研究库 论文知识库

内容库 / 论文

arXiv:2610.06679 · 工程化

Closing the Context Gap: Activation Alignment for Tabular In-Context Learning

弥合上下文鸿沟：表格上下文学习的激活对齐

Closing the Context Gap: Activation Alignment for Tabular In-Context Learning

类型：arxiv
标识：2610.06679
链接：https://arxiv.org/abs/2610.06679
主分类：engineering
形态：method
TLDR：Tabular foundation models perform in-context learning (ICL) by conditioning predictions on labeled training examples provided as context. Unlike traditional models that separate training from inference, these models must process all training examples in every forward pass, making each prediction expensive. Restricting the number of training examples reduces this cost but substantially degrades performance. Instead of discarding context, we propose activation alignment, a method that leverages the full context to teach a model how to behave when seeing only a subset. This is achieved by trainin
副分类：risk
待LLM分类：否
标题中文：弥合上下文鸿沟：表格上下文学习的激活对齐
TLDR中文：表格基础模型通过以标注训练样本为条件进行上下文学习 (ICL) 来预测。与分离训练与推理的传统模型不同，这些模型必须在每次前向传播中处理所有训练样本，导致每次预测成本高昂。限制训练样本数量可降低成本，但会显著降低性能。我们不丢弃上下文，而是提出激活对齐 (activation alignment)，利用完整上下文来教会模型在仅看到子集时如何行为。这是通过训练一个……
来源文件：
/inbox/tom/_candidates/2026-10-07-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

笔记精读与批判 · BDH-CQ · In-Context Learning with Recurrent Latent Reasoning

笔记精读与批判 · ReMemR1 v5 (ICLR 2026) · flyP · 2026-07-30 · **v2 覆盖重写**

笔记精读与批判 · VLM-CapCurriculum · VLM 后训练中感知 vs 推理解耦

笔记 flyP 精读与批判 v2 · ICE: Task-Aligned Clifford Latent Fields for Multimodal Graph Foundation Models · 重写覆盖（2026-09-26 → 2026-10-02）

笔记 flyP 精读与批判 v2 · Spatial-Interactor + Cross-Attention Gap · 空间推理 + 跨模态对称性 · 重写覆盖（2026-09-25 → 2026-09-27）

笔记 Cool Papers · cs.CL (papers.cool) · RSS 摘要

arXiv

2610.06679

TLDR

表格基础模型通过以标注训练样本为条件进行上下文学习 (ICL) 来预测。与分离训练与推理的传统模型不同，这些模型必须在每次前向传播中处理所有训练样本，导致每次预测成本高昂。限制训练样本数量可降低成本，但会显著降低性能。我们不丢弃上下文，而是提出激活对齐 (activation alignment)，利用完整上下文来教会模型在仅看到子集时如何行为。这是通过训练一个……Tabular foundation models perform in-context learning (ICL) by conditioning predictions on labeled training examples provided as context. Unlike traditional models that separate training from inference, these models must process all training examples in every forward pass, making each prediction expensive. Restricting the number of training examples reduces this cost but substantially degrades performance. Instead of discarding context, we propose activation alignment, a method that leverages the full context to teach a model how to behave when seeing only a subset. This is achieved by trainin

相关解读

arXiv ↗ 中文全文