr.press/v306/dong26n.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2605.19662", "title": "When Tabular Foundation Models Meet Strategic Tabular Data: A Prior Alignment Approach", "date": "2026-05-19", "updated": "2026-06-06", "category": 2, "related": [ 5 ], "topic": "决策反馈引起的数据变化", "status": "published", "venue": "ICML 2026", "summary": "SPN 用模拟个体策略性改动前后的样本对重建上下文，在推理时适配决策诱发的分布变化。在策略与普通表格任务上比较，并设置 80% 策略性输入的测试混合分布和多种操纵成本。收益依赖所设的个体响应与成本模型；模拟操纵不能直接代表现实博弈或部署后的行为变化。", "url": "https://arxiv.org/abs/2605.19662", "pdf": "https://arxiv.org/pdf/2605.19662", "code": null, "keywords": [], "section": "prediction-decisions", "topicId": "strategic-feedback", "sources": [ "https://arxiv.org/abs/2605.19662", "https://proceedings.mlr.press/v306/lv26g.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2605.18696", "title": "Ensembling Tabular Foundation Models - A Diversity Ceiling And A Calibration Trap", "date": "2026-05-18", "updated": "2026-05-18", "category": 2, "related": [ 1, 5 ], "topic": "集成预测与概率校准", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "比较六种 TFM 的集成，检验错误相关性及准确率改善与概率校准的冲突。以 153 个候选分类任务、独立验证集学习集成权重，并报告准确率、log-loss 与计算成本。单数据集只有一个种子，主表限全部方法完成的交集；未比较 TFM 与 GBDT 的异类集成，不能据此宣称所有集成都遇到上限。", "url": "https://arxiv.org/abs/2605.18696", "pdf": "https://arxiv.org/pdf/2605.18696", "code": null, "keywords": [], "section": "prediction-distributions", "topicId": "probability-calibration", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.04363", "title": "Mitigating Label Shift in Tabular In-Context Learning via Test-Time Posterior Adjustment", "date": "2026-05-06", "updated": "2026-05-24", "category": 2, "related": [ 5 ], "topic": "预测概率与分布优化", "status": "published", "venue": "ICML 2026", "summary": "DistPFN 按预测后验与训练类别频率的比例重加权，温度版本控制修正强度。在 253 个分类数据集、五个种子及人工标签比例偏移下评估，并补充 TableShift 比较。该调整是部分修正，不解决特征分布变化；人工反频率采样下的大增益不能代表任意真实漂移。", "url": "https://arxiv.org/abs/2605.04363", "pdf": "https://arxiv.org/pdf/2605.04363", "code": "https://github.com/seunghan96/DistPFN", "keywords": [], "section": "prediction-distributions", "topicId": "probability-calibration", "sources": [ "https://arxiv.org/abs/2605.04363", "https://proceedings.mlr.press/v306/lee26aw.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2604.25154", "title": "Model-Aware Data Cleaning for Tabular Foundation Models", "date": "2026-04-28", "updated": "2026-09-17", "category": 2, "related": [ 5, 6 ], "topic": "复杂数据下的预测优化", "status": "preprint", "venue": "预印本", "summary": "L2C-TFM 通过强化学习选择清洗序列，将 TFM 验证表现与数据分布稳定性纳入奖励。十个主要任务采用内层选择、外层测试及八种子评估，另比较十三任务上的清洗加模型与 AutoML。部分贪心结果标为测试集选择的乐观上界，不能作为可部署成绩；其他清洗系统为近似重实现，小数据清洗选择也更慢。", "url": "https://arxiv.org/abs/2604.25154", "pdf": "https://arxiv.org/pdf/2604.25154", "code": null, "keywords": [], "section": "prediction-quality", "topicId": "missing-noisy-data", "addedAt": "2026-10-12" }, { "id": "arxiv-2603.29928", "title": "ScoringBench: A Benchmark for Evaluating Tabular Foundation Models with Proper Scoring Rules", "date": "2026-03-31", "updated": "2026-03-31", "category": 2, "related": [ 1, 6 ], "topic": "预测概率与分布优化", "status": "preprint", "venue": "预印本", "summary": "ScoringBench 用恰当评分规则比较完整预测分布，并探索按这些规则微调。汇总去重回归任务，把直方图、分位数与采样输出转为统一表示，同时报告排名和效应量。输出转换包含离散化和平滑假设；部分微调排名改善的效应量很小，榜单为论文日期的快照。", "url": "https://arxiv.org/abs/2603.29928", "pdf": "https://arxiv.org/pdf/2603.29928", "code": "https://github.com/jonaslandsgesell/ScoringBench", "keywords": [], "section": "prediction-distributions", "topicId": "predictive-distributions", "addedAt": "2026-10-12" }, { "id": "arxiv-2603.26611", "title": "Benchmarking Tabular Foundation Models for Conditional Density Estimation in Regression", "date": "2026-03-27", "updated": "2026-03-27", "category": 2, "related": [ 1, 6 ], "topic": "预测概率与分布优化", "status": "published", "venue": "TMLR · 2026", "summary": "比较 TFM 与传统方法对回归条件密度的预测、校准和时间成本。39 个 OpenML／SDSS 任务采用多种样本规模、五次重复及密度评分，并单独分析测光红移。原生输出需转换为密度；大样本高维任务存在显存失败，离散或匹配已知分布的专用模型仍可能占优。", "url": "https://arxiv.org/abs/2603.26611", "pdf": "https://arxiv.org/pdf/2603.26611", "code": "https://github.com/rizbicki/tabDensityComparisons", "keywords": [], "section": "prediction-distributions", "topicId": "predictive-distributions", "addedAt": "2026-10-12" }, { "id": "arxiv-2602.04596", "title": "Uncertainty Decomposition for Bayes-Filtered Transformers via Bayesian Predictive Inference", "date": "2026-02-04", "updated": "2026-07-19", "category": 2, "related": [ 5 ], "topic": "预测不确定性分解", "status": "preprint", "venue": "预印本", "summary": "通过预测中心极限定理，将 Bayes-filtered Transformer 的不确定性分解为认知与数据噪声部分。验证合成生成过程的覆盖与条件诊断，并给出 TabPFN 和两个真实数据例子。理论所需条件未在 TabPFN 上实证验证，收敛率假设为 1；远离数据流形时认知不确定性也可能偏低。", "url": "https://arxiv.org/abs/2602.04596", "pdf": "https://arxiv.org/pdf/2602.04596", "code": "https://github.com/weiyaw/ud4pfn", "keywords": [ "TabPFN" ], "section": "prediction-distributions", "topicId": "predictive-distributions", "addedAt": "2026-10-12" }, { "id": "arxiv-2602.03018", "title": "From Zero to Hero: Advancing Zero-Shot Foundation Models for Tabular Outlier Detection", "date": "2026-02-03", "updated": "2026-05-29", "category": 2, "section": "prediction-quality", "topicId": "anomaly-detection", "related": [ 1, 5 ], "topic": "表格异常检测", "status": "published", "venue": "ICML 2026", "summary": "OutFormer 通过混合异常先验与自演进课程训练，实现新表格上的无标签异常检测。除 ADBench 外新增 OddBench、OvRBench，并做课程消融和类别特征实验；OvRBench 含从 TabArena 等分类集改造的任务。一对其余类别构造的异常不等于真实开放异常；类别数据先验与上下文规模仍有限。", "url": "https://arxiv.org/abs/2602.03018", "pdf": "https://arxiv.org/pdf/2602.03018", "code": "https://github.com/psorus/Outformer.git", "keywords": [ "TFM", "OUTFORMER" ], "aliases": [ "OUTFORMER" ], "sources": [ "https://arxiv.org/abs/2602.03018", "https://proceedings.mlr.press/v306/ding26s.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2601.22371", "title": "FIRE: Multi-fidelity Regression with Distribution-conditioned In-context Learning using Tabular Foundation Models", "date": "2026-01-29", "updated": "2026-01-29", "category": 2, "section": "prediction-distributions", "topicId": "predictive-distributions", "related": [ 1, 5 ], "topic": "条件分布与不确定性分解", "status": "published", "venue": "ICML 2026", "summary": "FIRE 将低保真预测分布作为高保真残差模型的输入，串联冻结 TFM 完成多保真回归。31 个任务涵盖合成函数、学习曲线和工程仿真，并与多保真 GP 及深度模型比较。仍受上下文与注意力成本限制，对轻量 GP 的速度优势更有限；实验尚未验证闭环多保真优化收益。", "url": "https://arxiv.org/abs/2601.22371", "pdf": "https://arxiv.org/pdf/2601.22371", "code": "https://github.com/rosenyu304/FIRE", "keywords": [], "aliases": [ "FIRE", "Fidelity-aware In-context REgression" ], "sources": [ "https://arxiv.org/abs/2601.22371", "https://decode.mit.edu/projects/fire/", "https://proceedings.mlr.press/v306/yu26au.html" ], "recognition": { "label": "Spotlight", "source": "https://decode.mit.edu/projects/fire/", "verifiedAt": "2026-10-12" }, "addedAt": "2026-10-12" }, { "id": "paper-3b87eab6aa", "title": "Conformal Prediction for Tabular Prior-Data Fitted Networks with Missing Data", "date": "2025", "updated": "2025", "category": 2, "related": [ 5 ], "topic": "缺失模式条件下的预测区间校准", "status": "workshop", "venue": "EurIPS 2025 Tabular Data Workshop", "summary": "将按缺失模式构造校准集的 CP-MDA-exact 接到 TabPFN 分位数预测上。论文摘要报告覆盖改善；作者公开 notebook 显示线性高斯模拟、MCAR 嵌套缺失及独立训练／校准／测试切分。原 PDF 当前不可访问；代码的额外分位数实验与主函数采用不同修正公式，且未见缺失模式只会告警，尚不足以支持任意缺失下的保证。", "url": "https://openreview.net/pdf?id=TnZcC7GXI5", "pdf": "https://openreview.net/pdf?id=TnZcC7GXI5", "code": "https://github.com/Fdvanleeuwen/TabPFN_CP-MDA", "keywords": [ "TabPFN" ], "section": "prediction-distributions", "topicId": "prediction-intervals", "recognition": { "label": "Oral", "source": "https://sites.google.com/view/eurips25-ai-td/home", "verifiedAt": "2026-10-12" }, "addedAt": "2026-10-12", "sources": [ "https://openreview.net/forum?id=TnZcC7GXI5", "https://github.com/Fdvanleeuwen/TabPFN_CP-MDA/blob/main/Simulation.ipynb", "https://github.com/Fdvanleeuwen/TabPFN_CP-MDA/blob/main/Functions_for_implementation.ipynb" ] }, { "id": "arxiv-2510.25154", "title": "TabMGP: Martingale Posterior with TabPFN", "date": "2025-10-29", "updated": "2026-05-28", "category": 2, "section": "prediction-distributions", "topicId": "predictive-distributions", "related": [ 3 ], "topic": "条件分布与不确定性分解", "status": "published", "venue": "ICML 2026", "summary": "TabMGP 用 TabPFN 连续预测构造鞅后验，将点预测扩展为统计量的不确定性估计。合成与真实小中样本任务检查路径稳定性、可信集覆盖和尺寸，并与贝叶斯及其他鞅后验比较。尚无保证 TabPFN 极限分布存在的完整证明；小样本高维分类可能欠覆盖，有限预算下必须检查路径收敛。", "url": "https://arxiv.org/abs/2510.25154", "pdf": "https://arxiv.org/pdf/2510.25154", "code": "https://github.com/weiyaw/tabmgp", "keywords": [], "aliases": [ "TabMGP", "Martingale Posterior" ], "sources": [ "https://arxiv.org/abs/2510.25154", "https://proceedings.mlr.press/v306/ng26a.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2510.02625", "title": "One Pipeline, Many Transformers: Pattern-Specific Imputation Specialists for Tabular Missing Data", "date": "2025-10-03", "updated": "2026-08-28", "category": 2, "related": [ 5 ], "topic": "复杂数据下的预测优化", "status": "preprint", "venue": "预印本", "summary": "TabImpute 把缺失位置编码为行列上下文预测任务，训练适配不同缺失模式的插补专家。MissBench 使用 42 个原始无缺失的数值表、11 种人工缺失模式，并比较通用与模式专用模型。专用模型需要预先匹配缺失机制，混合训练还出现遗忘；人工 MNAR 成绩不构成现实未知缺失机制的识别保证。", "url": "https://arxiv.org/abs/2510.02625", "pdf": "https://arxiv.org/pdf/2510.02625", "code": null, "keywords": [], "section": "prediction-quality", "topicId": "missing-noisy-data", "addedAt": "2026-10-12" }, { "id": "paper-dd682e2bac", "title": "Conformalized Quantile Regression", "date": "2019-05-08", "updated": "2019-05-08", "category": 2, "related": [], "topic": "回归分布的后验校准基础方法", "status": "published", "venue": "NeurIPS 2019", "summary": "CQR 在分位数回归之外使用独立校准集，构造能随输入变化宽度的预测区间。给出交换性条件下有限样本边际覆盖定理，并在 11 个回归数据集、20 次切分上比较区间长度与覆盖。保证是边际覆盖，不是每个样本或每个子群的条件覆盖；时间漂移等非交换数据不自动满足前提。", "url": "https://proceedings.neurips.cc/paper_files/paper/2019/hash/5103c3584b063c431bd1268e9b5e76fb-Abstract.html", "pdf": "https://proceedings.neurips.cc/paper_files/paper/2019/file/5103c3584b063c431bd1268e9b5e76fb-Paper.pdf", "code": "https://github.com/yromano/cqr", "keywords": [], "section": "prediction-distributions", "topicId": "prediction-intervals", "addedAt": "2026-10-12" }, { "id": "paper-d00034280f", "title": "Unveiling Prior-Data Fitted Networks on Causal Effect Estimation: Pre-Training or Fine-Tuning?", "date": "2026", "updated": "2026", "category": 3, "related": [ 5, 6 ], "topic": "因果效应模型的目标微调", "status": "published", "venue": "ICML 2026", "summary": "分析 PFN 未覆盖干预分布时的偏差，提出单点干预微调 PWF 与预算内干预分布选择 MSF。给出局部泛化条件，并在三类合成结构方程与 RealCause-Lalonde 半合成任务上比较近邻和远邻干预。理论依赖分布距离、光滑性与优化误差条件；微调需要目标干预样本，不是仅凭观测数据消除不可识别性。", "url": "https://proceedings.mlr.press/v306/wang26fl.html", "pdf": "https://raw.githubusercontent.com/mlresearch/v306/main/assets/wang26fl/wang26fl.pdf", "code": null, "keywords": [], "dateNote": "已确认 2026 年会议版；首发月份待核", "section": "causal-models", "topicId": "causal-adaptation", "sources": [ "https://proceedings.mlr.press/v306/wang26fl.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.12537", "title": "Market Impact Estimation via Causal Foundation Models", "date": "2026-10-08", "updated": "2026-10-08", "category": 3, "section": "causal-estimation", "topicId": "treatment-effects", "related": [ 1 ], "topic": "处理效应与干预分布", "status": "preprint", "venue": "预印本", "summary": "用连续时间因果 PFN 将市场冲击估计写成干预预测，并分析保形重校准。在三变量合成市场、ABIDES 仿真和加密市场事件上做干预信息消融与迁移检验。真实事件仅可能外生；迁移时原生区间严重欠覆盖，仿真反事实校准不等于真实交易中已验证因果效应。", "url": "https://arxiv.org/abs/2610.12537", "pdf": "https://arxiv.org/pdf/2610.12537", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "Causal Market Simulator" ], "sources": [ "https://arxiv.org/abs/2610.12537" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.39523", "title": "CIDER-FM: Foundation Models for Causal Inference from Diverse Experimental Regimes", "date": "2026-09-30", "updated": "2026-09-30", "category": 3, "related": [ 5 ], "topic": "观测与实验数据融合", "status": "preprint", "venue": "预印本", "summary": "CIDER-FM 用干预标记与三轴注意力融合观测和代理实验，预测条件干预分布。按可识别性划分合成图实验，并用 Causal Chambers 光学装置实测数据验证固定实验设计。评估最多十变量、每个实验仅一个干预目标，并假定共享总体和未干预机制；不保证一般不可识别问题被解决。", "url": "https://arxiv.org/abs/2609.39523", "pdf": "https://arxiv.org/pdf/2609.39523", "code": null, "keywords": [], "section": "causal-estimation", "topicId": "experimental-regimes", "addedAt": "2026-10-12" }, { "id": "arxiv-2609.03003", "title": "Causal Foundation Models", "date": "2026-09-02", "updated": "2026-09-02", "category": 3, "related": [ 1, 2, 5 ], "topic": "因果基础模型路线综述", "status": "preprint", "venue": "预印本", "summary": "系统介绍因果 PFN 的识别前提、架构与用法，并提供 CausalPFN、Do-PFN、CausalFM 的统一比较。在 RealCause-Lalonde 两个半合成队列、十次生成重复上比较效应误差与匹配 CPU 的运行时间。两队列具有共同数据背景且无混杂性由生成器保证；排名不足以代表真实未知混杂任务。", "url": "https://arxiv.org/abs/2609.03003", "pdf": "https://arxiv.org/pdf/2609.03003", "code": "https://github.com/layer6ai-labs/cfms", "keywords": [ "CausalPFN", "Do-PFN", "CausalFM" ], "section": "causal-models", "topicId": "causal-unified", "addedAt": "2026-10-12" }, { "id": "arxiv-2607.11510", "title": "DAG-FM: A Foundation Model for Causal Discovery under Heterogeneous Causal Mechanisms", "date": "2026-07-13", "updated": "2026-08-02", "category": 3, "related": [ 5 ], "topic": "因果结构发现与推理算法", "status": "preprint", "venue": "预印本", "summary": "DAG-FM 交替预测叶节点与父节点，以专家混合处理不同因果机制并构造无环图。测试异质合成机制、Sachs 与 Causal Chambers，并与传统和摊销因果发现模型比较。真实评测范围有限，部分基线显存失败；指标对反向边和环额外加罚，无环性不意味着真实因果方向已被识别。", "url": "https://arxiv.org/abs/2607.11510", "pdf": "https://arxiv.org/pdf/2607.11510", "code": null, "keywords": [], "section": "causal-discovery", "topicId": "causal-mechanisms", "addedAt": "2026-10-12" }, { "id": "arxiv-2606.26467", "title": "A Causal Foundation Model for Structure and Outcome Prediction", "date": "2026-06-25", "updated": "2026-06-25", "category": 3, "related": [ 2, 5 ], "topic": "结构发现、干预与反事实联合建模", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "TabPFN-CFM 联合预测结构与结果，可输入已知图信息并区分观测、干预、反事实查询。在合成与函数形式偏移任务上对照有图／无图输入；Amazon 与法学院数据的反事实答案由 DoWhy 生成。真实数据上的反事实评测仍依赖设定图与生成模型，不能当作实际干预真值或任意图识别的验证。", "url": "https://arxiv.org/abs/2606.26467", "pdf": "https://arxiv.org/pdf/2606.26467", "code": null, "keywords": [ "TabPFN" ], "section": "causal-models", "topicId": "causal-unified", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.15133", "title": "Causal Foundation Models with Continuous Treatments", "date": "2026-05-14", "updated": "2026-07-13", "category": 3, "related": [ 2, 5 ], "topic": "连续处理与剂量策略", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "CCPFN 通过连续处理先验与专门处理编码器预测个体处理响应曲线。在合成和半合成连续处理任务中，以五折比较响应曲线误差和决策误差。无混杂与正值性是核心前提；连续处理的大规模真实干预基准不足，不能据此确认现实剂量决策效果。", "url": "https://arxiv.org/abs/2605.15133", "pdf": "https://arxiv.org/pdf/2605.15133", "code": "https://github.com/layer6ai-labs/CCPFN-inference", "keywords": [], "section": "causal-estimation", "topicId": "continuous-iv", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.05993", "title": "TabCF: Distributional Control Function Estimation with Tabular Foundation Models", "date": "2026-05-07", "updated": "2026-05-07", "category": 3, "related": [ 2, 5 ], "topic": "工具变量与干预分布估计", "status": "preprint", "venue": "预印本", "summary": "TabCF 用 TFM 估计控制函数和条件结果分布，结合工具变量推断干预均值与分位数。合成实验重复 100 个种子，另展示四个经济学工具变量数据的响应曲线。真实案例主要验证经济解释的一致性，缺少干预真值；排除限制等工具变量假设不可由模型自动保证，未处理无效工具变量。", "url": "https://arxiv.org/abs/2605.05993", "pdf": "https://arxiv.org/pdf/2605.05993", "code": "https://github.com/GepingChen/TabCF", "keywords": [], "section": "causal-estimation", "topicId": "continuous-iv", "addedAt": "2026-10-12" }, { "id": "arxiv-2603.15928", "title": "Prior-data Fitted Networks for Causal Inference: a Simulation Study with Real-world Scenarios", "date": "2026-03-16", "updated": "2026-04-13", "category": 3, "related": [ 1, 2 ], "topic": "因果估计可用性与误差诊断", "status": "preprint", "venue": "预印本", "summary": "用两个临床数据背景构造模拟场景，比较 TabPFN 因果包装与 CausalPFN 的 ATE 偏差、覆盖和成本。按经验条件分布生成已知真值，发现部分估计器偏差与区间欠覆盖，并报告随样本增长的耗时。模拟只覆盖特定二元处理、二元结果与分类混杂设置；计算负担应归于所测完整估计流程，不能等同单次 TabPFN 推理。", "url": "https://arxiv.org/abs/2603.15928", "pdf": "https://arxiv.org/pdf/2603.15928", "code": "https://github.com/franciscomourao/PFNs_Causal_Inference", "keywords": [ "TabPFN", "CausalPFN" ], "section": "causal-models", "topicId": "causal-reliability", "addedAt": "2026-10-12" }, { "id": "arxiv-2603.12037", "title": "Frequentist Consistency of Prior-Data Fitted Networks for Causal Inference", "date": "2026-03-12", "updated": "2026-07-22", "category": 3, "related": [ 2, 5 ], "topic": "因果效应去偏与区间校准", "status": "published", "venue": "ICML 2026", "summary": "MP-OSPC 用鞅后验抽取辅助函数，再用影响函数的一步修正缓解 PFN 的 ATE 先验偏差。合成和半合成数据检验后验集中、分布距离与可信区间校准，并区分辅助估计集和最终估计集。渐近结论依赖 PFN 的集中条件，实测大样本下余项会反弹；不能概括成已证明现有 PFN 无条件一致。", "url": "https://arxiv.org/abs/2603.12037", "pdf": "https://arxiv.org/pdf/2603.12037", "code": "https://github.com/Valentyn1997/freq-cons-pfns", "keywords": [], "section": "causal-models", "topicId": "causal-reliability", "sources": [ "https://arxiv.org/abs/2603.12037", "https://proceedings.mlr.press/v306/melnychuk26a.html" ], "addedAt": "2026-10-12" }, { "id": "paper-8380790f74", "title": "Improving TabPFN's Synthetic Data Generation by Integrating Causal Structure", "date": "2026-03-10", "updated": "2026-07-09", "category": 3, "related": [ 5 ], "topic": "因果约束的数据生成与推理", "status": "published", "venue": "UAI 2026", "summary": "按完整或部分因果图限制 TabPFN 自回归生成的条件变量，减少生成顺序造成的依赖失真。比较生成数据的依赖保真与处理效应保留，区分已知图、真实部分图和从数据发现的图。已知图条件下的改善不代表自动发现图同样有效；原文报告发现图没有稳定增益，分布相似度也不构成隐私保证。", "url": "https://proceedings.mlr.press/v337/tugnoli26a.html", "pdf": "https://raw.githubusercontent.com/mlresearch/v337/main/assets/tugnoli26a/tugnoli26a.pdf", "code": "https://github.com/DavideTugnoli/tabpfn-causal-synthetic", "keywords": [ "TabPFN" ], "section": "causal-models", "topicId": "causal-knowledge", "addedAt": "2026-10-12" }, { "id": "paper-a15b9c81ae", "title": "Use What You Know: Causal Foundation Models with Partial Graphs", "date": "2026-02-16", "updated": "2026-09-27", "category": 3, "related": [ 5 ], "topic": "不完整因果知识注入", "status": "published", "venue": "ICML 2026", "summary": "Graph4CFM 用已知祖先、非祖先和未知关系编码部分图，通过注意力偏置与图表示注入因果先验。从不可仅凭观测数据识别的合成案例出发，对照等结构无图模型并扩展至复杂和半合成基准。实验假设因果充分性、无隐藏变量；精确贝叶斯后验的性质不自动传给神经网络近似。", "url": "https://proceedings.mlr.press/v306/reuter26a.html", "pdf": "https://raw.githubusercontent.com/mlresearch/v306/main/assets/reuter26a/reuter26a.pdf", "code": "https://github.com/ArikReuter/Graphs4CausalFoundationModels", "keywords": [], "section": "causal-models", "topicId": "causal-knowledge", "sources": [ "https://proceedings.mlr.press/v306/reuter26a.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2601.04110", "title": "Causal Data Augmentation for Robust Fine-Tuning of Tabular Foundation Models", "date": "2025", "updated": "2025", "category": 3, "related": [ 6 ], "topic": "因果数据增强辅助微调", "status": "workshop", "venue": "EurIPS 2025 Tabular Data Workshop", "summary": "CausalMixFT 在目标数据拟合结构因果模型，再混合真实与生成样本微调 Mitra。在 TabArena 的 33 个分类任务、十折上比较生成器与微调策略，分析验证和测试差距。结果基于所选 Mitra 与低样本任务，归一化 ROC-AUC 改善不能当作原始百分点；学习到的图不保证真实因果结构。", "url": "https://arxiv.org/abs/2601.04110", "pdf": "https://arxiv.org/pdf/2601.04110", "code": null, "keywords": [], "dateNote": "2025 Workshop 已公开；arXiv 于 2026-01-07 提交", "section": "causal-models", "topicId": "causal-adaptation", "recognition": { "label": "Oral", "source": "https://sites.google.com/view/eurips25-ai-td/home", "verifiedAt": "2026-10-12" }, "addedAt": "2026-10-12" }, { "id": "arxiv-2511.07236", "title": "Does TabPFN Understand Causal Structures?", "date": "2025-11-10", "updated": "2025-11-10", "category": 3, "related": [ 6 ], "topic": "冻结表格模型的因果适配器", "status": "workshop", "venue": "EurIPS 2025 Tabular Data Workshop", "summary": "冻结 TabPFNv2 编码器，训练因果 token 与双注意力解码器从中间表示预测邻接矩阵。在 4–20 变量的合成图任务训练，按层数和编码器消融，评估观测加干预样本上的边预测。仍需数百万可训练参数及因果监督；小图合成实验不能证明原版 TabPFN 已具备通用因果理解或识别能力。", "url": "https://arxiv.org/abs/2511.07236", "pdf": "https://arxiv.org/pdf/2511.07236", "code": null, "keywords": [ "TabPFN" ], "section": "causal-models", "topicId": "causal-adaptation", "addedAt": "2026-10-12" }, { "id": "paper-9fd1ee8bbf", "title": "Foundation Models for Causal Inference via Prior-Data Fitted Networks", "date": "2025-06-12", "updated": "2025-06-12", "category": 3, "related": [ 5 ], "topic": "可识别因果估计与模型封装", "status": "published", "venue": "ICLR 2026", "summary": "CausalFM 把可识别的因果设定写入 PFN 先验，分别构建后门、工具变量与前门估计。提供先验条件分析，在合成及 Jobs 半合成数据上比较对应识别设定的效应误差与推理成本。评估依赖已知生成机制；需要匹配相应识别前提，不能把不同先验的成绩合成一个任意场景通用的因果模型。", "url": "https://proceedings.iclr.cc/paper_files/paper/2026/hash/7f8cabaf2de70e1a9d3eb187f02bd58c-Abstract-Conference.html", "pdf": "https://proceedings.iclr.cc/paper_files/paper/2026/file/7f8cabaf2de70e1a9d3eb187f02bd58c-Paper-Conference.pdf", "code": "https://github.com/yccm/CausalFM-toolkit", "keywords": [ "CausalFM" ], "section": "causal-estimation", "topicId": "treatment-effects", "addedAt": "2026-10-12" }, { "id": "paper-e5eafe3ccc", "title": "CausalPFN: Amortized Causal Effect Estimation via In-Context Learning", "date": "2025-06-09", "updated": "2025-06-09", "category": 3, "related": [ 2, 5 ], "topic": "ATE/CATE与增益建模基础模型", "status": "published", "venue": "NeurIPS 2025", "summary": "CausalPFN 直接学习潜在结果条件期望的后验，借助上下文估计平均和异质处理效应。比较调参后的因果估计器，并分析有真值任务上的效应误差与不确定性。实现聚焦二元处理与强可忽略性；理论依赖适当先验和渐近样本，尚无一般有限样本保证，大表性能也会下降。", "url": "https://proceedings.neurips.cc/paper_files/paper/2025/hash/e3d3db07c1bfb63e1d0b998996de1d12-Abstract-Conference.html", "pdf": "https://proceedings.neurips.cc/paper_files/paper/2025/file/e3d3db07c1bfb63e1d0b998996de1d12-Paper-Conference.pdf", "code": "https://github.com/vdblm/CausalPFN", "keywords": [ "CausalPFN" ], "section": "causal-estimation", "topicId": "treatment-effects", "addedAt": "2026-10-12" }, { "id": "paper-93cf6a538b", "title": "Do-PFN: In-Context Learning for Causal Effect Estimation", "date": "2025-06-06", "updated": "2025-06-06", "category": 3, "related": [ 2, 5 ], "topic": "条件干预分布预测", "status": "published", "venue": "NeurIPS 2025", "summary": "Do-PFN 将观测表格和干预查询映射为条件干预结果分布，对先验内多种机制进行平均。在后门、前门等合成案例和半合成数据上比较结果分布、效应估计、先验外迁移与校准。缺少图输入不等于无需因果假设；不可识别时答案依赖先验，真实场景的先验失配仍待系统验证。", "url": "https://proceedings.neurips.cc/paper_files/paper/2025/hash/ff997469ac66cf893c4183efeb22212a-Abstract-Conference.html", "pdf": "https://proceedings.neurips.cc/paper_files/paper/2025/file/ff997469ac66cf893c4183efeb22212a-Paper-Conference.pdf", "code": "https://github.com/jr2021/Do-PFN", "keywords": [ "Do-PFN" ], "section": "causal-estimation", "topicId": "treatment-effects", "addedAt": "2026-10-12" }, { "id": "arxiv-2205.12934", "title": "Amortized Inference for Causal Structure Learning", "date": "2022-05-25", "updated": "2022-05-25", "category": 3, "related": [ 5 ], "topic": "跨数据集因果发现的基础方法", "status": "published", "venue": "NeurIPS 2022", "summary": "AVICI 用样本置换不变、变量置换等变的网络摊销因果图后验推断。在结构方程与半合成基因调控数据中主动改变图、机制、噪声和规模，比较结构及干预距离。泛化仍取决于模拟器覆盖，观测图恢复受识别条件限制；有限真实蛋白数据不代表所有领域。", "url": "https://arxiv.org/abs/2205.12934", "pdf": "https://arxiv.org/pdf/2205.12934", "code": "https://github.com/larslorch/avici", "keywords": [], "section": "causal-discovery", "topicId": "causal-graphs", "addedAt": "2026-10-12" }, { "id": "paper-dee7306e6b", "title": "Policy Learning With Observational Data", "date": "2017-02-09", "updated": "2017-02-09", "category": 3, "related": [ 2 ], "topic": "由因果估计到受约束策略", "status": "published", "venue": "Econometrica 2021", "summary": "将双重稳健效应评分用于约束策略类中的干预分配，分开辅助函数估计和策略优化。在明确识别与估计条件下给出遗憾界，并以 GAIN 项目及内生处理模拟展示方法。依赖识别、重叠、辅助估计质量及策略复杂度条件；离线价值最大化不是现实部署收益的自动保证。", "url": "https://doi.org/10.3982/ECTA15732", "pdf": "https://arxiv.org/pdf/1702.02896", "code": "https://github.com/grf-labs/policytree/tree/master/experiments/athey_wager_2021", "keywords": [], "section": "causal-estimation", "topicId": "causal-policy", "addedAt": "2026-10-12" }, { "id": "paper-af13c15dd9", "title": "Double/debiased machine learning for treatment and structural parameters", "date": "2016-07-30", "updated": "2016-07-30", "category": 3, "related": [ 2 ], "topic": "机器学习估计器的因果去偏", "status": "published", "venue": "The Econometrics Journal 2018", "summary": "DML 结合 Neyman 正交得分和交叉拟合，降低辅助预测模型的误差对低维因果参数推断的影响。给出收敛与推断条件，展示失业保险、401(k) 和制度工具变量三个经验案例。正交化不修复错误识别假设，仍要求辅助函数收敛、适当矩条件及重叠；不能把任意预测器直接包装成可靠因果估计。", "url": "https://doi.org/10.1111/ectj.12097", "pdf": "https://arxiv.org/pdf/1608.00060", "code": null, "keywords": [], "section": "causal-models", "topicId": "causal-reliability", "addedAt": "2026-10-12" }, { "id": "arxiv-2610.01725", "title": "RelICL: Training-free Relational Learning with Tabular Foundation Models", "date": "2026-10-01", "updated": "2026-10-01", "category": 4, "section": "semantic-integration", "topicId": "data-integration", "related": [ 1, 5 ], "topic": "端到端数据整合", "status": "preprint", "venue": "预印本", "summary": "RelICL 沿数据库模式图抽取并融合行表示，用冻结 TFM 完成跨表特征构建与预测。在 RelBench v2 的七个数据库、21 个实体分类／回归任务上评估，并更换 TabICL 与 TabPFN 主干。基线数字引用各自论文，调参与种子未统一复现；未覆盖推荐等其他任务，配对排名还取各模型共有任务。", "url": "https://arxiv.org/abs/2610.01725", "pdf": "https://arxiv.org/pdf/2610.01725", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "RelICL" ], "sources": [ "https://arxiv.org/abs/2610.01725" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.23574", "title": "PACE: Plug-and-Play Contextual Embedding for Feature Screening with Pretrained Tabular Foundation Models", "date": "2026-09-20", "updated": "2026-09-20", "category": 4, "section": "semantic-features", "topicId": "feature-search", "related": [ 2, 5 ], "topic": "任务反馈与特征搜索", "status": "preprint", "venue": "预印本", "summary": "PACE 将冻结列编码器作为特征筛选的表示适配器，再交给已有相关性或依赖评分函数。控制原始／嵌入表示的比较预算，在合成支持恢复、加噪 TALENT 与自然高维表上评估下游效果。列内注意力使样本表示产生依赖，尚无一般 sure-screening 保证；核或距离评分仍可能具有样本二次成本。", "url": "https://arxiv.org/abs/2609.23574", "pdf": "https://arxiv.org/pdf/2609.23574", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "PACE", "PACE-DC" ], "sources": [ "https://arxiv.org/abs/2609.23574" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.21894", "title": "LLMs as Feature Engineers for Text-and-Tabular Prediction", "date": "2026-09-18", "updated": "2026-09-18", "category": 4, "related": [ 2 ], "topic": "语义信息结构化与任务驱动特征生成", "status": "preprint", "venue": "预印本", "summary": "LLM 定义有取值约束的语义特征并逐行抽取，利用预测错误反馈迭代特征搜索。在众筹、书评和问答三个英语文本加表格任务上进行交叉验证与特征选择比较。任务范围较窄且依赖闭源 API；迭代搜索的交叉验证收益仍需独立外层测试确认，新增样本也需支付特征抽取成本。", "url": "https://arxiv.org/abs/2609.21894", "pdf": "https://arxiv.org/pdf/2609.21894", "code": null, "keywords": [], "section": "semantic-features", "topicId": "text-structure", "addedAt": "2026-10-12" }, { "id": "arxiv-2609.03529", "title": "KnowFeat: Knowledge-Guided Feature Engineering via LLM Agents", "date": "2026-09-03", "updated": "2026-09-03", "category": 4, "related": [ 1 ], "topic": "语义信息结构化与任务驱动特征生成", "status": "preprint", "venue": "预印本", "summary": "KnowFeat 将领域知识注入 LLM 特征工程，经代码执行、统计质量与模型效果三层筛选并保存来源。十二个公开分类集与两个反洗钱任务保留独立测试集，补充跨预测器和重复运行分析。缺失基线按可用方法排名，比较并不完全齐全；知识整理需要人工，收益受字段语义和原始特征空间影响。", "url": "https://arxiv.org/abs/2609.03529", "pdf": "https://arxiv.org/pdf/2609.03529", "code": null, "keywords": [], "section": "semantic-features", "topicId": "knowledge-features", "addedAt": "2026-10-12" }, { "id": "paper-20e0758264", "title": "CaRL-EM: Cost-Aware Reinforcement Learning for Entity Matching with LLMs", "date": "2026-07", "updated": "2026-07", "category": 4, "related": [ 1 ], "topic": "表格字段、实体对齐与跨表整合", "status": "published", "venue": "ACL 2026", "summary": "CaRL-EM 用强化学习选择实体匹配操作及模型大小，在候选集中逐步匹配或停止。在七个实体匹配基准比较 F1 与成本，并检验更换 LLM 后的控制策略迁移。仅覆盖最多一个真匹配、top-10 候选的 clean-clean 设置；训练用两档抽象成本，论文节省比例不等于任意 API 账单降幅。", "url": "https://aclanthology.org/2026.acl-long.1258/", "pdf": "https://aclanthology.org/2026.acl-long.1258.pdf", "code": null, "keywords": [], "section": "semantic-integration", "topicId": "entity-matching", "addedAt": "2026-10-12" }, { "id": "arxiv-2608.03565", "title": "Enhancing Tabular Learners with Context-Aware Semantic Embeddings", "date": "2026-07", "updated": "2026-10-09", "category": 4, "related": [ 1, 5 ], "topic": "语义与表格特征融合", "status": "published", "venue": "NeurIPS 2026", "summary": "CASE 用有标签上下文预填表格语言模型的 KV 缓存，抽取任务相关行嵌入，再压缩并融合原始列。在 CARTE、TextTab 与单折 TabArena 上比较多种下游预测器，并做上下文与语义消融。依赖额外训练的表格语言模型和嵌入成本；单折结果及特定语义任务不能代表完整 TabArena 协议。", "url": "https://arxiv.org/abs/2608.03565", "pdf": "https://arxiv.org/pdf/2608.03565", "code": "https://github.com/SAP-samples/case", "keywords": [ "TabICL" ], "dateNote": "日期按 2026 年 7 月 Workshop 版本", "section": "semantic-fusion", "topicId": "semantic-embeddings", "sources": [ "https://arxiv.org/abs/2608.03565", "https://neurips.cc/virtual/2026/poster/149526", "https://openreview.net/forum?id=pV2wVDrTqK" ], "aliases": [ "CASE", "Context-Aware Semantic Embeddings" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2607.24492", "title": "SINT-Flow: Schema Integration using Large Language Model Workflows", "date": "2026-07-27", "updated": "2026-07-27", "category": 4, "related": [], "topic": "表格字段、实体对齐与跨表整合", "status": "preprint", "venue": "预印本", "summary": "SINT-Flow 用表拆分、模式匹配、分组和属性合并等 LLM 操作形成统一数据库模式。SINT-Bench 包含十个整合任务、93 张表，提供逐操作真值并比较自一致性、复核循环与耗时。基准最多三类实体，拆分采用至少三列的启发式；大型异构库和复杂约束尚需验证。", "url": "https://arxiv.org/abs/2607.24492", "pdf": "https://arxiv.org/pdf/2607.24492", "code": "https://github.com/wbsg-uni-mannheim/SINT-Flow", "keywords": [], "section": "semantic-integration", "topicId": "schema-matching", "addedAt": "2026-10-12" }, { "id": "arxiv-2607.01548", "title": "Evolutionary Feature Engineering for Structured Data", "date": "2026-07-02", "updated": "2026-07-02", "category": 4, "related": [ 5 ], "topic": "语义信息结构化与任务驱动特征生成", "status": "published", "venue": "NeurIPS 2026", "summary": "EFE 让 LLM 演化 fit/transform 特征程序，用验证收益和特征复杂度共同选择。分别评估时间序列归一化和表格特征工程，表格侧对照决策树、LightGBM 与 TabPFN。表格改善主要体现在简单模型，TabPFN 上各方法平均排名接近；搜索耗时与反复验证成本应计入完整流程。", "url": "https://arxiv.org/abs/2607.01548", "pdf": "https://arxiv.org/pdf/2607.01548", "code": "https://github.com/egetaga/EFE", "keywords": [], "section": "semantic-features", "topicId": "feature-search", "sources": [ "https://arxiv.org/abs/2607.01548", "https://neurips.cc/virtual/2026/poster/150446", "https://openreview.net/forum?id=hz6wl9NrRG" ], "aliases": [ "EFE", "EFE-Tab", "EFE-Time" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2606.30371", "title": "MaDI-Bench: An End-to-End Data Integration Benchmark", "date": "2026-06-29", "updated": "2026-10-05", "category": 4, "related": [ 1 ], "topic": "表格字段、实体对齐与跨表整合", "status": "preprint", "venue": "预印本", "summary": "MaDI-Bench 将模式匹配、值归一、实体匹配和数据融合放入同一端到端评测。五个基础领域同时报告步骤指标与最终整合结果，并提供可控扰动变体。部分端到端指标使用人工流程输出作为银标，变体正例池也包含模型预测；应区分人工真值、银标和无参考指标。", "url": "https://arxiv.org/abs/2606.30371", "pdf": "https://arxiv.org/pdf/2606.30371", "code": "https://github.com/wbsg-uni-mannheim/MaDI-Bench", "keywords": [], "section": "semantic-integration", "topicId": "data-integration", "addedAt": "2026-10-12" }, { "id": "arxiv-2606.08800", "title": "Bridging Expert Knowledge and Automated Feature Engineering via Self-Evolution", "date": "2026-06-07", "updated": "2026-06-07", "category": 4, "related": [ 1 ], "topic": "语义信息结构化与任务驱动特征生成", "status": "preprint", "venue": "预印本", "summary": "FEST 从多模态样本与专家规范生成语义和确定性特征，再通过去重与树反馈迭代。比较品牌等二分类任务，并以 LLM 评分与人工专家研究评估特征对规范的覆盖。60–80% 专家覆盖率来自指定 LLM 评审协议，不等于全部人工验证；尚未覆盖多分类与回归，生成与抽取有额外成本。", "url": "https://arxiv.org/abs/2606.08800", "pdf": "https://arxiv.org/pdf/2606.08800", "code": null, "keywords": [], "section": "semantic-features", "topicId": "knowledge-features", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.12411", "title": "Predicting Decisions of AI Agents from Limited Interaction through Text-Tabular Modeling", "date": "2026-05-12", "updated": "2026-05-12", "category": 4, "related": [ 2, 5 ], "topic": "语义、数值与类别特征融合", "status": "preprint", "venue": "预印本", "summary": "把对话隐藏状态、语义嵌入和游戏数值状态组成表格，用 TabPFN 适配少量交互中的 Agent 行为。从 13 个模型的竞赛群体迁移到未见黑客松 Agent，比较 0–16 局适配及接收／报价预测。结论针对谈判游戏与所测群体；表格模型同时使用来源群体标签，和直接提示 LLM 的信息利用方式不同。", "url": "https://arxiv.org/abs/2605.12411", "pdf": "https://arxiv.org/pdf/2605.12411", "code": null, "keywords": [ "TabPFN" ], "section": "semantic-fusion", "topicId": "semantic-embeddings", "addedAt": "2026-10-12" }, { "id": "arxiv-2604.11835", "title": "Schema-Adaptive Tabular Representation Learning with LLMs for Generalizable Multimodal Clinical Reasoning", "date": "2026-04-12", "updated": "2026-04-12", "category": 4, "related": [ 1 ], "topic": "语义与表格特征融合", "status": "preprint", "venue": "预印本", "summary": "将字段名和值转成文本，用语义嵌入构造可跨模式迁移的表格表示，并与 MRI 表示融合。在 NACC 训练、ADNI 外部数据零样本测试，另评估多标签痴呆分类与表示消融。实现依赖 text-embedding-3-large，属于嵌入建模；MRI 子集较小，字段名语义和单一临床领域限制泛化，不能等同临床有效性验证。", "url": "https://arxiv.org/abs/2604.11835", "pdf": "https://arxiv.org/pdf/2604.11835", "code": null, "keywords": [], "section": "semantic-fusion", "topicId": "multimodal-schema", "addedAt": "2026-10-12" }, { "id": "arxiv-2506.10707", "title": "ConTextTab: A Semantics-Aware Tabular In-Context Learner", "date": "2025-06-12", "updated": "2025-06-12", "category": 4, "related": [ 1, 5 ], "topic": "语义与表格特征融合", "status": "published", "venue": "NeurIPS 2025", "summary": "ConTextTab 为文本、日期和数值使用不同编码，在表格原生 ICL 中联合语义与统计信息。203 个分类／回归任务采用固定测试切分，重点检查 CARTE 语义任务及预训练数据污染。未进行跨折验证，污染检索不能完全排除重合；在大表上增加上下文或集成仍不能解决全部性能差距。", "url": "https://arxiv.org/abs/2506.10707", "pdf": "https://arxiv.org/pdf/2506.10707", "code": "https://github.com/SAP-samples/sap-rpt-1-oss", "keywords": [ "ConTextTab", "TFM" ], "section": "semantic-fusion", "topicId": "text-tabular-models", "recognition": { "label": "Spotlight", "source": "https://arxiv.org/abs/2506.10707", "verifiedAt": "2026-10-12" }, "aliases": [ "ConTextTab", "SAP-RPT-1-OSS", "SAP RPT 1 OSS" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2505.18125", "title": "TabSTAR: A Tabular Foundation Model for Tabular Data with Text Fields", "date": "2025-05-23", "updated": "2025-05-23", "category": 4, "related": [ 1, 6 ], "topic": "语义与表格特征融合", "status": "published", "venue": "NeurIPS 2025", "summary": "TabSTAR 联合可更新文本编码器、目标 token 与表格结构，为含文本列的任务学习目标相关表示。跨表预训练后对每个目标任务做 LoRA 微调，并在十次切分上比较语义表格基线。不是冻结模型的一次上下文预测；部分基线因资源或类别限制缺失，归一化成绩不能直接作原始准确率提升。", "url": "https://arxiv.org/abs/2505.18125", "pdf": "https://arxiv.org/pdf/2505.18125", "code": "https://github.com/alanarazi7/TabSTAR", "keywords": [ "TabSTAR", "TFM" ], "section": "semantic-fusion", "topicId": "text-tabular-models", "aliases": [], "addedAt": "2026-10-12" }, { "id": "arxiv-2505.14415", "title": "Table Foundation Models: on knowledge pre-training for tabular learning", "date": "2025-05-20", "updated": "2025-06-30", "category": 4, "section": "semantic-fusion", "topicId": "semantic-embeddings", "related": [ 1, 6 ], "topic": "语义嵌入与特征拼接", "status": "published", "venue": "TMLR 2025", "summary": "TARTE 将字段语义变成可复用表格表示，可冻结接预测器，也可微调或用于残差提升。在 51 个富语义表格任务上比较表示复用、单表学习与跨表迁移。跨表比较中 TARTE 复用已微调表示，而基线联合拟合来源与目标表，访问条件不同；结果偏向有语义字段的任务。", "url": "https://arxiv.org/abs/2505.14415", "pdf": "https://arxiv.org/pdf/2505.14415", "code": "https://github.com/soda-inria/tarte-ai", "keywords": [ "TFM", "TARTE" ], "aliases": [ "TARTE" ], "sources": [ "https://arxiv.org/abs/2505.14415", "https://openreview.net/forum?id=QV4P8Csw17" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2412.08194", "title": "Magneto: Combining Small and Large Language Models for Schema Matching", "date": "2024-12-11", "updated": "2024-12-11", "category": 4, "related": [], "topic": "表格字段、实体对齐与跨表整合", "status": "published", "venue": "PVLDB 18(8)，2025", "summary": "Magneto 用小语言模型召回匹配列，再由大语言模型重排，并以合成同义列增强召回模型。在六组人工标注或构造的模式匹配数据上评估准确率、运行时间与组件消融。正确列未进入 top-k 时重排无法补救；合成同义列和提示方式影响迁移，LLM 重排费用仍需计入。", "url": "https://arxiv.org/abs/2412.08194", "pdf": "https://arxiv.org/pdf/2412.08194", "code": "https://github.com/VIDA-NYU/magneto-matcher", "keywords": [], "section": "semantic-integration", "topicId": "schema-matching", "addedAt": "2026-10-12" }, { "id": "arxiv-2402.16785", "title": "CARTE: Pretraining and Transfer for Tabular Learning", "date": "2024-02-26", "updated": "2024-05-31", "category": 4, "section": "semantic-fusion", "topicId": "text-tabular-models", "related": [ 1, 6 ], "topic": "文本与表格联合建模", "status": "published", "venue": "ICML 2024", "summary": "CARTE 把列名与单元格组织为带边语义的行图，支持不同字段表格间的迁移和联合学习。在 51 个语义丰富的表格任务上比较单表、跨表及预处理消融，使用目标任务微调与集成。优势与字符串语义有关，不能外推到匿名数值表；集成和微调成本应与最终成绩一起比较。", "url": "https://arxiv.org/abs/2402.16785", "pdf": "https://arxiv.org/pdf/2402.16785", "code": null, "keywords": [ "TFM", "CARTE" ], "aliases": [ "CARTE" ], "sources": [ "https://arxiv.org/abs/2402.16785", "https://proceedings.mlr.press/v235/kim24d.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2305.03403", "title": "Large Language Models for Automated Data Science: Introducing CAAFE for Context-Aware Automated Feature Engineering", "date": "2023-05-05", "updated": "2023-05-05", "category": 4, "related": [ 5 ], "topic": "语义信息结构化与任务驱动特征生成", "status": "published", "venue": "NeurIPS 2023", "summary": "CAAFE 让 LLM 根据数据说明生成可解释 Python 特征代码，并以验证表现反馈迭代。用 TabPFN 选择特征，再在逻辑回归、随机森林与 TabPFN 上比较传统自动特征工程。特征选择偏向所用评价器，部分昂贵基线只生成一次特征供其他模型使用；描述质量、API 与验证集过拟合仍需检查。", "url": "https://arxiv.org/abs/2305.03403", "pdf": "https://arxiv.org/pdf/2305.03403", "code": "https://github.com/noahho/CAAFE", "keywords": [], "section": "semantic-features", "topicId": "feature-search", "addedAt": "2026-10-12" }, { "id": "paper-62f5871d79", "title": "Text-to-Table: A New Way of Information Extraction", "date": "2021-09-06", "updated": "2021-09-06", "category": 4, "related": [], "topic": "语义信息结构化与任务驱动特征生成", "status": "published", "venue": "ACL 2022", "summary": "Text-to-Table 将文档信息抽取写成表格序列生成，加入表格格式约束与关系嵌入。将四个数据到文本数据集反向处理并过滤未在文本出现的表格内容，对比 BART 与 BERT 抽取基线。任务依赖监督训练与训练集中隐含模式；大表、实体别名、背景知识和隐式推理仍是失败来源，不是任意文本自动建库保证。", "url": "https://aclanthology.org/2022.acl-long.180/", "pdf": "https://aclanthology.org/2022.acl-long.180.pdf", "code": "https://github.com/shirley-wu/text_to_table", "keywords": [], "section": "semantic-features", "topicId": "text-structure", "addedAt": "2026-10-12" }, { "id": "arxiv-2610.13253", "title": "Tabular Foundation Models Are Effectively Shallow", "date": "2026-10-09", "updated": "2026-10-09", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 6 ], "topic": "模型机制与行为分析", "status": "preprint", "venue": "预印本", "summary": "用无标签校准的仿射映射替换连续 Transformer 层，测量多类 TFM 的可压缩深度。八个模型、171 个分类任务及回归实验比较半深度和单块配置，并测量 GPU 与缓存推理。极端压缩与半深度的精度损失不同；映射需校准、宽表开销增加，浅层代理实验不证明预训练深度无用。", "url": "https://arxiv.org/abs/2610.13253", "pdf": "https://arxiv.org/pdf/2610.13253", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "Affine Translator" ], "sources": [ "https://arxiv.org/abs/2610.13253" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.11784", "title": "Compile the Table: Query-Calibrated Operator Compression for Tabular In-Context Learning", "date": "2026-10-08", "updated": "2026-10-08", "category": 5, "section": "inference-efficiency", "topicId": "context-compression", "related": [ 6 ], "topic": "上下文压缩与信息迁移", "status": "preprint", "venue": "预印本", "summary": "QCOC 联合聚类 KV 原型并校准取值，把固定上下文压缩为可复用的查询记忆。以单估计器 TabICLv2 在 64 个 CC18 任务及长表上比较同保留量压缩、检索和服务成本。服务提速不含一次性完整缓存与编译，需多次查询摊销；Covertype 上局部检索明显更好。", "url": "https://arxiv.org/abs/2610.11784", "pdf": "https://arxiv.org/pdf/2610.11784", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "QCOC", "Query-Calibrated Operator Compression" ], "sources": [ "https://arxiv.org/abs/2610.11784" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.12005", "title": "Test-Time Compute for Tabular Foundation Models: Mechanisms, Gains, and Limits", "date": "2026-10-08", "updated": "2026-10-08", "category": 5, "related": [ 2, 6 ], "topic": "表格推理策略", "status": "preprint", "venue": "预印本", "summary": "拆分参数适配、预测聚合和上下文构建，并以 DiagScale 只调整注意力相似度中的少量缩放参数。在 51 个 TabArena 任务比较三个主干的适配，再分析不同聚合与选样组合。聚合和上下文实验主要基于 TabPFN-3；组合方案的内层折共享已训练适配器，集成选择分数应谨慎解释，仍以外层测试为准。", "url": "https://arxiv.org/abs/2610.12005", "pdf": "https://arxiv.org/pdf/2610.12005", "code": "https://github.com/kanghui-learning/test-time-compute-for-tabular-foundation-models", "keywords": [], "section": "inference-strategies", "topicId": "test-time-compute", "addedAt": "2026-10-12" }, { "id": "arxiv-2610.07996", "title": "TICDA: Tabular In-Context Data Attribution", "date": "2026-10-06", "updated": "2026-10-06", "category": 5, "section": "context-selection", "topicId": "representative-context", "related": [ 2 ], "topic": "代表性选择与上下文优化", "status": "preprint", "venue": "预印本", "summary": "TICDA 在冻结 TFM 的隐表示上拟合替代模型，近似上下文样本的删除影响。在 38 个 TabArena 分类任务比较穷举留一法、20% 人工标签错误、上下文精简和主动学习。近似假设删一行不改变其他表示；忠实度只验证到 4096 行上下文，人工错标不代表所有真实标签噪声。", "url": "https://arxiv.org/abs/2610.07996", "pdf": "https://arxiv.org/pdf/2610.07996", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "TICDA" ], "sources": [ "https://arxiv.org/abs/2610.07996" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.08314", "title": "The Standardization Trap: Certifying Joint Label Processing in Tabular Foundation Models", "date": "2026-10-06", "updated": "2026-10-06", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 2 ], "topic": "模型机制与行为分析", "status": "preprint", "venue": "预印本", "summary": "在标签标准化流形上构造检验，区分真正联合标签处理与预处理造成的表观非线性。对五种 TFM、四个回归任务、256 行上下文测量证书，并通过注意力等消融定位交互来源。覆盖任务与上下文较小；检出联合交互不等于解释完整算法，也不证明这种交互导致预测优势。", "url": "https://arxiv.org/abs/2610.08314", "pdf": "https://arxiv.org/pdf/2610.08314", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "Standardization Trap", "Joint Label Processing" ], "sources": [ "https://arxiv.org/abs/2610.08314" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.06679", "title": "Closing the Context Gap: Activation Alignment for Tabular In-Context Learning", "date": "2026-10-05", "updated": "2026-10-05", "category": 5, "related": [ 6 ], "topic": "大表推理效率", "status": "preprint", "venue": "预印本", "summary": "用完整上下文教师和合成无标签查询训练轻量激活对齐器，补偿短上下文学生的信息损失。在 38 个 TabArena 分类集、两种主干和五次切分中比较不同上下文比例。训练对齐器仍需访问完整上下文教师；五个任务对两主干均无平均收益，回归与其他模型迁移未验证。", "url": "https://arxiv.org/abs/2610.06679", "pdf": "https://arxiv.org/pdf/2610.06679", "code": "https://github.com/yoel-zeldes/tabalign", "keywords": [], "section": "inference-efficiency", "topicId": "context-compression", "addedAt": "2026-10-12" }, { "id": "arxiv-2610.05352", "title": "Rethinking Tabular Foundation Models On Data Streams", "date": "2026-10-04", "updated": "2026-10-04", "category": 5, "section": "context-selection", "topicId": "stream-context", "related": [ 1, 2 ], "topic": "流式更新与记忆管理", "status": "preprint", "venue": "预印本", "summary": "以预测后才获得标签的数据流协议，比较 TFM 的记忆保留、漂移适应与延迟成本。仅纳入真实时间顺序的流，并设置多种合成漂移与标签延迟，对照传统流学习和 AutoML。TFM 用 GPU、流模型用 CPU，成本是完整系统比较；每类合成漂移对应不同生成器，不能直接归因漂移类型。", "url": "https://arxiv.org/abs/2610.05352", "pdf": "https://arxiv.org/pdf/2610.05352", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "TFM Data Streams" ], "sources": [ "https://arxiv.org/abs/2610.05352" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.37989", "title": "TabFM-Auto: Self-Evolving Pipelines for Tabular Foundation Models", "date": "2026-09-29", "updated": "2026-09-29", "category": 5, "related": [ 1, 2, 4 ], "topic": "表格推理策略", "status": "preprint", "venue": "预印本", "summary": "TabFM-Auto 用 LLM 编程代理在冻结 TabFM 外搜索清洗、特征、上下文与后处理流程。在 TabArena 和 MLE-Bench-Tabular 评估，TabArena 每数据集只在第 0 折训练集搜索，并另报告该折独立测试。其余折测试行可能出现在搜索用训练集，全部折成绩不能视为独立嵌套评估；还需计入最高六小时 GPU 搜索及 API 成本。", "url": "https://arxiv.org/abs/2609.37989", "pdf": "https://arxiv.org/pdf/2609.37989", "code": null, "keywords": [ "TabFM" ], "section": "inference-strategies", "topicId": "pipeline-search", "addedAt": "2026-10-12" }, { "id": "arxiv-2609.36337", "title": "Adapting Linear-Time Architectures for Tabular In-Context Learning", "date": "2026-09-28", "updated": "2026-09-28", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 1 ], "topic": "模型机制与行为分析", "status": "preprint", "venue": "预印本", "summary": "在统一预算下比较线性注意力与递归主干，用写入率衰减和最终状态读取改善 DeltaNet 长度泛化。匹配约 1250 万参数、训练先验和预算，并在缩减 OpenML／TabArena 及长合成上下文上分析状态漂移。主要限行级主干和最多 20 个输入特征，长上下文证据偏合成；尚不能外推到现代双轴大型 TFM。", "url": "https://arxiv.org/abs/2609.36337", "pdf": "https://arxiv.org/pdf/2609.36337", "code": "https://github.com/schnurrd/ICL-Architectures", "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "DeltaNet", "Linear-Time Tabular ICL" ], "sources": [ "https://arxiv.org/abs/2609.36337" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.31306", "title": "Benchmarking Attention for Tabular Foundation Models", "date": "2026-09-25", "updated": "2026-10-09", "category": 5, "section": "inference-efficiency", "topicId": "memory-quantization", "related": [ 1 ], "topic": "内存优化与注意力量化", "status": "published", "venue": "NeurIPS 2026 · Evaluations and Datasets Track", "summary": "针对表格行列两轴注意力，比较算子后端、张量布局和硬件形状的吞吐与显存。A100、H100、B200 上做合成张量测试，另以 H100 的 TabPFN-v2.6 推理作端到端验证。核函数速度不等同完整系统速度；端到端仅一个模型与 GPU，量化推理后端还需另查精度影响。", "url": "https://arxiv.org/abs/2609.31306", "pdf": "https://arxiv.org/pdf/2609.31306", "code": "https://github.com/SAP-samples/tabular-attention-benchmark", "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "Tabular Attention Benchmark" ], "sources": [ "https://arxiv.org/abs/2609.31306", "https://neurips.cc/virtual/2026/poster/139839", "https://openreview.net/forum?id=1MbSbrblLi" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.28208", "title": "Support-Compiled Feature Folding: More Evidence at Lower Memory Across Tabular Foundation Models", "date": "2026-09-23", "updated": "2026-09-25", "category": 5, "section": "inference-efficiency", "topicId": "context-compression", "related": [ 4 ], "topic": "上下文压缩与信息迁移", "status": "preprint", "venue": "预印本", "summary": "SCFF 把宽表分成核心与候选特征，在有限宽度下编码折叠，再检查尾部信息是否增益。18 个宽表任务、五折和六种主干分别比较精度、端到端时间与峰值显存，并报告不同适配契约。降低显存不必然加速，TabICL 两版本的中位耗时反而增加；不同主干的容量扩展与同宽比较应分开。", "url": "https://arxiv.org/abs/2609.28208", "pdf": "https://arxiv.org/pdf/2609.28208", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "SCFF", "Support-Compiled Feature Folding" ], "sources": [ "https://arxiv.org/abs/2609.28208" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2609.13031", "title": "Attention Quantization for Tabular Foundation Models", "date": "2026-09-11", "updated": "2026-09-11", "category": 5, "related": [ 1 ], "topic": "大表推理效率", "status": "preprint", "venue": "预印本", "summary": "对 TFM 注意力的 Q、K、V 做 FP8 量化，并协调支持行与查询行的量化尺度。在完整 TabArena 和长表上评估 TabPFN-v3，另补 TabICLv2、不同头维度与硬件实验。最终方案仅在超过 8192 行时启用量化，最高约 1.7 倍提速依赖硬件和表形状；不是所有表都会受益。", "url": "https://arxiv.org/abs/2609.13031", "pdf": "https://arxiv.org/pdf/2609.13031", "code": null, "keywords": [], "section": "inference-efficiency", "topicId": "memory-quantization", "addedAt": "2026-10-12" }, { "id": "arxiv-2609.07956", "title": "Streaming Hierarchical Inference with Tabular Foundation Models", "date": "2026-09-07", "updated": "2026-09-07", "category": 5, "section": "context-selection", "topicId": "stream-context", "related": [ 1 ], "topic": "流式更新与记忆管理", "status": "workshop", "venue": "ECML PKDD 2026 Streaming Continual Learning Workshop", "summary": "HINT 用边缘近邻记忆处理高置信样本，将难例及检索上下文卸载给云端 TFM。按预测后更新协议，在多个流的前两万条记录比较准确率与通信量折中。截短数据流和指定卸载策略限制覆盖；离线通信量或计算耗时不等于真实网络中的尾延迟与端云费用。", "url": "https://arxiv.org/abs/2609.07956", "pdf": "https://arxiv.org/pdf/2609.07956", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "HINT", "Streaming Hierarchical Inference" ], "sources": [ "https://arxiv.org/abs/2609.07956" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2608.17856", "title": "ARASH: Adaptive Retrieval And Shot Selection for Tabular Prediction", "date": "2026-08-18", "updated": "2026-08-18", "category": 5, "related": [], "topic": "上下文样本优化", "status": "preprint", "venue": "预印本", "summary": "ARASH 按局部聚类的邻近性与标签纯度，自适应选择检索上下文和示例数量。在 CC18 与 Combo 的固定 80/20 切分上比较 TabPFN、TabDPT 及语言模型，另用独立任务调阈值。主要使用单一切分种子，消融集中于较小 Combo；高纯度直接预测的收益依赖邻域质量与类别分布。", "url": "https://arxiv.org/abs/2608.17856", "pdf": "https://arxiv.org/pdf/2608.17856", "code": null, "keywords": [], "section": "context-selection", "topicId": "retrieval-shots", "addedAt": "2026-10-12" }, { "id": "arxiv-2608.16429", "title": "Localized TabICLv2: Scaling Tabular In-Context Learning through k-NN", "date": "2026-07", "updated": "2026-07", "category": 5, "related": [ 6 ], "topic": "上下文样本优化", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "Localized TabICLv2 在行表征空间检索近邻，微调行表征和 ICL 头以降低局部化损失。38 个 TabArena 分类任务采用三种子切分，并比较固定 k=32、完整上下文和检索基线。这是自定切分而非完整官方协议；未直接对照 LoCalPFN，检索开销可能削弱小数据或大查询批次的速度优势。", "url": "https://arxiv.org/abs/2608.16429", "pdf": "https://arxiv.org/pdf/2608.16429", "code": null, "keywords": [ "TabICL" ], "dateNote": "日期按 2026 年 7 月 Workshop 版本", "section": "context-selection", "topicId": "retrieval-shots", "addedAt": "2026-10-12" }, { "id": "arxiv-2607.27546", "title": "Memory Efficient Tabular Foundation Models", "date": "2026-07", "updated": "2026-07", "category": 5, "related": [ 1 ], "topic": "大表推理效率", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "比较 K-Means、GPTQ 与 AWQ 对 TabPFN 权重的 INT4 后训练量化。在 30 个受模型容量约束的 OpenML 任务、两个版本上评估预测表现及 BZIP2 后的模型文件大小。推理前会反量化，原文明确没有运行时显存改善；压缩文件比例不能写成推理内存或速度收益。", "url": "https://arxiv.org/abs/2607.27546", "pdf": "https://arxiv.org/pdf/2607.27546", "code": null, "keywords": [ "TabPFN" ], "dateNote": "日期按 2026 年 7 月 Workshop 版本", "section": "inference-efficiency", "topicId": "memory-quantization", "addedAt": "2026-10-12" }, { "id": "arxiv-2606.18677", "title": "Bounded Context Management for Tabular Foundation Models on Stream Learning", "date": "2026-06-17", "updated": "2026-06-17", "category": 5, "related": [ 2 ], "topic": "上下文样本优化", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "CURE 将有限上下文分成近期和长期记忆，以预测熵控制入库、同类冗余控制淘汰。七个数据流使用 1000 行预算与预测后更新协议，并在四种 TFM 上检验策略迁移。主要是即时标签协议，部分数据的行序未必代表真实时间流；准确率优势需与同主干 FIFO、记忆容量和服务成本一起判断。", "url": "https://arxiv.org/abs/2606.18677", "pdf": "https://arxiv.org/pdf/2606.18677", "code": "https://github.com/morcellinus/CURE-ICML-FMSD", "keywords": [], "section": "context-selection", "topicId": "stream-context", "recognition": { "label": "Spotlight Oral", "source": "https://icml-structured-fm-workshop.github.io/accepted-papers/", "verifiedAt": "2026-10-12" }, "addedAt": "2026-10-12" }, { "id": "arxiv-2606.11473", "title": "CRUMB: Efficient Prior Fitted Network Inference via Distributionally Matched Context Batching", "date": "2026-06-09", "updated": "2026-06-09", "category": 5, "related": [], "topic": "上下文样本优化", "status": "published", "venue": "NeurIPS 2026", "summary": "先聚类整批无标签测试查询，再用 MMD 选择匹配的训练子集，共享短上下文批量推理。作者在 51 个 TabArena 数据集上比较等上下文预算方法，并扩展到多种 PFN。已核实 NeurIPS 2026 主会录用；实验剔除了超时运行，重复种子也非独立任务，因此精度与效率优势仍需完整复核；不能直接当作逐条在线服务。", "url": "https://arxiv.org/abs/2606.11473", "pdf": "https://arxiv.org/pdf/2606.11473", "code": null, "aliases": [ "CRUMB" ], "keywords": [ "TabArena", "MMD", "上下文选择", "批量推理" ], "sources": [ "https://arxiv.org/abs/2606.11473", "https://arxiv.org/html/2606.11473v1", "https://neurips.cc/Downloads/2026", "https://openreview.net/forum?id=z3XMLStVPC", "https://neurips.cc/virtual/2026/poster/148277" ], "section": "context-selection", "topicId": "representative-context", "addedAt": "2026-10-12" }, { "id": "arxiv-2606.05441", "title": "GOTabPFN: From Feature Ordering to Compact Tokenization for Tabular Foundation Models on High-Dimensional Data", "date": "2026-06-03", "updated": "2026-06-07", "category": 5, "section": "inference-efficiency", "topicId": "context-compression", "related": [ 2, 4 ], "topic": "上下文压缩与信息迁移", "status": "published", "venue": "ICML 2026", "summary": "GOTabPFN 用关系驱动的特征排序与局部压缩，将高维表折成较少元特征后输入冻结 TabPFN。在八个生物医学高维小样本任务上做 5×5 交叉验证，并对照不同压缩表示。收益来自排序与压缩组合，不能理解为仅换列顺序即可增强不变模型；增加预处理成本，仍继承主干容量限制。", "url": "https://arxiv.org/abs/2606.05441", "pdf": "https://arxiv.org/pdf/2606.05441", "code": "https://github.com/zadid6pretam/GOTabPFN", "keywords": [], "aliases": [ "GOTabPFN", "GO-LR", "NSC" ], "sources": [ "https://arxiv.org/abs/2606.05441", "https://proceedings.mlr.press/v306/habib26a.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2605.27254", "title": "LUCoS: Latent Unsupervised Context Selection for Tabular Foundation Models", "date": "2026-05-26", "updated": "2026-05-26", "category": 5, "related": [], "topic": "上下文样本优化", "status": "preprint", "venue": "预印本", "summary": "LUCoS 在无监督 PFN 的行嵌入中选代表样本，先选后标注，再交给独立监督 TFM。在 67 个不超过十类的 CC18 任务上比较不同标注预算与表示、选择器消融。最小预算的收益主要来自覆盖而非潜在表示；ZEUS 比较因性能与显存问题提前停止，只覆盖部分任务，应与完整主比较分开。", "url": "https://arxiv.org/abs/2605.27254", "pdf": "https://arxiv.org/pdf/2605.27254", "code": "https://github.com/ari-dasci/S-LUCoS", "keywords": [], "section": "context-selection", "topicId": "retrieval-shots", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.21288", "title": "A Mechanistic Study of Tabular Foundation Models", "date": "2026-05-20", "updated": "2026-05-20", "category": 5, "related": [ 2, 6 ], "topic": "表格推理策略", "status": "preprint", "venue": "预印本", "summary": "通过逐层探针、读出替换、排列干预与压力测试，研究 TabPFN、Mitra、TabICL 的表征和读出机制。覆盖 49 个分类与十个回归任务，但高成本攻击和消融使用明确子集。观察到的晚层投票或原型读出针对所测版本和探针；不等于模型整体是固定权重算法，也不直接证明这些机制导致泛化优势。", "url": "https://arxiv.org/abs/2605.21288", "pdf": "https://arxiv.org/pdf/2605.21288", "code": null, "keywords": [], "section": "inference-strategies", "topicId": "inference-mechanisms", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.12904", "title": "VIP-COp: Black-box Context Optimization for Tabular Foundation Models", "date": "2026-05-13", "updated": "2026-09-29", "category": 5, "related": [ 2 ], "topic": "上下文样本优化", "status": "preprint", "venue": "预印本", "summary": "VIP-COp 在黑盒模型外估计样本和特征的重要性，迭代采样、验证并分配上下文预算。在 38 个 TALENT 分类任务及异常检测设置比较上下文选择，并分析优化后的截断排名。需要有标签验证集和多次模型调用；估计重要性是所用模型与验证目标下的预测贡献，不是因果贡献。", "url": "https://arxiv.org/abs/2605.12904", "pdf": "https://arxiv.org/pdf/2605.12904", "code": null, "keywords": [], "section": "context-selection", "topicId": "representative-context", "addedAt": "2026-10-12", "aliases": [ "VIP-COP: Context Optimization for Tabular Foundation Models" ] }, { "id": "arxiv-2605.06510", "title": "Is One Layer Enough? Understanding Inference Dynamics in Tabular Foundation Models", "date": "2026-05-07", "updated": "2026-05-07", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 1, 6 ], "topic": "模型机制与行为分析", "status": "published", "venue": "ICML 2026", "summary": "逐层探测和干预六种 TFM，研究迭代细化与深度冗余，并用共享层循环验证参数复用。主分析覆盖所选二分类任务，附多分类与回归，另以 nanoTabPFN 做受控循环模型训练。单层循环仍执行多次，计算量可与六层模型相当；冗余探针不能推出直接删层无损或无需预训练。", "url": "https://arxiv.org/abs/2605.06510", "pdf": "https://arxiv.org/pdf/2605.06510", "code": "https://github.com/amirbalef/is_one_layer_enough", "keywords": [], "aliases": [ "Is One Layer Enough", "Looped Single-Layer TFM" ], "sources": [ "https://arxiv.org/abs/2605.06510", "https://proceedings.mlr.press/v306/balef26a.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2602.05649", "title": "End-to-End Compression for Tabular Foundation Models", "date": "2026-02-05", "updated": "2026-05-29", "category": 5, "related": [ 6 ], "topic": "大表推理效率", "status": "published", "venue": "ICML 2026", "summary": "TACO 联合训练上下文压缩器与预测器，将大训练集转换为较短潜在记忆。在 TabArena 分类子集与合成形状扫描中，分开测试缓存、分批查询、压缩率和显存。压缩器与预测器需联合训练，不是任意现成 TFM 的即插即用后处理；最高提速与显存比例依赖压缩率、缓存及查询批次。", "url": "https://arxiv.org/abs/2602.05649", "pdf": "https://arxiv.org/pdf/2602.05649", "code": "https://github.com/machinelearningnuremberg/TACO", "keywords": [], "section": "inference-efficiency", "topicId": "context-compression", "recognition": { "label": "Spotlight", "source": "https://arxiv.org/abs/2602.05649", "verifiedAt": "2026-10-12" }, "sources": [ "https://arxiv.org/abs/2602.05649", "https://proceedings.mlr.press/v306/zabergja26a.html" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2406.05207", "title": "Retrieval & Fine-Tuning for In-Context Tabular Models", "date": "2024-06-07", "updated": "2024-06-07", "category": 5, "related": [ 6 ], "topic": "上下文样本优化", "status": "published", "venue": "NeurIPS 2024", "summary": "LoCalPFN 为每个查询检索近邻，再用共享局部上下文做目标数据微调。95 个经过特征、类别和缺失限制筛选的 TabZilla 任务，区分纯检索与检索加微调。原实验基于早期 TabPFN 分类主干，不能直接外推最新版本或回归；应计入每任务微调与查询检索成本。", "url": "https://arxiv.org/abs/2406.05207", "pdf": "https://arxiv.org/pdf/2406.05207", "code": null, "keywords": [ "TabPFN" ], "section": "context-selection", "topicId": "retrieval-shots", "addedAt": "2026-10-12" }, { "id": "arxiv-2402.11137", "title": "TuneTables: Context Optimization for Scalable Prior-Data Fitted Networks", "date": "2024-02-17", "updated": "2024-10-21", "category": 5, "related": [ 2, 6 ], "topic": "大表推理效率", "status": "published", "venue": "NeurIPS 2024", "summary": "TuneTables 将大目标数据集的信息优化进紧凑软上下文，并组合特征处理和类别扩展。通过验证集选择配置、三折 OpenML 测试，分别比较样本压缩、微调、软提示与较大数据任务。需要每任务优化且各方法搜索策略不同；原版 PFN 的特征与类别约束不能直接代表当前模型限制。", "url": "https://arxiv.org/abs/2402.11137", "pdf": "https://arxiv.org/pdf/2402.11137", "code": "https://github.com/penfever/TuneTables", "keywords": [], "section": "inference-efficiency", "topicId": "context-compression", "addedAt": "2026-10-12" }, { "id": "arxiv-2610.13818", "title": "TabMF: Learning Dataset Representations with Tabular Foundation Models", "date": "2026-10-09", "updated": "2026-10-12", "category": 6, "section": "adaptation-target", "topicId": "parameter-finetuning", "related": [ 4, 5 ], "topic": "参数微调与训练策略", "status": "preprint", "venue": "预印本", "summary": "TabMF 将已有表格基础模型改为整个有标签数据集的编码器：结合架构专用读出与合成数据视图的对比后训练，联合优化骨干和表示头。以 TabICLv2、TabPFN-2.5 为骨干，在真实未见分类数据上进行六类评估，支持数据集检索、基准压缩和分布偏移分析。表示依赖指定目标列，回归与其他骨干的泛化尚未确立。", "url": "https://arxiv.org/abs/2610.13818", "pdf": "https://arxiv.org/pdf/2610.13818", "code": null, "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "TabMF" ], "sources": [ "https://arxiv.org/abs/2610.13818", "https://arxiv.org/html/2610.13818" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.01435", "title": "Distillation of Tabular Foundation Models into Efficient Predictors", "date": "2026-10-01", "updated": "2026-10-01", "category": 6, "related": [ 1, 5 ], "topic": "表格预测能力蒸馏", "status": "preprint", "venue": "预印本", "summary": "研究 TFM 蒸馏中的教师监督构造与合成查询覆盖，将预测能力转移到 TabM 和 XGBoost。两个教师、两个学生在 TabArena 与 TALENT 比较，并补充排除设计折后的分析及一次性训练成本。教师为训练行生成软标签与最终测试应严格分开；更快学生需先付出教师查询与蒸馏成本，结论仅覆盖所测教师学生组合。", "url": "https://arxiv.org/abs/2610.01435", "pdf": "https://arxiv.org/pdf/2610.01435", "code": null, "keywords": [], "section": "adaptation-distillation", "topicId": "distillation-recipes", "addedAt": "2026-10-12" }, { "id": "arxiv-2610.01143", "title": "Parameter-Efficient Distributionally Robust Adaptation of Tabular Foundation Models under Subpopulation Shift", "date": "2026-10-01", "updated": "2026-10-01", "category": 6, "related": [ 2, 5 ], "topic": "数据分布适配", "status": "preprint", "venue": "预印本", "summary": "DR-TFM 以小型查询缩放网络配合分布鲁棒目标，降低子群比例变化下的最差组错误。六类二分类数据及 ACS 州内／跨州设置报告平均组和最差组准确率，区分估计组与真组标签基线。不使用真实组标签仍依赖所设模糊集合和分组方式；最差组改善不等于满足所有公平性定义或任意分布偏移。", "url": "https://arxiv.org/abs/2610.01143", "pdf": "https://arxiv.org/pdf/2610.01143", "code": null, "keywords": [], "section": "adaptation-distribution", "topicId": "subpopulation-shift", "addedAt": "2026-10-12" }, { "id": "arxiv-2609.04901", "title": "Adaptation Interfaces for In-Context Tabular Foundation Models in Time-to-Event Prediction", "date": "2026-09-04", "updated": "2026-09-12", "category": 6, "section": "adaptation-target", "topicId": "parameter-finetuning", "related": [ 1, 2 ], "topic": "参数微调与训练策略", "status": "preprint", "venue": "预印本", "summary": "比较时间展开分类、冻结 TFM 加生存头等接口，将表格基础模型适配到删失事件时间预测。在多种规模的单风险及四个竞争风险任务上评估 concordance 与 IBS。文中监督适配主要训练任务头，骨干仍冻结；IBS 同时受区分与概率误差影响，不能只称校准指标，也不能代替临床效用。", "url": "https://arxiv.org/abs/2609.04901", "pdf": "https://arxiv.org/pdf/2609.04901", "code": "https://github.com/kaylode/survival-fm", "keywords": [ "TFM", "表格基础模型" ], "aliases": [ "Survival FM", "TabFM survival adaptation" ], "sources": [ "https://arxiv.org/abs/2609.04901" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2605.18654", "title": "Pocket Foundation Models: Distilling TFMs into CPU-Ready Gradient-Boosted Trees", "date": "2026-05-18", "updated": "2026-05-18", "category": 6, "related": [ 1, 2 ], "topic": "表格预测能力蒸馏", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "Pocket Foundation Models 用折外教师软标签训练 CPU 可部署的树学生，并比较单教师与多教师。在 153 个全部配置均能完成的分类任务上评估精度保留、校准和推理费用。每配置只有一个种子，表中标准差是跨数据集差异；AUC 比值不是超越随机水平的能力保留率，也不能忽略教师预处理成本。", "url": "https://arxiv.org/abs/2605.18654", "pdf": "https://arxiv.org/pdf/2605.18654", "code": "https://github.com/Lexsi-Labs/TabTune", "keywords": [], "section": "adaptation-distillation", "topicId": "tree-distillation", "addedAt": "2026-10-12" }, { "id": "arxiv-2605.06047", "title": "TFM-Retouche: A Lightweight Input-Space Adapter for Tabular Foundation Models", "date": "2026-05-07", "updated": "2026-05-08", "category": 6, "related": [ 5 ], "topic": "目标数据适配", "status": "workshop", "venue": "ICML 2026 FMSD Workshop", "summary": "TFM-Retouche 冻结主干，在输入端训练残差适配器，并根据验证结果选择回退原模型。以 TabICLv2 在 51 个 TabArena-Lite 任务上比较调参与集成配置。八折 bagging 不是八次独立外层测试；耗时采用 A10，与其他硬件榜单不宜直比，验证回退也不保证测试集必然无损。", "url": "https://arxiv.org/abs/2605.06047", "pdf": "https://arxiv.org/pdf/2605.06047", "code": null, "keywords": [], "section": "adaptation-target", "topicId": "input-adapters", "addedAt": "2026-10-12" }, { "id": "arxiv-2601.09654", "title": "Exploring Fine-Tuning for Tabular Foundation Models", "date": "2026-01-14", "updated": "2026-01-14", "category": 6, "related": [ 2 ], "topic": "目标表格数据微调", "status": "preprint", "venue": "预印本", "summary": "比较多种 TFM 的冻结推理、监督微调、episodic 微调和参数高效适配。在三组分类基准的共同可运行子集分析规模、维度、不平衡、校准与组间差异。部分模型不支持 PEFT，敏感属性由人工定义且跨数据集不同；共同子集和配置差异限制全面排名及公平性外推。", "url": "https://arxiv.org/abs/2601.09654", "pdf": "https://arxiv.org/pdf/2601.09654", "code": null, "keywords": [], "section": "adaptation-target", "topicId": "parameter-finetuning", "addedAt": "2026-10-12" }, { "id": "arxiv-2506.08982", "title": "On Finetuning Tabular Foundation Models", "date": "2025-06-10", "updated": "2025-06-11", "category": 6, "related": [ 5 ], "topic": "目标表格数据微调", "status": "preprint", "venue": "预印本", "summary": "比较 TabPFNv2 的完整微调、LoRA 与输入表示微调，并分析注意力如何随任务变化。在可装入单张 80GB GPU 的学术数据及 TabReD 子集上比较，TabReD 使用多次抽样和种子。容量限制与预处理差异影响模型可比性；部分实际任务微调变差，注意力近邻探针不能证明整个模型等同于 kNN。", "url": "https://arxiv.org/abs/2506.08982", "pdf": "https://arxiv.org/pdf/2506.08982", "code": "https://github.com/yandex-research/tabpfn-finetuning", "keywords": [ "TabPFN" ], "section": "adaptation-target", "topicId": "parameter-finetuning", "addedAt": "2026-10-12" }, { "id": "arxiv-2405.13396", "title": "Fine-tuned In-Context Learning Transformers are Excellent Tabular Data Classifiers", "date": "2024-05-22", "updated": "2025-01-23", "category": 6, "section": "adaptation-target", "topicId": "parameter-finetuning", "related": [ 1, 5 ], "topic": "参数微调与训练策略", "status": "preprint", "venue": "预印本", "summary": "TabForestPFN 将森林生成先验与按任务划分支持集、查询集的 episodic 微调结合。在过滤后的 TabZilla 和 WhyTrees 分类任务上比较冻结与微调，并以验证集早停。各基线调参预算不完全相同，森林先验的零样本收益并不一致；实验不支持对所有数据规模或回归任务外推。", "url": "https://arxiv.org/abs/2405.13396", "pdf": "https://arxiv.org/pdf/2405.13396", "code": null, "keywords": [ "TFM", "TabForestPFN" ], "aliases": [ "TabForestPFN", "TabForest" ], "sources": [ "https://arxiv.org/abs/2405.13396" ], "addedAt": "2026-10-12" }, { "id": "arxiv-2610.14998", "title": "Recurrent In-Context Learning for Tabular Foundation Models", "date": "2026-10-10", "updated": "2026-10-10", "category": 5, "section": "inference-strategies", "topicId": "test-time-compute", "related": [ 1 ], "topic": "推理时计算与组合策略", "status": "preprint", "venue": "预印本", "summary": "Tab-Loop 将 TabICLv2 的 12 个独立上下文学习块换成 4 个共享块循环执行。在相同预训练计算和 12 次块调用下，以少 62% 的参数取得接近的预测表现，评估覆盖 200 个 TALENT 数据集和 38 个 TabArena 分类数据集。继续增加循环可扩展计算深度，但参数减少不等于推理计算或延迟同比下降。", "url": "https://arxiv.org/abs/2610.14998", "pdf": "https://arxiv.org/pdf/2610.14998", "code": null, "keywords": [ "TFM", "ICL", "循环推理", "权重共享", "TabArena", "TALENT" ], "aliases": [ "Tab-Loop", "Tab Loop" ], "sources": [ "https://arxiv.org/abs/2610.14998", "https://arxiv.org/pdf/2610.14998" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.14989", "title": "Characterizing the Separability Gap in Prior-Data Fitted Networks", "date": "2026-10-10", "updated": "2026-10-10", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 2, 6 ], "topic": "模型机制与行为分析", "status": "preprint", "venue": "预印本", "summary": "分析 PFN 先验中输入分布与标签机制相互依赖时，未标注样本为何会改变预测；并训练 SemiTabICL 显式利用未标注上下文。在 150 个 semi-TALENT 分类数据集和 23 个 UCI 任务上检验收益。使用时需区分独立未标注数据与测试查询，先验失配也可能使模型过度拟合输入分布。", "url": "https://arxiv.org/abs/2610.14989", "pdf": "https://arxiv.org/pdf/2610.14989", "code": null, "keywords": [ "TFM", "半监督学习", "贝叶斯推断", "未标注数据" ], "aliases": [ "SemiTabICL", "separability gap" ], "sources": [ "https://arxiv.org/abs/2610.14989", "https://arxiv.org/html/2610.14989" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.13764", "title": "How Well Do Prior-Data Fitted Networks Amortize Bayesian Inference?", "date": "2026-10-09", "updated": "2026-10-09", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 1, 2 ], "topic": "模型机制与行为分析", "status": "preprint", "venue": "预印本", "summary": "通过可计算参考后验与独立模型间分歧分析 PFN 的摊销推断误差，并比较深度集成和同一参数盆地内采样。除小模型与 TabArena 外，完整规模 TabICLv2 集成也在筛选后的 255 个 TALENT 分类／回归任务上评估。生产先验下只能观测可移除的误差部分；参考 MCMC 也有有限预算，集成需计入多模型训练和推理成本。", "url": "https://arxiv.org/abs/2610.13764", "pdf": "https://arxiv.org/pdf/2610.13764", "code": null, "keywords": [ "PFN", "贝叶斯推断", "深度集成", "TabArena" ], "aliases": [ "PFN amortization error" ], "sources": [ "https://arxiv.org/abs/2610.13764", "https://arxiv.org/html/2610.13764" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.13644", "title": "SelectArena: A Modern Benchmark for Feature Selection on Real-world Tabular Data", "date": "2026-10-09", "updated": "2026-10-09", "category": 1, "section": "industry-benchmarks", "topicId": "general-benchmarks", "related": [ 4, 5 ], "topic": "通用评测与分布泛化", "status": "preprint", "venue": "预印本", "summary": "SelectArena 统一特征选择的任务、特征预算、下游预测器和计算限制。在 71 个真实分类／回归任务上比较 15 种方法和五个预测器，包含 TabArena 任务与 TFM。部分方法不能按预算返回特征时采用随机回退，榜单因此衡量整套处理流程；方法筛选也不代表所有特征选择研究。", "url": "https://arxiv.org/abs/2610.13644", "pdf": "https://arxiv.org/pdf/2610.13644", "code": null, "keywords": [ "TabArena", "特征选择", "评测协议", "复现" ], "aliases": [ "SelectArena" ], "sources": [ "https://arxiv.org/abs/2610.13644", "https://arxiv.org/pdf/2610.13644" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.12904", "title": "GIST: Generator-state-Informed Supervision for Tabular Foundation Models", "date": "2026-10-09", "updated": "2026-10-09", "category": 1, "section": "industry-benchmarks", "topicId": "general-benchmarks", "related": [ 2 ], "topic": "通用评测与分布泛化", "status": "preprint", "venue": "预印本", "summary": "GIST 根据已知合成机制为 TFM 提供条件统计量监督，减少学习目标中的观测噪声。在匹配架构的 TabArena 与 TALENT 实验中比较原始监督和统计目标，并测试训练分布外任务。生成机制和监督设计共同构成方法；TALENT 重新划分，分布外比较的可运行任务随骨干不同，跨骨干结果不可直接排名。", "url": "https://arxiv.org/abs/2610.12904", "pdf": "https://arxiv.org/pdf/2610.12904", "code": null, "keywords": [ "TFM", "监督目标", "梯度方差", "TabArena", "TALENT" ], "aliases": [ "GIST", "Generator-state-Informed Supervision" ], "sources": [ "https://arxiv.org/abs/2610.12904", "https://arxiv.org/html/2610.12904" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.12390", "title": "Long Text to Predictive Features: LLM-Guided Blockwise Feature Engineering via Executable Program Search", "date": "2026-10-08", "updated": "2026-10-08", "category": 4, "section": "semantic-features", "topicId": "feature-search", "related": [ 1, 2 ], "topic": "任务反馈与特征搜索", "status": "preprint", "venue": "预印本", "summary": "LLM-BlockFE 通过增量程序搜索、预测验证反馈与回溯，把长文本转成可固化的特征程序，线上执行无需 LLM。两组公开和两组私有任务比较，另报告五个风控场景的同样本 KS 对照及业务 A/B 结果；线上 A/B 的随机分组与统计细节披露不足，整体成本仍需计入离线程序搜索。", "url": "https://arxiv.org/abs/2610.12390", "pdf": "https://arxiv.org/pdf/2610.12390", "code": null, "keywords": [ "语义结构化", "特征工程", "程序搜索", "金融风控" ], "aliases": [ "LLM-BlockFE", "BlockFE" ], "sources": [ "https://arxiv.org/abs/2610.12390", "https://arxiv.org/html/2610.12390" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.10832", "title": "MotherTree: Meta-learning on synthetic data improves decision tree training", "date": "2026-10-07", "updated": "2026-10-07", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5, 6 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "MotherTree 用 Transformer 根据新表格一次生成可独立运行的硬决策树，并可进一步优化生成的树参数。它在合成任务上学习建树方式，无需参考树作为监督；在树大小匹配的两组基准上与传统、可微及蒸馏树比较，优势主要出现在小样本二分类。生成树与把教师模型蒸馏成树是不同路线。", "url": "https://arxiv.org/abs/2610.10832", "pdf": "https://arxiv.org/pdf/2610.10832", "code": null, "keywords": [ "决策树", "元学习", "模型生成", "可解释预测" ], "aliases": [ "MotherTree", "MotherTree-Tune" ], "sources": [ "https://arxiv.org/abs/2610.10832", "https://arxiv.org/html/2610.10832" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.10068", "title": "Efficient Provably Private Classification with a Tabular Foundation Model", "date": "2026-10-07", "updated": "2026-10-07", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "PrivTab 将敏感训练行压缩为加入校准噪声的私有摘要，再据此预测，避免每个数据集重新进行私有优化。实验使用 33 个 TabArena 分类数据集，另以五个案例计入私有预处理成本。隐私分析以每人一行为前提；主基准的理想化标准化不能直接视为端到端隐私保证。", "url": "https://arxiv.org/abs/2610.10068", "pdf": "https://arxiv.org/pdf/2610.10068", "code": null, "keywords": [ "TFM", "差分隐私", "上下文压缩", "TabArena" ], "aliases": [ "PrivTab" ], "sources": [ "https://arxiv.org/abs/2610.10068", "https://arxiv.org/html/2610.10068" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.09487", "title": "Correspondences as Decisions: JevNexus for Decision-Centric Schema Matching", "date": "2026-10-07", "updated": "2026-10-07", "category": 4, "section": "semantic-integration", "topicId": "schema-matching", "related": [ 1, 2 ], "topic": "字段匹配与模式整合", "status": "preprint", "venue": "预印本", "summary": "JevNexus 先检索候选字段，再用 Jev 输出配对判断并融合字段／取值证据，只对有歧义的候选重排。六组基准共 561 个匹配案例中，与 Magneto 的 MRR 和首位命中率差异未达统计显著；在同一 RTX PRO 6000 主机、排除模型启动时间的比较中，平均延迟约低至其 1/7.75。", "url": "https://arxiv.org/abs/2610.09487", "pdf": "https://arxiv.org/pdf/2610.09487", "code": "https://github.com/RazeenLI/JevNexus", "keywords": [ "Jev", "模式匹配", "结构化决策", "选择性重排" ], "aliases": [ "JevNexus", "DeMa" ], "sources": [ "https://arxiv.org/abs/2610.09487", "https://arxiv.org/html/2610.09487", "https://github.com/RazeenLI/JevNexus" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.08078", "title": "ProximalFM: Amortized Proximal Causal Inference under Hidden Confounding", "date": "2026-10-06", "updated": "2026-10-06", "category": 3, "section": "causal-estimation", "topicId": "treatment-effects", "related": [ 2, 6 ], "topic": "处理效应与干预分布", "status": "preprint", "venue": "预印本", "summary": "ProximalFM 在满足近端因果识别条件的合成机制上学习，通过可观测代理处理隐混杂；训练目标显式计入蒙特卡洛监督噪声，推理输出条件平均处理效应及后验。在 12 组真实协变量构造的 288 个半合成实例上评估；分布偏移下区间过度自信，超过训练上下文后收益趋缓，物理系统实验未全面胜出。", "url": "https://arxiv.org/abs/2610.08078", "pdf": "https://arxiv.org/pdf/2610.08078", "code": "https://github.com/ChristopheMuller/ProximalFM", "keywords": [ "因果基础模型", "隐混杂", "代理变量", "CATE" ], "aliases": [ "ProximalFM", "ProximalFM-FT" ], "sources": [ "https://arxiv.org/abs/2610.08078", "https://arxiv.org/html/2610.08078", "https://github.com/ChristopheMuller/ProximalFM" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2609.33114", "title": "Can Tabular Foundation Models Amortize Statistical Inference?", "date": "2026-09-27", "updated": "2026-09-27", "category": 6, "section": "adaptation-target", "topicId": "parameter-finetuning", "related": [ 2, 5 ], "topic": "参数微调与训练策略", "status": "preprint", "venue": "预印本", "summary": "TabCon 冻结预测主干，通过残差专家与经过监督及强化学习训练的路由器输出条件均值置信区间。64 类合成机制及 13 个 OpenML 数据集构造的半合成任务比较覆盖率、长度与成本。结果是所测机制的经验覆盖，不能视为任意分布的有限样本保证；均值置信区间不同于未来观测的预测区间。", "url": "https://arxiv.org/abs/2609.33114", "pdf": "https://arxiv.org/pdf/2609.33114", "code": null, "keywords": [ "TFM", "强化学习后训练", "稀疏专家", "置信区间" ], "aliases": [ "TabCon" ], "sources": [ "https://arxiv.org/abs/2609.33114", "https://arxiv.org/html/2609.33114" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2609.16309", "title": "Agentic Search Spaces for Tabular Machine Learning", "date": "2026-09-14", "updated": "2026-09-14", "category": 5, "section": "inference-strategies", "topicId": "pipeline-search", "related": [ 1, 4 ], "topic": "自动管线与策略搜索", "status": "preprint", "venue": "预印本", "summary": "让代码智能体为表格模型一次生成可复用的模块选项，再由传统超参数搜索按数据集选择组合；对 TabICLv2 主要扩展预处理与推理模块。研究在 45 个数据集及 TabArena 上评估迁移收益。需区分生成成本与调参预算，且 TabICLv2 的默认对照未调参，不能把所有比较都解读为等预算算法优势。", "url": "https://arxiv.org/abs/2609.16309", "pdf": "https://arxiv.org/pdf/2609.16309", "code": null, "keywords": [ "AutoML", "代码生成", "超参数优化", "TabICL", "TabArena" ], "aliases": [ "Agentic Search Spaces" ], "sources": [ "https://arxiv.org/abs/2609.16309", "https://arxiv.org/html/2609.16309" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2608.27882", "title": "SOMTab: Set-Order Mamba for Efficient Tabular In-Context Learning", "date": "2026-08-28", "updated": "2026-08-28", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "SOMTab 将无序表格映射为固定潜在槽位，用 Mamba 构造行列表示，同时保留最终按查询检索标签上下文的注意力。研究比较分类性能、显存与推理成本，并补充 TabArena 结果；其中榜单与自测使用不同 GPU，速度图只能作为参考，不能视为同硬件加速验证。", "url": "https://arxiv.org/abs/2608.27882", "pdf": "https://arxiv.org/pdf/2608.27882", "code": null, "keywords": [ "TFM", "Mamba", "状态空间模型", "TabArena" ], "aliases": [ "SOMTab", "Set-Order Mamba", "DCH-TailMix" ], "sources": [ "https://arxiv.org/abs/2608.27882", "https://arxiv.org/html/2608.27882" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2608.20841", "title": "Foundation Models for Partial Causal Identification", "date": "2026-08-21", "updated": "2026-08-21", "category": 3, "section": "causal-models", "topicId": "causal-reliability", "related": [ 2 ], "topic": "效应去偏与可靠性", "status": "preprint", "venue": "预印本", "summary": "当观测数据无法唯一确定因果效应时，研究用覆盖离散结构因果模型的规范先验训练 PFN，以后验支持范围表示可识别集合。理论给出特定条件下的渐近关系，实验主要是具有解析界的二变量二值系统；不能据此认定已解决高维真实数据的有限样本因果界估计。", "url": "https://arxiv.org/abs/2608.20841", "pdf": "https://arxiv.org/pdf/2608.20841", "code": null, "keywords": [ "因果基础模型", "部分识别", "反事实边界", "隐混杂" ], "aliases": [ "Partial Causal Identification" ], "sources": [ "https://arxiv.org/abs/2608.20841", "https://arxiv.org/html/2608.20841" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2608.18919", "title": "Lost in Aggregation: How Benchmarks Overlook Irreplaceable Model Strengths", "date": "2026-08-19", "updated": "2026-08-19", "category": 1, "section": "industry-benchmarks", "topicId": "general-benchmarks", "related": [ 2 ], "topic": "通用评测与分布泛化", "status": "preprint", "venue": "预印本", "summary": "基于 TabArena 结果，区分模型在每个数据集上是否具有不可替代的最优表现，分析平均排名、胜率和 Elo 遗漏的信息。论文发现常用聚合指标更偏向稳定表现及避免失误，可能掩盖少数任务上的独特优势；该分析聚焦预测性能，不包含部署成本与集成收益。", "url": "https://arxiv.org/abs/2608.18919", "pdf": "https://arxiv.org/pdf/2608.18919", "code": null, "keywords": [ "TabArena", "模型比较", "榜单聚合", "统计评测" ], "aliases": [ "Lost in Aggregation", "peak performance frontier" ], "sources": [ "https://arxiv.org/abs/2608.18919", "https://arxiv.org/html/2608.18919" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2608.18849", "title": "GEAR: Generative Expansion and Real Anchoring for Two-Stage Distillation of Tabular Foundation Models", "date": "2026-08-19", "updated": "2026-08-25", "category": 6, "section": "adaptation-distillation", "topicId": "distillation-recipes", "related": [ 1, 5 ], "topic": "教师信号与学生比较", "status": "preprint", "venue": "预印本", "summary": "GEAR 先在合成特征上查询 TFM 软标签训练学生，再用真实标签与折外教师预测校正，避免训练行同时进入教师上下文造成标签泄漏。以 TabPFN、TabICL、TabDPT 为教师，在 TALENT 和 TabArena 上检验 MLP 与树学生。合成样本数量和生成质量存在折中，部署收益须结合教师 GPU 与学生 CPU 的计时条件判断。", "url": "https://arxiv.org/abs/2608.18849", "pdf": "https://arxiv.org/pdf/2608.18849", "code": null, "keywords": [ "TFM", "两阶段蒸馏", "合成查询", "TabArena", "TALENT" ], "aliases": [ "GEAR", "Generative Expansion and Real Anchoring" ], "sources": [ "https://arxiv.org/abs/2608.18849", "https://arxiv.org/html/2608.18849" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2608.17957", "title": "Understanding the Surprising Generalization Properties of Tabular Foundation Models", "date": "2026-08-18", "updated": "2026-08-18", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 1, 4 ], "topic": "模型机制与行为分析", "status": "published", "venue": "NeurIPS 2026", "summary": "通过单表与多表训练、列清洗和受控消融，研究真实数据训练的 TFM 如何跨数据集泛化；列级任务多样性比简单整表去重更有解释力，并在 51 个 TabArena 任务上复核。学习到的检索机制得到实验支持，但未排除其他机制；结果主要来自行注意力模型，不能直接推广到所有合成先验模型。", "url": "https://arxiv.org/abs/2608.17957", "pdf": "https://arxiv.org/pdf/2608.17957", "code": null, "keywords": [ "TFM", "任务多样性", "检索机制", "TabArena" ], "aliases": [ "TFM Generalization Properties" ], "sources": [ "https://arxiv.org/abs/2608.17957", "https://arxiv.org/html/2608.17957", "https://neurips.cc/virtual/2026/poster/150354", "https://openreview.net/forum?id=ikYK94iVvj" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2608.16319", "title": "Advancing Open and Reproducible Relational Learning: RelArena-α, TabPFN-Rel and RPI", "date": "2026-08-17", "updated": "2026-09-07", "category": 1, "section": "industry-models", "topicId": "relational-releases", "related": [ 4, 5 ], "topic": "关系数据模型", "status": "preprint", "venue": "预印本", "summary": "RelArena-α 统一关系预测的时间截点、调参和评价；TabPFN-Rel 将关系聚合成表格，RPI 提供任务接口。初版在 21 个 RelBench 实体任务、单随机种子下区分模型与系统比较。RDBLearn 对照未包含 LimiX；API 与开源版文本输入不同，部分系统预算独立，旧快照结果也不可直接比较。", "url": "https://arxiv.org/abs/2608.16319", "pdf": "https://arxiv.org/pdf/2608.16319", "code": "https://github.com/PriorLabs/relarena", "keywords": [ "关系数据", "TabPFN", "TabArena", "评测协议", "特征聚合" ], "aliases": [ "RelArena", "RelArena-alpha", "RelArena-α", "TabPFN-Rel", "RPI" ], "sources": [ "https://arxiv.org/abs/2608.16319", "https://arxiv.org/html/2608.16319", "https://github.com/PriorLabs/relarena" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2607.26628", "title": "Understanding Context Sampling in TabPFN on Small Tabular Datasets", "date": "2026-07-29", "updated": "2026-07-29", "category": 5, "section": "context-selection", "topicId": "representative-context", "related": [ 2 ], "topic": "代表性选择与上下文优化", "status": "preprint", "venue": "预印本", "summary": "在 15 个小型 OpenML 数据集上反复抽取 TabPFN-3 上下文，比较样本数、特征覆盖、多样性和选择成本。较大上下文使预测更稳定；只匹配特征均值可能损害多样性，K-Means 和最远点选择也未稳定胜过随机抽样。结果基于固定测试切分，不能直接外推到大表或所有分布偏移。", "url": "https://arxiv.org/abs/2607.26628", "pdf": "https://arxiv.org/pdf/2607.26628", "code": "https://github.com/mohammed1916/tabmx", "keywords": [ "TabPFN", "上下文采样", "原型选择", "小样本" ], "aliases": [ "TabPFN context sampling" ], "sources": [ "https://arxiv.org/abs/2607.26628", "https://arxiv.org/html/2607.26628", "https://github.com/mohammed1916/tabmx" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2606.08679", "title": "Rank Intervals for Leaderboards: A Hierarchical Framework for Model Evaluation", "date": "2026-06-07", "updated": "2026-06-07", "category": 1, "section": "industry-benchmarks", "topicId": "general-benchmarks", "related": [ 2 ], "topic": "通用评测与分布泛化", "status": "preprint", "venue": "预印本", "summary": "用成对检验构造单任务排名置信区间，再以保形方法给出新任务的排名预测区间。模拟数据、TabArena 归档的 44 个模型／51 个任务及 PromptEval 评估覆盖与区间长度。保证依赖检验有效性及任务可交换性，并要求模型完整评估；归档结果不代表动态实时榜单。", "url": "https://arxiv.org/abs/2606.08679", "pdf": "https://arxiv.org/pdf/2606.08679", "code": null, "keywords": [ "TabArena", "排名区间", "保形推断", "评测不确定性" ], "aliases": [ "Rank Intervals for Leaderboards" ], "sources": [ "https://arxiv.org/abs/2606.08679", "https://arxiv.org/html/2606.08679" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2606.03681", "title": "Speedrunning Tabular Foundation Model Pretraining", "date": "2026-06-02", "updated": "2026-06-02", "category": 1, "section": "industry-benchmarks", "topicId": "general-benchmarks", "related": [ 5 ], "topic": "通用评测与分布泛化", "status": "preprint", "venue": "预印本", "summary": "建立基于 nanoTabPFN 的训练竞速协议，固定单张 L40S 和目标精度，研究优化器、实现及训练策略的加速。评测采用 TabArena 的 38 个分类任务，但下采样至最多 100 特征、1,000 行；记录仅统计训练时间，不含先验生成与评测，不能视为完整 TabArena 排名或大型模型训练成本。", "url": "https://arxiv.org/abs/2606.03681", "pdf": "https://arxiv.org/pdf/2606.03681", "code": "https://github.com/borawhocodess/modded-nanotabpfn", "keywords": [ "TabArena", "训练效率", "优化器", "复现协议" ], "aliases": [ "modded-nanoTabPFN", "nanoTabPFN speedrun" ], "sources": [ "https://arxiv.org/abs/2606.03681", "https://arxiv.org/html/2606.03681", "https://github.com/borawhocodess/modded-nanotabpfn" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2606.02384", "title": "TabPrep: Closing the Feature Engineering Gap in Tabular Benchmarks", "date": "2026-06-01", "updated": "2026-06-01", "category": 4, "section": "semantic-features", "topicId": "feature-search", "related": [ 1, 5 ], "topic": "任务反馈与特征搜索", "status": "published", "venue": "NeurIPS 2026 · Evaluations and Datasets Track", "summary": "TabPrep 用有预算限制的特征生成器补充数值交互、类别条件效应和伪类别结构。合成诊断及统一 TabArena 协议比较线性、树、TabM 与 TabPFN-2.5，显式测试调参与集成。仅覆盖四个模型家族，扩展特征可能增加后续计算；这是特征工程基线而非最优特征或收益上界。", "url": "https://arxiv.org/abs/2606.02384", "pdf": "https://arxiv.org/pdf/2606.02384", "code": "https://github.com/atschalz/tabprep", "keywords": [ "TabArena", "特征工程", "类别编码", "结构模式" ], "aliases": [ "TabPrep", "CatPrep" ], "sources": [ "https://arxiv.org/abs/2606.02384", "https://arxiv.org/html/2606.02384", "https://github.com/atschalz/tabprep", "https://neurips.cc/virtual/2026/poster/139065", "https://openreview.net/forum?id=nMW7DYLGTo" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2605.30729", "title": "SemStruct: Contextualizing Semantic Embeddings with Structural Information for Schema Matching", "date": "2026-05-29", "updated": "2026-05-29", "category": 4, "section": "semantic-integration", "topicId": "schema-matching", "related": [ 1 ], "topic": "字段匹配与模式整合", "status": "published", "venue": "KDD 2026", "summary": "SemStruct 冻结语言编码器，在列、行、取值构成的异质图上训练 GNN 以改善字段匹配。Valentine 与 SOTAB-SM 实验及结构消融，训练利用源表的合成视图而不见目标匹配表。使用待评估来源表进行自监督适配，不能直接解释为对完全未见数据库的零样本迁移；仍需 GNN 训练。", "url": "https://arxiv.org/abs/2605.30729", "pdf": "https://arxiv.org/pdf/2605.30729", "code": null, "keywords": [ "模式匹配", "语义嵌入", "图神经网络", "表格结构" ], "aliases": [ "SemStruct" ], "sources": [ "https://arxiv.org/abs/2605.30729", "https://arxiv.org/html/2605.30729", "https://research.ibm.com/publications/semstruct-contextualizing-semantic-embeddings-with-structural-information-for-schema-matching" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2605.28418", "title": "Explaining Tabular Foundation Model Differences Through Meta-Features", "date": "2026-05-27", "updated": "2026-07-01", "category": 1, "section": "industry-benchmarks", "topicId": "general-benchmarks", "related": [ 2, 5 ], "topic": "通用评测与分布泛化", "status": "preprint", "venue": "预印本", "summary": "分析 51 个 TabArena 数据集的元特征是否能解释或预测树、神经网络和 TFM 的差异，并使用多重检验及留一数据集验证。整体未形成稳健的通用选模规则；相关性与预测能力不同，筛选出的稳健特征也未在每个验证折内重新选择，不能当作完整嵌套验证的路由器效果。", "url": "https://arxiv.org/abs/2605.28418", "pdf": "https://arxiv.org/pdf/2605.28418", "code": null, "keywords": [ "TabArena", "元特征", "模型选择", "统计检验" ], "aliases": [ "Revisiting Metafeatures to Explain Model Differences on Tabular Data" ], "sources": [ "https://arxiv.org/abs/2605.28418", "https://arxiv.org/html/2605.28418" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2605.11262", "title": "Latent Chain-of-Thought Improves Structured-Data Transformers", "date": "2026-05-11", "updated": "2026-05-19", "category": 5, "section": "inference-strategies", "topicId": "test-time-compute", "related": [ 1 ], "topic": "推理时计算与组合策略", "status": "preprint", "venue": "预印本", "summary": "将隐藏表示压缩成反馈 token 并重复送回 Transformer，研究结构化数据上的潜在多步推理。实验覆盖 27 个表格与 9 个时序数据集，并扩展至 nanoTabPFN；TabArena 部分使用低维小数据子集，其结果不能直接外推到完整基准或更大规模 TabPFN。", "url": "https://arxiv.org/abs/2605.11262", "pdf": "https://arxiv.org/pdf/2605.11262", "code": null, "keywords": [ "潜在推理", "TFM", "反馈Token", "TabArena" ], "aliases": [ "Latent CoT", "nanoTabPFN-CoT" ], "sources": [ "https://arxiv.org/abs/2605.11262", "https://arxiv.org/html/2605.11262" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2605.06290", "title": "Data Language Models: A New Foundation Model Class for Tabular Data", "date": "2026-05-07", "updated": "2026-05-07", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 4, 6 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "Schema-1 融合列名语义、列统计、原始单元格与缺失结构，面向预测、填补和数据领域识别。报告描述冻结主干后的任务适配与连续微调；CC18 比较仅用 18 个数据集，外部基线复用旧论文结果，未重新运行，也未覆盖较新 TFM，因此不能据此判断其相对当前模型的整体优势。", "url": "https://arxiv.org/abs/2605.06290", "pdf": "https://arxiv.org/pdf/2605.06290", "code": null, "keywords": [ "TFM", "列语义", "缺失数据", "多源表格" ], "aliases": [ "Schema-1", "Schema1", "Data Language Model", "DLM" ], "sources": [ "https://arxiv.org/abs/2605.06290", "https://arxiv.org/html/2605.06290" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2511.02802", "title": "TabTune: A Unified Library for Inference and Fine-Tuning Tabular Foundation Models", "date": "2025-11-04", "updated": "2025-12-02", "category": 6, "section": "adaptation-target", "topicId": "parameter-finetuning", "related": [ 1, 5 ], "topic": "参数微调与训练策略", "status": "preprint", "venue": "预印本", "summary": "TabTune 为不同表格基础模型统一预处理、推理和微调接口，涵盖元学习式适配、监督微调及参数高效微调，并检查校准、公平性与计算开销。论文在 TALENT、OpenML-CC18、TabZilla 的可共同运行子集上比较策略；它是适配工具与评测框架，不代表所有模型支持相同训练方式。", "url": "https://arxiv.org/abs/2511.02802", "pdf": "https://arxiv.org/pdf/2511.02802", "code": "https://github.com/Lexsi-Labs/TabTune", "keywords": [ "TFM", "微调", "LoRA", "适配评测" ], "aliases": [ "TabTune" ], "sources": [ "https://arxiv.org/abs/2511.02802", "https://arxiv.org/html/2511.02802", "https://github.com/Lexsi-Labs/TabTune" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2605.10616", "title": "MulTaBench: Benchmarking Multimodal Tabular Learning with Text and Image", "date": "2026-05-11", "updated": "2026-09-27", "category": 1, "section": "industry-benchmarks", "topicId": "semantic-relational-benchmarks", "related": [ 4, 6 ], "topic": "语义表格评测与技术路线比较", "status": "published", "venue": "NeurIPS 2026 · Evaluations and Datasets Track", "recognition": { "label": "Spotlight", "source": "https://arxiv.org/abs/2605.10616", "verifiedAt": "2026-10-14" }, "summary": "MulTaBench 收集 40 个文本／图像与表格联合预测任务，比较冻结嵌入和按目标微调的表示，并额外开放 40 个候选数据集。以编码器末三层 LoRA、五类表格模型及五个随机种子构造筛选协议，再用新模型和不同表示规模复核。数据集有意筛选出受益于表示微调的任务，适合研究多模态适配，排名带有筛选偏差。", "url": "https://arxiv.org/abs/2605.10616", "pdf": "https://arxiv.org/pdf/2605.10616", "code": "https://github.com/alanarazi7/MulTaBench", "keywords": [ "多模态表格", "文本特征", "图像特征", "LoRA", "目标感知表示" ], "aliases": [ "MulTaBench", "Target-Aware Representations", "TAR" ], "sources": [ "https://arxiv.org/abs/2605.10616", "https://arxiv.org/html/2605.10616v2", "https://github.com/alanarazi7/MulTaBench", "https://neurips.cc/virtual/2026/poster/138893", "https://openreview.net/forum?id=yOd4AxlEbb" ], "addedAt": "2026-10-14" }, { "id": "arxiv-2610.19814", "title": "When Does Depth Help? Understanding Layerwise Inference in Tabular Foundation Models", "date": "2026-10-14", "updated": "2026-10-14", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 1, 6 ], "topic": "模型机制与行为分析", "status": "preprint", "venue": "预印本", "summary": "以 TabPFN-3 在 272 个分类数据集上的逐层读出为主，分析深层何时仍改善预测，并在其他表格基础模型中检查相似现象。预测标签提前稳定不代表后续层无用，较难区分的查询仍可能受益；重复单层或打乱层序通常损失性能。提供推理深度优化的机制依据，尚未给出可直接部署的自适应早退规则。", "url": "https://arxiv.org/abs/2610.19814", "pdf": "https://arxiv.org/pdf/2610.19814", "code": null, "keywords": [ "TFM", "推理深度", "层间表征" ], "aliases": [ "Layerwise Inference", "When Does Depth Help" ], "sources": [ "https://arxiv.org/abs/2610.19814", "https://arxiv.org/html/2610.19814" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.17065", "title": "TILE: Bridging LLMs and Table Foundation Models for Tabular Learning with Strings", "date": "2026-10-12", "updated": "2026-10-12", "category": 4, "section": "semantic-fusion", "topicId": "text-tabular-models", "related": [ 1, 5 ], "topic": "文本与表格联合建模", "status": "preprint", "venue": "预印本", "summary": "TILE 将整行文本送入 LLM，从中间层按列提取表征并压缩，再与原始数值、类别特征一同输入表格学习器。主实验在 100 张含字符串的表上对比 30 个基线，并用另建的 STRABLE 数据验证；PCA 仅拟合训练集。收益依赖文本中的预测信息，需额外的嵌入计算，不能直接外推到纯数值表。", "url": "https://arxiv.org/abs/2610.17065", "pdf": "https://arxiv.org/pdf/2610.17065", "code": null, "keywords": [ "LLM", "TFM", "字符串特征", "语义嵌入" ], "aliases": [ "TILE" ], "sources": [ "https://arxiv.org/abs/2610.17065", "https://arxiv.org/html/2610.17065" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.16990", "title": "TabGift: A Context-Efficient Foundation Model for Tabular Data Generation", "date": "2026-10-12", "updated": "2026-10-12", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "TabGift 用表格编码器提取少量上下文，再以潜空间扩散模型生成表征，并由针对目标数据集训练的解码器还原表格。在 18 个 TinyTALENT 数据集、16–512 条上下文下评估生成保真度、下游效用与经验隐私指标。小上下文是主要优势场景；仍需训练目标解码器，隐私指标也不构成差分隐私保证。", "url": "https://arxiv.org/abs/2610.16990", "pdf": "https://arxiv.org/pdf/2610.16990", "code": null, "keywords": [ "TFM", "表格生成", "扩散模型", "小上下文" ], "aliases": [ "TabGift" ], "sources": [ "https://arxiv.org/abs/2610.16990", "https://arxiv.org/html/2610.16990" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.16763", "title": "Relational Learning as World Modeling", "date": "2026-10-12", "updated": "2026-10-12", "category": 1, "section": "industry-models", "topicId": "relational-releases", "related": [ 2, 3, 5 ], "topic": "关系数据模型", "status": "preprint", "venue": "预印本", "summary": "TabWM 把动作后的关系数据库变化表示为操作、位置和值，利用上下文示例预测完整编辑，再滚动模拟动作以支持规划。WorldRel 在 512 个未见合成世界上测试，并引入真实数据库的结构开展扩展实验。连接关系学习、状态转移与决策；真实来源的表结构仍搭配模拟动态，尚非真实业务干预效果的验证。", "url": "https://arxiv.org/abs/2610.16763", "pdf": "https://arxiv.org/pdf/2610.16763", "code": "https://github.com/oooranz/TabWM", "keywords": [ "关系基础模型", "世界模型", "关系数据库", "动作规划" ], "aliases": [ "TabWM", "WorldRel" ], "sources": [ "https://arxiv.org/abs/2610.16763", "https://arxiv.org/html/2610.16763", "https://github.com/oooranz/TabWM", "https://huggingface.co/oooranz/TabWM" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.15699", "title": "TabMosaic: Unified In-Context Diffusion for Tabular Generative Tasks", "date": "2026-10-11", "updated": "2026-10-11", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 5 ], "topic": "通用表格模型", "status": "preprint", "venue": "预印本", "summary": "TabMosaic 用统一上下文扩散完成混合类型单元格的联合与条件生成，并缓存上下文。48 张 TabArena 表的 Lite 首个官方外层划分、三个种子及 20 个基线，主要补全设置为 30% 随机缺失。不是完整多外层划分评估，原文未提出统计显著性结论；先验下目标与经验隐私指标不能保证真实数据校准或隐私。", "url": "https://arxiv.org/abs/2610.15699", "pdf": "https://arxiv.org/pdf/2610.15699", "code": null, "keywords": [ "TFM", "TabArena", "表格生成", "条件扩散", "缺失补全" ], "aliases": [ "TabMosaic" ], "sources": [ "https://arxiv.org/abs/2610.15699", "https://arxiv.org/html/2610.15699" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.17994", "title": "Improving Neural Symbolic Regression with a Diverse Prior and Post-training Alignment", "date": "2026-10-13", "updated": "2026-10-13", "category": 2, "section": "prediction-decisions", "topicId": "symbolic-regression", "related": [ 5, 6 ], "topic": "显式表达式与符号回归", "status": "preprint", "venue": "预印本", "summary": "RASR 将数值样本直接映射为显式公式，先监督学习表达式，再用强化学习对齐留出集拟合质量与公式长度，推理时优化候选公式的常数。采用 TabICLv2 输入先验，在 SRBench 2.0 的 12 个黑箱数据集及物理公式任务上评估。当前输入上限为 2,048 行、11 个特征，短公式也不自动等于领域可解释性。", "url": "https://arxiv.org/abs/2610.17994", "pdf": "https://arxiv.org/pdf/2610.17994", "code": "https://github.com/automl/rasr", "keywords": [ "符号回归", "强化学习", "后训练", "TabICLv2" ], "aliases": [ "RASR" ], "sources": [ "https://arxiv.org/abs/2610.17994", "https://arxiv.org/html/2610.17994", "https://github.com/automl/rasr" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.17874", "title": "Searching for What Matters: Open-Ended Discovery of Multiple Hypotheses with Agentic Data Exploration", "date": "2026-10-13", "updated": "2026-10-13", "category": 3, "section": "causal-discovery", "topicId": "causal-mechanisms", "related": [ 4, 5 ], "topic": "异质机制与结构推理", "status": "preprint", "venue": "预印本", "summary": "HypAgent 通过残差化、候选假设筛选和迭代细化，减少反复发现同一强信号，探索结构化数据中的多个机制。OpenHypBench 包含 15 个合成任务、351 个植入机制，代理只看观测数据，评测器用隐藏结构因果模型检验假设。实验支持多机制发现能力，但不能据此把真实观测关联认定为因果关系。", "url": "https://arxiv.org/abs/2610.17874", "pdf": "https://arxiv.org/pdf/2610.17874", "code": null, "keywords": [ "因果机制", "数据探索", "Agent", "假设发现" ], "aliases": [ "HypAgent", "OpenHypBench" ], "sources": [ "https://arxiv.org/abs/2610.17874", "https://arxiv.org/html/2610.17874" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.16801", "title": "When Repeated Measurements Hurt Tabular Foundation Models", "date": "2026-10-12", "updated": "2026-10-12", "category": 5, "section": "inference-strategies", "topicId": "inference-mechanisms", "related": [ 2, 6 ], "topic": "模型机制与行为分析", "status": "preprint", "venue": "预印本", "summary": "固定输入位置、增加重复测量时，更多上下文可能改善已测位置却损害新输入的预测。研究比较 TabPFN、TabFM、TabDPT、LimiX-2、TabICL 等六个模型，覆盖合成、半合成及 10 个真实重复测量数据集。结果指向重复计入分布证据与注意力尺度；平均测量或调整行数尺度可缓解部分情形，但并非普适修复。", "url": "https://arxiv.org/abs/2610.16801", "pdf": "https://arxiv.org/pdf/2610.16801", "code": "https://github.com/drizmiz/PFN-Repeated-Measurement-Supp", "keywords": [ "TFM", "重复测量", "上下文分布", "LimiX-2" ], "aliases": [ "PFN Repeated Measurements", "Repeated Measurements" ], "sources": [ "https://arxiv.org/abs/2610.16801", "https://arxiv.org/html/2610.16801", "https://github.com/drizmiz/PFN-Repeated-Measurement-Supp" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2606.30336", "title": "FlexTab: A Flexible Encoder-Decoder Architecture for In-Context Learning Across Diverse Tabular Tasks", "date": "2026-06-29", "updated": "2026-10-12", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "related": [ 2, 4, 5 ], "topic": "通用表格模型", "status": "published", "venue": "NeurIPS 2026", "summary": "FlexTab 将不依赖预测目标的行编码器与任务解码器分离，以共享表示连接分类、回归、异常检测、聚类、实体匹配、检索和关系实体分类。评测覆盖 TabArena-Lite、TALENT-Tiny、语义表格及专门任务基准。作者报告了轻度训练与评测数据重合；宽表的单行压缩瓶颈和分阶段训练也限制跨任务结论。", "url": "https://arxiv.org/abs/2606.30336", "pdf": "https://arxiv.org/pdf/2606.30336", "code": "https://github.com/SAP-samples/flextab", "keywords": [ "TFM", "编码器解码器", "TabArena", "实体匹配", "异常检测" ], "aliases": [ "FlexTab", "FlexTab-AD" ], "sources": [ "https://arxiv.org/abs/2606.30336", "https://arxiv.org/html/2606.30336", "https://github.com/SAP-samples/flextab", "https://neurips.cc/virtual/2026/poster/153378" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2606.09004", "title": "LATTEArena: An Evaluation Framework for LLM-powered Tabular Feature Engineering (Extended Version)", "date": "2026-06-08", "updated": "2026-10-12", "category": 1, "section": "industry-benchmarks", "topicId": "semantic-relational-benchmarks", "related": [ 4, 5 ], "topic": "语义与关系数据评测", "status": "preprint", "venue": "预印本", "summary": "LATTEArena 将 15 类 LLM 表格特征工程方法拆为六个设计维度，在统一执行与评估流程中比较 24 种配置，同时记录预测收益、令牌开销和执行成功率，开放逾 4,000 条执行日志。适合分析提示、搜索和历史示例各自的作用；比较的是标准化组件组合，并非逐项原样复现全部原方法，结论受数据筛选与模型预算约束。", "url": "https://arxiv.org/abs/2606.09004", "pdf": "https://arxiv.org/pdf/2606.09004", "code": "https://github.com/goodenhak/LATTEArena", "keywords": [ "LLM", "特征工程", "评测协议", "推理成本" ], "aliases": [ "LATTEArena", "LATTE Arena" ], "sources": [ "https://arxiv.org/abs/2606.09004", "https://arxiv.org/html/2606.09004", "https://github.com/goodenhak/LATTEArena", "https://goodenhak.github.io/LATTEArena/" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2607.03926", "title": "TabQueryBench: A Query-Centric Benchmark for Synthetic Tabular Data", "date": "2026-07-04", "updated": "2026-10-11", "category": 1, "section": "industry-benchmarks", "topicId": "semantic-relational-benchmarks", "related": [ 2, 4 ], "topic": "语义与关系数据评测", "status": "preprint", "venue": "预印本", "summary": "TabQueryBench 从公开 SQL 工作负载提取模板，检验合成表格的查询保真度。四十九个数据集、十一种生成器、四十四模板和每表百余可执行查询，比较局部与全局统计。查询以真实训练表为参照，衡量对源表的模拟效用而非未见分布泛化；限于单表，不含关系连接，也不证明隐私。", "url": "https://arxiv.org/abs/2607.03926", "pdf": "https://arxiv.org/pdf/2607.03926", "code": "https://github.com/TabQueryBench/TabQueryBench", "keywords": [ "合成表格", "SQL", "评测协议", "查询保真度" ], "aliases": [ "TabQueryBench" ], "sources": [ "https://arxiv.org/abs/2607.03926", "https://arxiv.org/html/2607.03926", "https://github.com/TabQueryBench/TabQueryBench" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2609.07629", "title": "Open Tabular Insight Extraction: Where Do We Stand, and Where Should We Go?", "date": "2026-09-07", "updated": "2026-10-12", "category": 1, "section": "industry-benchmarks", "topicId": "tabular-surveys", "related": [ 4, 5 ], "topic": "路线综述", "status": "preprint", "venue": "预印本", "summary": "OpenTI 将跨表检索、数据整合、分析执行与结果解释放到开放数据环境中的同一问题下，系统梳理机器学习、数据库、自然语言处理等领域的 160 项研究。58 个被评估系统中有 41 个不具备表检索能力，许多基准预先暴露文件或列信息。是一份任务与评测框架综述，用于发现端到端分析缺口，并非新模型的性能证明。", "url": "https://arxiv.org/abs/2609.07629", "pdf": "https://arxiv.org/pdf/2609.07629", "code": null, "keywords": [ "综述", "表格检索", "数据分析Agent", "Text-to-SQL" ], "aliases": [ "OpenTI", "Open Tabular Insight Extraction" ], "sources": [ "https://arxiv.org/abs/2609.07629", "https://arxiv.org/html/2609.07629" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.18037", "title": "Weighting-Based Strategies for Classification: Analysis and Local Peak Dynamics with CART Trees", "date": "2026-10-13", "updated": "2026-10-13", "category": 2, "section": "prediction-distributions", "topicId": "probability-calibration", "related": [], "topic": "概率校准与预测一致性", "status": "preprint", "venue": "预印本", "summary": "研究类别不平衡时，样本加权的收益来自学习器变化还是最终阈值变化。在模型复杂度随样本量增长的渐近条件下，加权与不加权可通过阈值平移获得相同表现；对有限深度 CART 的线性回归函数情形给出精确率—召回率曲线分析。更复杂的合成设置仍存在加权收益，并提出数据指示量；不能解释为有限数据上加权普遍无效。", "url": "https://arxiv.org/abs/2610.18037", "pdf": "https://arxiv.org/pdf/2610.18037", "code": null, "keywords": [ "类别不平衡", "CART", "样本加权", "决策阈值" ], "aliases": [ "Weighting-Based Strategies", "Local Peak Dynamics" ], "sources": [ "https://arxiv.org/abs/2610.18037" ], "addedAt": "2026-10-15" }, { "id": "arxiv-2610.22242", "title": "SurvICL: a Foundation Model for in-context Survival Analysis", "date": "2026-10-15", "updated": "2026-10-15", "category": 1, "section": "industry-applications", "topicId": "biomedical-applications", "topic": "生命科学与医疗预测", "related": [ 2, 5 ], "status": "preprint", "venue": "预印本", "summary": "在 TabICLv2 架构上接入删失标记与连续分位数输出，用合成的完整事件时间监督生存预测。作者将 103 个数据集分为 40 个开发集与 63 个留出集，报告留出集上 IBS 总体排名领先；小样本对 Cox 的差异未达统计显著，大样本仍以 SurvivalBoost 更强。适用于右删失单事件数据，尚未覆盖竞争风险与时变协变量。", "url": "https://arxiv.org/abs/2610.22242", "pdf": "https://arxiv.org/pdf/2610.22242", "code": null, "keywords": [ "生存分析", "右删失", "分位数", "TabICLv2", "概率预测" ], "aliases": [ "SurvICL" ], "sources": [ "https://arxiv.org/abs/2610.22242", "https://arxiv.org/html/2610.22242v1" ], "addedAt": "2026-10-17" }, { "id": "arxiv-2610.21950", "title": "SelectPFN: Amortized Bayesian Variable Selection for Sparse Nonlinear Regression with Prior-Data Fitted Networks", "date": "2026-10-15", "updated": "2026-10-15", "category": 1, "section": "industry-models", "topicId": "tabular-releases", "topic": "通用表格模型", "related": [ 2, 5 ], "status": "preprint", "venue": "预印本", "summary": "将变量筛选与预测分成共享骨干的两次计算：先输出变量入选概率，再仅用被选列预测。作者在 7200 个合成任务和四组真实特征矩阵上报告筛选收益；真实响应实验中平均 RMSE 仍比全特征 TabPFN 高约 30%。适合研究稀疏变量选择，精确错误发现率保证依赖有效 knockoff，不能把入选变量直接解释为因果变量。", "url": "https://arxiv.org/abs/2610.21950", "pdf": "https://arxiv.org/pdf/2610.21950", "code": null, "keywords": [ "变量选择", "稀疏回归", "贝叶斯推断", "knockoff", "错误发现率" ], "aliases": [ "SelectPFN", "SelectPFN-L" ], "sources": [ "https://arxiv.org/abs/2610.21950", "https://arxiv.org/html/2610.21950v1" ], "addedAt": "2026-10-17" }, { "id": "arxiv-2610.17503", "title": "Foundation Models for Event Prediction Across Heterogeneous Event Logs", "date": "2026-10-13", "updated": "2026-10-13", "category": 5, "section": "context-selection", "topicId": "retrieval-shots", "topic": "检索与近邻选择", "related": [ 1, 2 ], "status": "preprint", "venue": "预印本", "summary": "FEP 将历史事件序列编码为查询和键，把上下文中已发生的下一事件作为值，通过交叉注意力预测事件类型和间隔。作者在 696 个事件数据集训练、50 个留出系统评估，无目标任务微调时接近强局部模型，并优于所测事件基础模型。当前只利用事件类型和时间，缺少业务属性；延迟实验测的是缓存上下文后的查询成本。", "url": "https://arxiv.org/abs/2610.17503", "pdf": "https://arxiv.org/pdf/2610.17503", "code": null, "keywords": [ "事件日志", "上下文学习", "交叉注意力", "下一事件预测", "结构化事件数据" ], "aliases": [ "FEP", "Foundation Event Predictor" ], "sources": [ "https://arxiv.org/abs/2610.17503", "https://arxiv.org/html/2610.17503v1" ], "addedAt": "2026-10-17" }, { "id": "arxiv-2608.10837", "title": "TACTICL: Task-Aware Compression of Tabular ICL Models", "date": "2026-08-11", "updated": "2026-08-11", "category": 6, "section": "adaptation-distillation", "topicId": "distillation-recipes", "topic": "教师信号与学生比较", "related": [ 1, 5 ], "status": "preprint", "venue": "预印本", "summary": "按目标任务验证集贪心选择 TabPFN 层，用小型 MLP 拟合被替换层，再只微调适配器。在 TabArena 的 47 个任务上，作者报告分类替换 25%–50% 层时 AUC 中位比值至少 0.995，推理约加速 1.25–1.63 倍；回归在 50% 替换时 RMSE 约增 16%。为保留上下文能力的压缩提供路线，但搜层、训练和跨任务迁移成本需另算。", "url": "https://arxiv.org/abs/2608.10837", "pdf": "https://arxiv.org/pdf/2608.10837", "code": "https://github.com/Hebog/tfm_compression", "keywords": [ "TabArena", "TabPFN", "层替换", "适配器", "压缩", "蒸馏" ], "aliases": [ "TACTICL", "Task-Aware Compression" ], "sources": [ "https://arxiv.org/abs/2608.10837", "https://arxiv.org/html/2608.10837v1", "https://github.com/Hebog/tfm_compression" ], "addedAt": "2026-10-17" }, { "id": "arxiv-2605.22335", "title": "Learning Causal Orderings for In-Context Tabular Prediction", "date": "2026-05-21", "updated": "2026-05-21", "category": 3, "section": "causal-discovery", "topicId": "causal-mechanisms", "topic": "异质机制与结构推理", "related": [ 2, 5 ], "status": "published", "venue": "NeurIPS 2026", "summary": "TabOrder 联合学习变量排序与受该排序约束的预测注意力，用缺失项似然训练，无需真值因果图标签。作者在合成因果数据、Sachs 信号数据及 56 个 OpenML 任务测试，报告高缺失率下的补全优势和合成机制干预下的预测稳定性。因果解释依赖可识别性等假设，输出是变量顺序而非完整因果图；实际表格实验另做了训练集微调。", "url": "https://arxiv.org/abs/2605.22335", "pdf": "https://arxiv.org/pdf/2605.22335", "code": null, "keywords": [ "TabOrder", "因果排序", "缺失补全", "注意力约束", "干预" ], "aliases": [ "TabOrder", "Inducing Causal Order through Tabular In-Context Learning" ], "sources": [ "https://arxiv.org/abs/2605.22335", "https://arxiv.org/pdf/2605.22335", "https://neurips.cc/Downloads/2026", "https://openreview.net/forum?id=Htb1iiN3Fx", "https://www.jilles.nl/pubs/" ], "addedAt": "2026-10-17" }, { "id": "arxiv-2605.07840", "title": "RelAgent: LLM Agents as Data Scientists for Relational Learning", "date": "2026-05-08", "updated": "2026-05-08", "category": 4, "section": "semantic-features", "topicId": "feature-search", "topic": "任务反馈与特征搜索", "related": [ 1, 5 ], "status": "published", "venue": "NeurIPS 2026", "summary": "让 LLM 探索关系库、编写 SQL 特征并按验证反馈选择树模型；部署时只执行选定 SQL 和预测器。作者在 RelBench V1/V2 与 4DBInfer 报告有竞争力的分类回归表现，去除字段语义或反馈会退化。每任务多次搜索和复用验证标签带来额外预算与过拟合风险；时间过滤部分依赖提示约束，不能把可读 SQL 等同于已证明无泄漏。", "url": "https://arxiv.org/abs/2605.07840", "pdf": "https://arxiv.org/pdf/2605.07840", "code": "https://anonymous.4open.science/r/RelAgent", "keywords": [ "关系数据", "特征工程", "SQL", "LLM Agent", "验证反馈" ], "aliases": [ "RelAgent" ], "sources": [ "https://arxiv.org/abs/2605.07840", "https://arxiv.org/html/2605.07840v1", "https://neurips.cc/Downloads/2026", "https://jw9730.github.io/publications.html" ], "addedAt": "2026-10-17" }, { "id": "arxiv-2610.00907", "title": "STEER: Reducing Inference Cost in Relational Foundation Models through Semantically Informed Sampling", "date": "2026-10-01", "updated": "2026-10-01", "category": 5, "related": [ 1, 4 ], "topic": "关系上下文采样", "section": "context-selection", "topicId": "representative-context", "status": "preprint", "venue": "预印本", "summary": "用 LLM 根据数据库模式和任务说明给外键边分级，再按概率筛选关系模型的采样路径，每个任务复用同一配置。作者在 8 个 RelBench 数据库、24 项任务及 RT、RT-J、Griffin 上报告上下文平均缩减约四成，多数预测指标接近原模型。代码已公开；上下文规模下降不能直接等同端到端提速，收益依赖模式语义和排序质量。", "url": "https://arxiv.org/abs/2610.00907", "pdf": "https://arxiv.org/pdf/2610.00907", "code": "https://github.com/lids-lab/steer", "aliases": [ "STEER" ], "keywords": [ "RelBench", "RFM", "关系数据", "语义采样", "RT-J", "Griffin" ], "sources": [ "https://arxiv.org/abs/2610.00907", "https://arxiv.org/html/2610.00907v1", "https://github.com/lids-lab/steer" ], "addedAt": "2026-10-18" }, { "id": "arxiv-2610.20514", "title": "Bayesian Learning to Select Fair Classifiers", "date": "2026-10-14", "updated": "2026-10-14", "category": 2, "related": [], "topic": "公平分类器选择", "section": "prediction-decisions", "topicId": "scores-routing", "status": "preprint", "venue": "预印本", "summary": "将固定分类器、LLM 或人工判断者视为候选，用 Thompson Sampling 权衡准确率、公平性差距和查询成本，并处理延迟标签。作者在含 4 个表格任务的 7 个主实验数据集上做离线重采样比较。部分候选阈值及权衡系数网格参考评测总体，不能视为完整部署泛化验证；选择不根据当前样本特征变化，也不提供因果公平保证。", "url": "https://arxiv.org/abs/2610.20514", "pdf": "https://arxiv.org/pdf/2610.20514", "code": null, "keywords": [ "Thompson Sampling", "模型选择", "公平性", "延迟反馈", "多臂老虎机" ], "sources": [ "https://arxiv.org/abs/2610.20514", "https://arxiv.org/html/2610.20514v1" ], "addedAt": "2026-10-18" }, { "id": "arxiv-2609.12136", "title": "The Anatomy and Boundary of Adaptation under Temporal Tabular Shift", "date": "2026-09-10", "updated": "2026-09-10", "category": 6, "related": [ 2, 5 ], "topic": "时间漂移下的冻结模型适配", "section": "adaptation-distribution", "topicId": "subpopulation-shift", "status": "preprint", "venue": "预印本", "summary": "审计冻结表格模型的流式残差校正，区分整体去偏、标签新鲜度、局部校正与排序校准。作者在 8 条 TabReD 数据流上观察到收益与负迁移，并揭示批量近邻混入未来标签的问题，可用于设计严格时间评测。结果来自预计算回放，部分模型选择和门控为回溯分析，复现资料待公开，尚不支持通用在线适配保证。", "url": "https://arxiv.org/abs/2609.12136", "pdf": "https://arxiv.org/pdf/2609.12136", "code": null, "aliases": [ "Temporal Tabular Shift" ], "keywords": [ "TabReD", "时间漂移", "冻结模型", "残差校正", "流式适配", "数据泄漏" ], "sources": [ "https://arxiv.org/abs/2609.12136", "https://arxiv.org/html/2609.12136v1" ], "addedAt": "2026-10-18" }, { "id": "arxiv-2610.22770", "title": "Unit-level aggregation breaks coverage in causal foundation models", "date": "2026-10-16", "updated": "2026-10-16", "category": 3, "related": [ 2, 6 ], "topic": "群体因果效应区间与冻结模型输出头", "section": "causal-models", "topicId": "causal-reliability", "status": "preprint", "venue": "预印本", "summary": "指出独立平均个体效应的边际分布会遗漏跨样本误差相关性，导致群体效应区间过窄；提出共享潜变量及群体输出头。作者在 6 类合成／半合成基准上验证，冻结 CausalPFN 后训练的输出头改善覆盖率，但部分场景仍未达到名义覆盖，点估计也可能变差。可用于群体决策的不确定性审计，不能替代混杂控制或提供通用校准保证。", "url": "https://arxiv.org/abs/2610.22770", "pdf": "https://arxiv.org/pdf/2610.22770", "code": "https://github.com/gabeschulman-nyu/subset-cate-coverage", "aliases": [ "subset-cate-coverage" ], "keywords": [ "CausalPFN", "subset-CATE", "群体效应", "区间覆盖", "误差协方差", "冻结模型", "输出头" ], "sources": [ "https://arxiv.org/abs/2610.22770", "https://arxiv.org/html/2610.22770v1", "https://github.com/gabeschulman-nyu/subset-cate-coverage" ], "addedAt": "2026-10-19" }, { "id": "arxiv-2610.23073", "title": "Path to Prior-Data Fitted Relational Models via Learned Latent Multi-table Representations", "date": "2026-10-16", "updated": "2026-10-16", "category": 1, "related": [ 5 ], "topic": "关系编码器与冻结上下文预测器", "section": "industry-models", "topicId": "relational-releases", "status": "preprint", "venue": "预印本", "summary": "用 Griffin 编码多表关系，再把冻结表征交给冻结 TabICLv2，在目标任务上免梯度更新。作者在 16 项关系分类任务、5 个种子上报告合成训练表征的收益，提供关系编码与上下文预测组合的路线。不同课程的总预算与数据量尚未完全匹配，不能把增益单独归因于课程顺序；尚缺回归和强端到端基线比较。", "url": "https://arxiv.org/abs/2610.23073", "pdf": "https://arxiv.org/pdf/2610.23073", "code": null, "aliases": [ "Griffin + TabICL", "Latent Multi-table Representations" ], "keywords": [ "Griffin", "TabICLv2", "关系数据", "多表表征", "冻结编码器", "上下文学习", "课程学习" ], "sources": [ "https://arxiv.org/abs/2610.23073", "https://arxiv.org/html/2610.23073v1" ], "addedAt": "2026-10-19" }, { "id": "arxiv-2610.23272", "title": "DeDiCa: A Scalable Deconfounding Discrete Causal Generative Model", "date": "2026-10-16", "updated": "2026-10-16", "category": 3, "related": [], "topic": "已知图上的离散因果生成与精确推断", "section": "causal-estimation", "topicId": "treatment-effects", "status": "preprint", "venue": "预印本", "summary": "将已知因果图写成张量化贝叶斯网络，在观测数据上拟合一次后，对满足代理识别条件的干预查询精确边缘化隐混杂。作者在 4 个合成／半合成图基准上报告 CPU 推断效率，可作为因果基础模型的结构化替代方案。优势依赖已知图、离散变量和可识别性；与 CausalPFN 的比较输入信息不同，且须计入逐数据集拟合成本。", "url": "https://arxiv.org/abs/2610.23272", "pdf": "https://arxiv.org/pdf/2610.23272", "code": null, "aliases": [ "DeDiCa" ], "keywords": [ "离散因果模型", "隐混杂", "代理变量", "贝叶斯网络", "张量化", "CausalPFN", "树宽" ], "sources": [ "https://arxiv.org/abs/2610.23272", "https://arxiv.org/html/2610.23272v1" ], "addedAt": "2026-10-19" } ], "collectionUpdates": [ { "date": "2026-10-12", "added": [ "arxiv-2205.12934", "arxiv-2207.01848", "arxiv-2305.03403", "arxiv-2402.11137", "arxiv-2402.14335", "arxiv-2402.16785", "arxiv-2405.13396", "arxiv-2406.05207", "arxiv-2406.19380", "arxiv-2407.20060", "arxiv-2410.18164", "arxiv-2412.08194", "arxiv-2502.05564", "arxiv-2505.14415", "arxiv-2505.18125", "arxiv-2506.05584", "arxiv-2506.08982", "arxiv-2506.10707", "arxiv-2506.16791", "arxiv-2509.03505", "arxiv-2510.02625", "arxiv-2510.21204", "arxiv-2510.25154", "arxiv-2511.02818", "arxiv-2511.07236", "arxiv-2511.08667", "arxiv-2601.04110", "arxiv-2601.09654", "arxiv-2601.22371", "arxiv-2602.03018", "arxiv-2602.04596", "arxiv-2602.05649", "arxiv-2602.11139", "arxiv-2603.12037", "arxiv-2603.15928", "arxiv-2603.26611", "arxiv-2603.29928", "arxiv-2604.11835", "arxiv-2604.12596", "arxiv-2604.21696", "arxiv-2604.25154", "arxiv-2605.04363", "arxiv-2605.05993", "arxiv-2605.06047", "arxiv-2605.06510", "arxiv-2605.12292", "arxiv-2605.12411", "arxiv-2605.12904", "arxiv-2605.13986", "arxiv-2605.15133", "arxiv-2605.18147", "arxiv-2605.18383", "arxiv-2605.18635", "arxiv-2605.18654", "arxiv-2605.18696", "arxiv-2605.19662", "arxiv-2605.21288", "arxiv-2605.27254", "arxiv-2605.31272", "arxiv-2606.04485", "arxiv-2606.05441", "arxiv-2606.07345", "arxiv-2606.08800", "arxiv-2606.11473", "arxiv-2606.15314", "arxiv-2606.18677", "arxiv-2606.26467", "arxiv-2606.30371", "arxiv-2606.30410", "arxiv-2607.01548", "arxiv-2607.11510", "arxiv-2607.24492", "arxiv-2607.27546", "arxiv-2608.01400", "arxiv-2608.03565", "arxiv-2608.06004", "arxiv-2608.16429", "arxiv-2608.17856", "arxiv-2608.25774", "arxiv-2609.03003", "arxiv-2609.03529", "arxiv-2609.03880", "arxiv-2609.04540", "arxiv-2609.04901", "arxiv-2609.07956", "arxiv-2609.13031", "arxiv-2609.17488", "arxiv-2609.17895", "arxiv-2609.21894", "arxiv-2609.22866", "arxiv-2609.23574", "arxiv-2609.27679", "arxiv-2609.28208", "arxiv-2609.29769", "arxiv-2609.31306", "arxiv-2609.36108", "arxiv-2609.36337", "arxiv-2609.37959", "arxiv-2609.37989", "arxiv-2609.39523", "arxiv-2610.01143", "arxiv-2610.01435", "arxiv-2610.01725", "arxiv-2610.05352", "arxiv-2610.06625", "arxiv-2610.06679", "arxiv-2610.06693", "arxiv-2610.06751", "arxiv-2610.07406", "arxiv-2610.07559", "arxiv-2610.07996", "arxiv-2610.08314", "arxiv-2610.08694", "arxiv-2610.10317", "arxiv-2610.10321", "arxiv-2610.11784", "arxiv-2610.12005", "arxiv-2610.12537", "arxiv-2610.13253", "arxiv-2610.13654", "arxiv-2610.13818", "paper-20e0758264", "paper-3b87eab6aa", "paper-612809da03", "paper-62f5871d79", "paper-8380790f74", "paper-93cf6a538b", "paper-9fd1ee8bbf", "paper-a15b9c81ae", "paper-af13c15dd9", "paper-baf74a758c", "paper-d00034280f", "paper-dd682e2bac", "paper-dee7306e6b", "paper-e5eafe3ccc", "report-nvidia-kumo-tabular" ], "updated": [] }, { "date": "2026-10-14", "added": [ "arxiv-2511.02802", "arxiv-2605.06290", "arxiv-2605.10616", "arxiv-2605.11262", "arxiv-2605.28418", "arxiv-2605.30729", "arxiv-2606.02384", "arxiv-2606.03681", "arxiv-2606.08679", "arxiv-2607.26628", "arxiv-2608.16319", "arxiv-2608.17957", "arxiv-2608.18849", "arxiv-2608.18919", "arxiv-2608.20841", "arxiv-2608.27882", "arxiv-2609.16309", "arxiv-2609.33114", "arxiv-2610.08078", "arxiv-2610.09487", "arxiv-2610.10068", "arxiv-2610.10832", "arxiv-2610.12390", "arxiv-2610.12904", "arxiv-2610.13644", "arxiv-2610.13764", "arxiv-2610.14989", "arxiv-2610.14998" ], "updated": [ "arxiv-2511.08667", "arxiv-2605.12292", "arxiv-2605.13986", "arxiv-2606.30410", "arxiv-2607.01548", "arxiv-2608.03565", "arxiv-2609.31306" ] }, { "date": "2026-10-15", "added": [ "arxiv-2606.09004", "arxiv-2606.30336", "arxiv-2607.03926", "arxiv-2609.07629", "arxiv-2610.15699", "arxiv-2610.16763", "arxiv-2610.16801", "arxiv-2610.16990", "arxiv-2610.17065", "arxiv-2610.17874", "arxiv-2610.17994", "arxiv-2610.18037", "arxiv-2610.19814" ], "updated": [ "arxiv-2205.12934", "arxiv-2207.01848", "arxiv-2305.03403", "arxiv-2402.11137", "arxiv-2402.14335", "arxiv-2402.16785", "arxiv-2405.13396", "arxiv-2406.05207", "arxiv-2406.19380", "arxiv-2407.20060", "arxiv-2410.18164", "arxiv-2412.08194", "arxiv-2502.05564", "arxiv-2505.14415", "arxiv-2505.18125", "arxiv-2506.05584", "arxiv-2506.08982", "arxiv-2506.10707", "arxiv-2506.16791", "arxiv-2509.03505", "arxiv-2510.02625", "arxiv-2510.21204", "arxiv-2510.25154", "arxiv-2511.02818", "arxiv-2511.07236", "arxiv-2511.08667", "arxiv-2601.04110", "arxiv-2601.09654", "arxiv-2601.22371", "arxiv-2602.03018", "arxiv-2602.04596", "arxiv-2602.05649", "arxiv-2602.11139", "arxiv-2603.12037", "arxiv-2603.15928", "arxiv-2603.26611", "arxiv-2603.29928", "arxiv-2604.11835", "arxiv-2604.12596", "arxiv-2604.21696", "arxiv-2604.25154", "arxiv-2605.04363", "arxiv-2605.05993", "arxiv-2605.06047", "arxiv-2605.06510", "arxiv-2605.12292", "arxiv-2605.12411", "arxiv-2605.12904", "arxiv-2605.13986", "arxiv-2605.15133", "arxiv-2605.18147", "arxiv-2605.18383", "arxiv-2605.18635", "arxiv-2605.18654", "arxiv-2605.18696", "arxiv-2605.19662", "arxiv-2605.21288", "arxiv-2605.27254", "arxiv-2605.30729", "arxiv-2605.31272", "arxiv-2606.02384", "arxiv-2606.04485", "arxiv-2606.05441", "arxiv-2606.07345", "arxiv-2606.08679", "arxiv-2606.08800", "arxiv-2606.11473", "arxiv-2606.15314", "arxiv-2606.18677", "arxiv-2606.26467", "arxiv-2606.30371", "arxiv-2606.30410", "arxiv-2607.01548", "arxiv-2607.11510", "arxiv-2607.24492", "arxiv-2607.27546", "arxiv-2608.01400", "arxiv-2608.03565", "arxiv-2608.06004", "arxiv-2608.16319", "arxiv-2608.16429", "arxiv-2608.17856", "arxiv-2608.25774", "arxiv-2609.03003", "arxiv-2609.03529", "arxiv-2609.03880", "arxiv-2609.04540", "arxiv-2609.04901", "arxiv-2609.07956", "arxiv-2609.13031", "arxiv-2609.17488", "arxiv-2609.17895", "arxiv-2609.21894", "arxiv-2609.22866", "arxiv-2609.23574", "arxiv-2609.27679", "arxiv-2609.28208", "arxiv-2609.29769", "arxiv-2609.31306", "arxiv-2609.33114", "arxiv-2609.36108", "arxiv-2609.36337", "arxiv-2609.37959", "arxiv-2609.37989", "arxiv-2609.39523", "arxiv-2610.01143", "arxiv-2610.01435", "arxiv-2610.01725", "arxiv-2610.05352", "arxiv-2610.06625", "arxiv-2610.06679", "arxiv-2610.06693", "arxiv-2610.06751", "arxiv-2610.07406", "arxiv-2610.07559", "arxiv-2610.07996", "arxiv-2610.08078", "arxiv-2610.08314", "arxiv-2610.08694", "arxiv-2610.10317", "arxiv-2610.10321", "arxiv-2610.11784", "arxiv-2610.12005", "arxiv-2610.12390", "arxiv-2610.12537", "arxiv-2610.12904", "arxiv-2610.13253", "arxiv-2610.13644", "arxiv-2610.13654", "arxiv-2610.13764", "arxiv-2610.13818", "paper-20e0758264", "paper-3b87eab6aa", "paper-612809da03", "paper-62f5871d79", "paper-8380790f74", "paper-93cf6a538b", "paper-9fd1ee8bbf", "paper-a15b9c81ae", "paper-af13c15dd9", "paper-baf74a758c", "paper-d00034280f", "paper-dd682e2bac", "paper-dee7306e6b", "paper-e5eafe3ccc", "report-nvidia-kumo-tabular" ] }, { "date": "2026-10-16", "added": [ "arxiv-2605.09424", "arxiv-2605.18702", "arxiv-2607.04809", "arxiv-2607.05103", "arxiv-2610.13724", "arxiv-2610.17867", "arxiv-2610.19536", "arxiv-2610.20342", "arxiv-2610.20418", "arxiv-2610.20462" ], "updated": [] }, { "date": "2026-10-17", "added": [ "arxiv-2605.07840", "arxiv-2605.22335", "arxiv-2608.10837", "arxiv-2610.17503", "arxiv-2610.21950", "arxiv-2610.22242" ], "updated": [] }, { "date": "2026-10-18", "added": [ "arxiv-2609.12136", "arxiv-2610.00907", "arxiv-2610.20514" ], "updated": [ "arxiv-2606.11473" ] }, { "date": "2026-10-19", "added": [ "arxiv-2610.22770", "arxiv-2610.23073", "arxiv-2610.23272" ], "updated": [] } ] }