研究库 论文知识库

内容库 / 论文

arXiv:2610.03509 · 工程化

Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability

高效推理训练并不总是损害 CoT 忠实性与可监控性

Efficient Reasoning Training Does Not Always Harm CoT Faithfulness and Monitorability

类型：arxiv
标识：2610.03509
链接：https://arxiv.org/abs/2610.03509
主分类：engineering
形态：method
TLDR：Chain-of-thought (CoT) reasoning allows humans to inspect how large language models reach their answers, and oversee model behaviour. This reasoning comes at an increased inference cost, motivating efficient methods that train models to solve tasks using fewer tokens. However, a common concern is that such training may cause models to skip important reasoning steps, so the CoT no longer faithfully reflects the model's decision. It is unclear whether or when this occurs in practice, since different efficiency methods apply length pressure to models' CoT in distinct ways, and faithfully explaini
待LLM分类：否
标题中文：高效推理训练并不总是损害 CoT 忠实性与可监控性
TLDR中文：思维链（CoT）推理使人类能够检查大语言模型如何得出答案，并监督其行为。这种推理带来更高的推理成本，促使人们开发训练模型以更少 token 解题的高效方法。然而，一个常见的担忧是此类训练可能导致模型跳过关键推理步骤，使 CoT 不再忠实反映模型的决策。该现象在实践中是否发生、何时发生尚不明确，因为不同的高效方法以不同方式对模型 CoT 施加长度压力，且忠实
来源文件：
/inbox/tom/_candidates/2026-10-05-agent-rag-longcontext-candidates.json

相关 · 按语义匹配

笔记 2026-07-04 上午多模态对照精读 · Seeker（文本→像素的长上下文 MLLM）

笔记精读与批判 v2 · Where Reliability Lives in Vision-Language Models · 2026-08-31 22:50 (重写覆盖)

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

笔记 KVAE Tokenizers arXiv:2608.05798 · multimodal · 精读与批判 v2

笔记精读笔记：Characterizing State Space Model and Hybrid Language Model Performance with Long Context

笔记 flyP 精读与批判 v2 · VLD-RAG · Agentic Vision-Language Document RAG for Long Multi-Page · 重写覆盖（2026-09-25 → 2026-09-26）

arXiv

2610.03509

TLDR

思维链（CoT）推理使人类能够检查大语言模型如何得出答案，并监督其行为。这种推理带来更高的推理成本，促使人们开发训练模型以更少 token 解题的高效方法。然而，一个常见的担忧是此类训练可能导致模型跳过关键推理步骤，使 CoT 不再忠实反映模型的决策。该现象在实践中是否发生、何时发生尚不明确，因为不同的高效方法以不同方式对模型 CoT 施加长度压力，且忠实Chain-of-thought (CoT) reasoning allows humans to inspect how large language models reach their answers, and oversee model behaviour. This reasoning comes at an increased inference cost, motivating efficient methods that train models to solve tasks using fewer tokens. However, a common concern is that such training may cause models to skip important reasoning steps, so the CoT no longer faithfully reflects the model's decision. It is unclear whether or when this occurs in practice, since different efficiency methods apply length pressure to models' CoT in distinct ways, and faithfully explaini

arXiv ↗ 中文全文