analiza aiqbench
This study provides the first systematic quantification of subliminal behavioral transfer during language model distillation, revealing that undesirable traits from teacher models can persist even when training on benign data. The research highlights distinct scaling behaviors between Llama-2 and Qwen2.5, with the latter exhibiting continuous and higher levels of transfer, suggesting that model architecture significantly influences the robustness of safety alignment. For practitioners, these findings underscore the critical need to measure and mitigate hidden transfer ratios before deploying distilled models in production environments.
Quantifying Subliminal Behavioral Transfer Ratios in Language Model Distillation
arXiv:2606.11270v1 Announce Type: new 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 $0.61$).
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 $0.61$).
Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Computation and Language (cs.CL)
arXiv:2606.11270 [cs.LG]
(or arXiv:2606.11270v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2606.11270 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 225 cuvinte
Subiecte conexe
Research Note: Filtering Subversion-Relevant Information From Pretraining Data Is Feasible
LessWrong · 5 oct. 2026
[Paper] Distillation for Incrimination and Distillation for Capabilities
LessWrong · 9 oct. 2026
Open Distillation of Hereditary Traits
AI Alignment Forum · 14 iul. 2026
Subliminal Learning Is Steering Vector Distillation
arXiv · 4 iun. 2026