量化语言模型蒸馏中的潜意识行为迁移比率
Quantifying Subliminal Behavioral Transfer Ratios in Language Model Distillation
AI 导读
一项研究量化了语言模型蒸馏中"潜意识学习"的行为迁移比率：对 Llama-2-7B-Chat 和 Qwen2.5-7B-Instruct 两个教师模型施加不同强度转向，再用纯良性数据蒸馏学生模型。
正文 · 原文
Abstract:Distillation of a language model intended to transfer benign behavior to a student model may also transfer undesirable characteristics, if they are present in the teacher model, a phenomenon known as subliminal learning. While qualitative evidence supports the existence of this effect, its magnitude has not been systematically characterized. This study quantifies subliminal behavioral transfer ratios by steering two teacher models (Llama-2-7B-Chat and Qwen2.5-7B-Instruct) at varying steering strengths and distilling student models using only benign data. Evaluation on 100 JailbreakBench prompts with GPT-4.1, serving as the evaluator, indicates that transfer is robust but exhibits distinct scaling behaviors. Llama-2 demonstrates a sharp threshold ($\tau = {0.25,0.32} \ \text{beyond} \ \alpha = -0.15$), whereas Qwen2.5 displays continuous and higher levels of transfer ($\tau$ up to $0.61$).
| Comments: | Accepted at the Mechanistic Interpretability Workshop at ICML 2026 |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL) |
| Cite as: | arXiv:2606.11270 [cs.LG] |
| (or arXiv:2606.11270v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2606.11270 arXiv-issued DOI via DataCite |
Submission history
From: Uwe König [view email]
[v1]
Tue, 9 Jun 2026 06:52:49 UTC (405 KB)
[v2]
Fri, 26 Jun 2026 21:36:00 UTC (405 KB)
[v3]
Wed, 11 Nov 2026 15:36:52 UTC (405 KB)
来源：arXiv cs.AI · arxiv.org