From Detection to Refusal: Safer LLMs via Circuit-Guided Weight Scaling
Source
arxiv.org
Author
Kuan-Lin Chu, Chung-En Sun, Tsui-Wei Weng
Date
Context
A circuit-guided safety study investigates how a LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → turns recognition of harmful input into refusal. Targeted interventions identify components involved in detecting the input, carrying the safety signal and generating the refusal. The researchers then strengthen selected components by scaling existing weights, preserving the architecture.
The edited models improved measured safety against regenerated attacks across six tested models. But five also refused more benign, suspicious-looking prompts. The findings support a targeted way to alter refusal behavior, with attack resistance and unnecessary rejection measured together.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
articleMLLMs Fail to Refuse when Using Tools AgenticallyRikiya Takehi, Ryo Hachiuma, Shaona Ghosh, Dan Zhao, Yu-Chiang Frank Wang, Yusuke Hirota
articleBeyond the Sycophancy Score: How Task, Model, and Pressure Shape LLM YieldingGuang Yang, Homa Hosseinmardi, Fengchen Liu, Amir Ghasemian
articleRefusing the Impossible: A Taxonomy and Benchmark for Code Hallucination in Large Language ModelsVishnu Asutosh Dasu, Ashish Kundu, Gang Tan