← Back to all news
arXiv Machine Learning

Probing and Steering for Safety-Relevant Behaviours in Large Language Models

The paper introduces activation probes—directions in a language model’s residual stream—that can both detect and steer safety‑relevant behaviors such as deception, high‑stakes awareness, and sycophancy. By evaluating different estimators (difference‑of‑means vs. logistic regression) and token‑pooling schemes (last, mean, learned attention) on Llama‑3.1‑8B‑Instruct, the authors find that logistic regression excels at detection while difference‑of‑means excels at steering, except for sycophancy where the pattern reverses. The study highlights that the optimal pooling method, layer choice, and steering metric vary by behavior and data, and provides exported probes that enable an interactive interface for scoring and steering text.

Related stories

arXiv AI
Oct 12

Attention Sees, MLP Refuses: Tracing the Attention-to-MLP Safety Pathway in LLMs

The paper investigates how safety‑aligned large language models (LLMs) can still be tricked by jailbreaks or over‑refuse benign requests. By analyzing residual‑stream updates in decoder‑only Transformers, the authors find that harmfulness is more readily decoded from attention updates—especially in early layers—while the MLP component is primarily responsible for writing refusal responses. Post‑training enhances MLP‑driven refusal writing without improving harmfulness decodability, and causal interventions confirm that downstream MLP updates are crucial for the refusal shift caused by upstream attention changes. Building on these insights, the authors propose Sentinel, a system that combines an attention‑derived perception score with an MLP‑derived decision score to better detect jailbreaks and over‑refusals across multiple models.

By Zixuan Weng, Haoqi He, Weifei Jin, Yue Huang, Yang Bai
llmssafety
More like this →