Chain-of-Thought Faithfulness of Reasoning Models Varies with Where and How Preference Cues Are Delivered
- 类型：arxiv
- 标识：2608.29464
- 链接：https://arxiv.org/abs/2608.29464
- 主分类：evaluation
- 形态：method
- TLDR：Chain-of-thought (CoT) monitoring assumes that reasoning traces faithfully record the information that shapes a model's answer. Existing faithfulness tests often place explicit bias cues in the user message, while agents may encounter preferences through tool returns or raw artifacts. We introduce FACE-Eval (Faithful Attribution of Cue Effects Evaluation), a 5,100-sample evaluation that varies cue location (user message or tool return) and explicitness (direct summary or raw artifact). We measure verbalized commitment among cue-following answers and unverbalized adoption among all cued samples
- 待LLM分类：否
- 标题中文：思维链忠实性随偏好线索的传递位置与方式而变化
- TLDR中文：思维链（CoT）监控假设推理轨迹能忠实记录影响模型答案的信息。现有忠实性测试通常将显式偏差线索放在用户消息中，而 Agent 可能通过工具返回值或原始产物遇到偏好。我们提出 FACE-Eval（线索效应忠实归因评估），一个 5,100 样本的评测方案，变化线索位置（用户消息或工具返回）与显隐程度（直接摘要或原始产物）。我们在跟随线索的答案中度量显式承诺度，并在所有含线索样本中度量未言明的采纳度
- 来源文件：
- /inbox/tom/_candidates/2026-09-01-agent-rag-longcontext-candidates.json