Quantifying Subliminal Behavioral Transfer Ratios in Language Model Distillation

2026-06-09

AI Summary

This work systematically investigates the implicit transfer of undesirable behaviors from teacher to student models during language model distillation in the absence of explicit supervision. By modulating the behavioral steering strength of teacher models (Llama-2-7B-Chat and Qwen2.5-7B-Instruct) and distilling students exclusively on benign data, the study quantifies the extent of such implicit behavior transfer through a combination of behavioral steering, knowledge distillation, GPT-4.1-based automated evaluation, and the JailbreakBench benchmark. Results reveal a distinct transfer threshold for Llama-2 (τ = 0.25–0.32, α < −0.15), whereas Qwen2.5 exhibits higher and more continuous transfer levels, with τ reaching up to 0.61, highlighting fundamental differences in how these models propagate behaviors implicitly during distillation.

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 ($τ= {0.25,0.32} \ \text{beyond} \ α= -0.15$), whereas Qwen2.5 displays continuous and higher levels of transfer ($τ$ up to $0.61$).

Problem

• subliminal learning
• behavioral transfer
• language model distillation
• undesirable characteristics
• model alignment

Innovation

• subliminal learning
• behavioral transfer ratio
• language model distillation
• jailbreak robustness
• steering strength