← 研踪 PaperTrailarXiv:2611.11892 · v1
When Answer Decodability Is Not a Decision: Separating Final-Answer and Imminent-Emission Signals in Chain-of-Thought
首次提交 2026-11-09 13:14
版本更新 2026-11-09 13:14
雷达观测 2026-11-10 13:18
摘要
Answer probes applied to chain-of-thought hidden states are often read as evidence that a language model has already decided its final response. This reading overlooks a competing signal, the candidate the model is about to mention. We introduce emission-controlled dual-target probing (ECDP), which fits the same low-rank reader family separately to the imminent mention and to the final answer at identical positions, with target-specific nulls and thresholds fixed before any fit. On correct DecisionDepth (DD) rollouts in which a distractor is mentioned before the final answer, Qwen3.5-4B decodes the next mention at .966 (null .234) but the final answer only at .221 (null .186), an emission gap of .696 [.648, .744]; Llama-3.1-8B gives .575 [.514, .632]. This pattern is not universal. In a pre-specified evaluation of Qwen3-8B with thinking enabled on held-out MMLU questions, using a full-prefix attention reader over every thinking position, both targets are decodable and the final answer is read far more strongly (mention-minus-answer lift -.398 [-.459, -.336] at layer 22 over 467 questions). Model, task, positions, and reader change together between these settings, and an attempted matched comparison on DD produced no valid statistics because the imminent-mention probe did not train reliably, so we do not identify what reverses the order. On DD, planted-signal controls show that the reader recovers an injected final-answer signal at one tenth of full dose. Answer-candidate decodability is therefore target-relative and protocol-dependent: a strong answer readout does not by itself establish when a model commits, and a probe that fails to train is not evidence that the information is absent. ECDP is a practical control for distinguishing imminent text emission from evidence of a settled decision.
作者备注：20 pages, 6 figures, 3 tables
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-09 13:14 · 最新观测版本
When Answer Decodability Is Not a Decision: Separating Final-Answer and Imminent-Emission Signals in Chain-of-Thought
Answer probes applied to chain-of-thought hidden states are often read as evidence that a language model has already decided its final response. This reading overlooks a competing signal, the candidate the model is about to mention. We introduce emission-controlled dual-target probing (ECDP), which fits the same low-rank reader family separately to the imminent mention and to the final answer at identical positions, with target-specific nulls and thresholds fixed before any fit. On correct DecisionDepth (DD) rollouts in which a distractor is mentioned before the final answer, Qwen3.5-4B decodes the next mention at .966 (null .234) but the final answer only at .221 (null .186), an emission gap of .696 [.648, .744]; Llama-3.1-8B gives .575 [.514, .632]. This pattern is not universal. In a pre-specified evaluation of Qwen3-8B with thinking enabled on held-out MMLU questions, using a full-prefix attention reader over every thinking position, both targets are decodable and the final answer is read far more strongly (mention-minus-answer lift -.398 [-.459, -.336] at layer 22 over 467 questions). Model, task, positions, and reader change together between these settings, and an attempted matched comparison on DD produced no valid statistics because the imminent-mention probe did not train reliably, so we do not identify what reverses the order. On DD, planted-signal controls show that the reader recovers an injected final-answer signal at one tenth of full dose. Answer-candidate decodability is therefore target-relative and protocol-dependent: a strong answer readout does not by itself establish when a model commits, and a probe that fails to train is not evidence that the information is absent. ECDP is a practical control for distinguishing imminent text emission from evidence of a settled decision.
查看此版本解读 →查看此版本原文 ↗