Recognition–Refusal Misalignment in LLMsRecognition–Refusal Misalignment in LLMs
问题:模型明明「知道」一个问题不可答,为什么还是会编一个答案?是表征没编码「不可答」,还是编码了却没接到拒答行为上?
证据位置:论文在隐藏状态里找到一条线性方向,能把可答与「结构上不可能」的数学/代码题分开;识别方向与安全拒答方向接近正交,因此作者把问题解释为路由失败,而不是编码失败。
作者自陈局限:实验集中在数学与代码的结构性不可答题,不是所有「不知道」;线性探针能分开,不等于模型在生成时一定使用这条方向;拒答还受提示、语言与安全策略影响;论文没有给出通用修复方案。
它没主张的:不是所有幻觉都来自路由;不是「加一个拒答头」就能解决;不是对安全对齐的全面评估;不是跨语言、跨领域的普适结论。Problem: if a model represents that a question is unanswerable, why does it still invent an answer? Is the representation missing, or is it encoded but not routed to refusal?
Where the evidence sits: the paper finds a single linear direction in hidden states that separates answerable from structurally impossible math and code prompts; the recognition direction is close to orthogonal to the safety-refusal direction, so the authors explain the failure as routing rather than encoding.
What the authors say about limits: experiments focus on structurally unanswerable math and code, not every kind of “don’t know”; a linear probe separating the directions does not prove the model uses that direction when generating; refusal also depends on prompts, language and safety policy; the paper offers no general fix.
What it does not claim: that every hallucination is a routing failure; that adding a refusal head solves it; a comprehensive evaluation of safety alignment; a universal cross-language, cross-domain conclusion.
关键参数Key specs
- 发表 / venue
- EMNLP 2026 Main
- ACL 编号 / anthology
- 2026.emnlp-main.1078
- 任务 / tasks
- 数学与代码的结构性不可答题
- 关键发现 / finding
- 单一线性方向可分;识别与拒答近正交
- 解释 / explanation
- routing failure,而非 encoding failure
- 复现状态 / reproduction
- 未复现
官方原文怎么说What the source says
a single linear direction in the hidden state separates answerable from structurally impossible math and code prompts
better explained as a routing failure than as an encoding failure
资料与许可Source and licence
- 资料时间Source checked
- 2026-11-18
- 许可 / 类型Licence / kind
- conference-paper
- 许可说明Licence evidence
- https://aclanthology.org/2026.emnlp-main.1078/
- 类目Category
- 推理Reasoning
论文里的数字怎么读How to read the numbers
先分清测什么、怎么算、和谁比、有没有数据污染。拿不准就回原文看实验设置。Before citing a score, separate what it measures, how it is scored, what it compares against and whether contamination is controlled. If unsure, go back to the experiment setup in the paper. 看评测基准See benchmark cards · 看数据集卡See dataset cards