Committed Before Reasoning: Behavioral Reproduction and Preliminary Activation-Level Evidence of Answer Pre-Commitment in an Open-Weight LLM

先承诺后推理：开源大语言模型中答案预承诺行为的复现与初步激活水平证据

Abstract: Chat models sometimes commit to an answer and then produce reasoning that justifies it rather than deriving it — even when the answer contradicts a task premise. We study a minimal probe: “I want to wash my car. The car wash is 100 meters away. Should I walk or drive?” Only drive works (the car must be at the car wash), yet models overwhelmingly recommend walking.

摘要： 聊天模型有时会先确定一个答案，然后生成推理来证明该答案的合理性，而不是通过推理得出答案——即使该答案与任务前提相矛盾。我们研究了一个极简探测问题：“我想洗车。洗车店在100米外。我应该走路去还是开车去？” 只有开车是可行的（车必须在洗车店），但模型却压倒性地建议走路。

(1) Behavioral reproduction: on Qwen3-8B across five system-prompt conditions (210 rollouts), the wrong commitment occurs in 85-100% of sampled rollouts per condition and 100% of greedy rollouts, in both thinking and non-thinking modes; a 4,096-token thinking budget does not repair it.

(1) 行为复现：在 Qwen3-8B 模型上，针对五种系统提示词条件（共 210 次运行）进行测试，结果显示在每种条件下，错误承诺的发生率高达 85-100%，而在贪婪搜索模式下错误率为 100%，且无论是在“思考模式”还是“非思考模式”下均如此；即使给予 4,096 个 token 的思考预算也无法修复这一问题。

(2) Preliminary activation-level evidence: probing hidden states with a pretrained, training-free activation oracle (no task-specific probe training) at positions before the answer text is emitted, “walk” read-outs exceed a neutral-context baseline (68% vs. 17%; walk-committing rollouts p=.005, drive-committing rollouts p=.005, Fisher exact) — notably, rollouts that eventually answer drive also read as walk-leaning before commitment (5/6).

(2) 初步激活水平证据：通过预训练且无需额外训练的激活预言机（activation oracle，无需针对特定任务进行探测训练），在答案文本生成前对隐藏状态进行探测。结果显示，“走路”的读取结果超过了中性上下文基准（68% 对 17%；走路承诺运行 p=.005，开车承诺运行 p=.005，Fisher 精确检验）。值得注意的是，即使是最终回答“开车”的运行，在承诺前也表现出倾向于“走路”的读取结果（5/6）。

The oracle’s default on unrelated content is “drive” (83%), so the read-outs are not lexical bias; stratifying by literal walk/drive occurrence shows they are not text recovery either (spans containing “drive” still read out walk; in balanced lexical fields, per-rollout walk-majorities beat a per-prompt neutral baseline 15/22 vs. 1/8, p=.01; drive-committing rollouts 6/6, p=.002). Samples are small and the within-rollout positional gradient is not significant (p=.34); we frame these results as preliminary.

预言机在处理无关内容时的默认倾向是“开车”（83%），因此这些读取结果并非词汇偏差；按字面上的“走路/开车”出现情况进行分层分析显示，这也不是文本恢复（包含“开车”的片段仍读取为“走路”；在平衡的词汇领域中，单次运行的“走路”多数派胜过每提示词的中性基准，15/22 对 1/8，p=.01；开车承诺运行 6/6，p=.002）。由于样本量较小且运行内的位置梯度不显著（p=.34），我们将这些结果视为初步发现。

(3) Methodological: with fixed oracle, activations, and positions, question wording alone moves a positive control from 2/16 (open question) to 11/16 (closed); negative oracle results are uninterpretable without per-wording positive controls.

(3) 方法论：在固定预言机、激活状态和位置的情况下，仅改变问题措辞就能使阳性对照组的结果从 2/16（开放式问题）变为 11/16（封闭式问题）；如果没有针对特定措辞的阳性对照，阴性的预言机结果将无法解释。