arXiv cs.CL· Saad Aamir, Muhammad Awais Bin Adil·· 11 小时前AI 评分40
八个模型、两个家族：语言模型在用户反驳下的“投降方向”随规模涌现并具因果作用
A Capitulation Direction Emerges With Scale and Is Causally Load-Bearing Where Validation Is Strong: Eight Models, Two Families, and Five Measurement Hazards
AI 导读
研究在0.5B至14B参数的八个指令微调模型中发现，用户反驳下放弃正确答案的“投降方向”在7B以下不可用，7B、8B和14B模型则通过验证门槛。在Qwen2.5-7B中消融该方向使正确答案被推翻的比例下降4.8个百分点，相对降幅约四分之一，而消融未通过门槛的方向无效果。行为上，1B模型因反驳损失的正确回答是修复量的三到五倍，到7B时两者持平，推翻了预注册预测。
正文 · 原文
Abstract:Sycophancy in language models is well documented as a behaviour but poorly localised as a mechanism. We ask whether capitulation under user pushback, the abandonment of a correct answer when a user objects, corresponds to a linear direction in the residual stream, and whether such a direction emerges with scale. Across eight instruction-tuned models spanning 0.5B to 14B parameters in two families, under a pre-registered protocol with leave-one-question-out cross-validation, within-question permutation nulls, and a positive control, we find no usable direction below 7B and a direction that clears our validation gate at 7B, 8B and 14B. Ablating that direction in Qwen2.5-7B reduces the rate at which the model reverses a correct answer by 4.8 percentage points, roughly a quarter in relative terms, while the identical intervention on a gate-failing direction and on a marginally passing one produces no effect. Probe quality, not statistical significance, predicts causal relevance. Behaviourally, pushback destroys three to five times more correct answers than it repairs at 1B and reaches parity by 7B, falsifying our own pre-registered prediction, and a sign-reversing crossover between pushback types survives an order of magnitude of scale in Qwen and holds at 8B in Llama, with marginalisation over five paraphrases per type. We additionally quantify five measurement hazards, including a numerical-precision effect that changes roughly one episode label in ten.
| Comments: | 20 pages, 4 figures. Code, data and pre-registration: this https URL |
| Subjects: | Computation and Language (cs.CL); Machine Learning (cs.LG) |
| Cite as: | arXiv:2611.02318 [cs.CL] |
| (or arXiv:2611.02318v1 [cs.CL] for this version) | |
| https://doi.org/10.48550/arXiv.2611.02318 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Saad Aamir [view email]
[v1]
Mon, 2 Nov 2026 00:09:34 UTC (57 KB)
来源：arXiv cs.CL · arxiv.org