小模型难以真正学会行为倾向，尝试失败但揭示了关键问题。
Disposition Distillation at Small Scale: A Three-Arc Negative Result
- 用四阶段蒸馏法训练小模型的行为倾向
- 所有方法均导致内容损失或风格模仿，无有效提升
- 发现模型自信与正确性严重脱节，适合研究可信生成的团队
我们尝试通过四阶段全MIT蒸馏流程，将自我验证、不确定性承认、反馈整合等行为倾向注入0.6B至2.3B参数的小型语言模型中，并在推理时干预注意力头及使用冻结基座的侧车模块进行实验。初期内部报告称在Qwen3-0.6B上取得MCAS +33.9点和HumanEval +15.3点提升，但二次验证后两者均被证伪：HumanEval增益实为预测长度（n_predict=512）导致的截断误差，在n_predict=1024时转为-8.0点；MCAS增益在严格对照下消失。随后三个路径验证：(1) 三种模型家族在两个领域上的SFT/DPO LoRA；(2) o_proj层推理时注意力头调温；(3) 冻结基座侧车读取最后一隐状态h_last，结果均未提升判断测量的行为倾向，反而损害内容质量或陷入风格模仿。该失败在五种模型（Qwen3-0.6B、Qwen3-1.7B、Qwen3.5-0.8B、Gemma 4 E2B、SmolLM2-1.7B-Instruct）上一致。分布内交叉验证AUC=0.683，新提示下降至0.516（随机水平）。我们贡献了一个三弧负结果及其机制分析，提出线性h_last探测器的双失败模式分类，并建立可复现的虚假阳性转为可发表负结果的验证流程。独立发现：Gemma 4 E2B在厨艺任务中自信与正确性近乎完全解耦（断言不对称性-0.009；模型始终以91%置信度断言）。
原文摘要 · Abstract (English)
We set out to train behavioral dispositions (self-verification, uncertainty acknowledgment, feedback integration) into small language models (0.6B to 2.3B effective parameters) through a four-stage all-MIT distillation pipeline, with follow-on experiments on inference-time attention-head interventions and a frozen-base confidence-gated sidecar. An internal draft reported +33.9-point MCAS and +15.3-point HumanEval gains on a Qwen3-0.6B student; a second-pass sanity check falsified both numbers before publication. The HumanEval delta was a truncation artifact (n_predict=512) that inverted to -8.0 points at n_predict=1024; the MCAS gain disappeared under apples-to-apples scoring. That falsification triggered three subsequent arcs. Across (1) SFT/DPO LoRA on three model families and two domains, (2) inference-time attention-head tempering on o_proj, and (3) a training-free frozen-base sidecar reading the final-token hidden state h_last, we find no operator that moves judge-measured disposition without damaging content or collapsing into stylistic mimicry. The failure is consistent across five models (Qwen3-0.6B, Qwen3-1.7B, Qwen3.5-0.8B, Gemma 4 E2B, and SmolLM2-1.7B-Instruct). A within-distribution cross-validation pass (AUC=0.683) collapsed to chance on fresh prompts (AUC=0.516). We contribute a three-arc negative result with mechanism, a two-failure-mode taxonomy for linear h_last probes, and an honest falsification pipeline that converts the class of false positives we ourselves produced into publishable negatives. As an independent finding, Gemma 4 E2B exhibits near-complete confidence-correctness decoupling on the Chef domain (assertion asymmetry -0.009; the model asserts at 91% regardless of correctness).
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。