Reproduce “MASteer: Multi-Agent Adaptive Steer Strategy for End-to-End LLM Trustworthiness Repair” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Changqing Li, Tianlin Li, Xiaohan Zhang, Aishan Liu, Li Pan
MASteer introduces an end-to-end framework that leverages multi-agent systems to automate the generation of contrastive steer samples and construct adaptive steering strategies for Large Language Models. The framework enhances LLM trustworthiness across issues like truthfulness, fairness, and safety, showing an average improvement of 15.36% on LLaMA-3.1-8B-Chat while preserving general model capabilities.
Key insights
- Automating steer sample generation via a multi-agent system addresses the manual bottleneck in representation engineering, enabling scalable trustworthiness repair.
- Adaptive selection of steer strategies at inference time, based on anchor vectors and dynamic strengths, is essential for effective and generalizable repair that preserves LLM general capabilities.
- Interventions at middle Transformer layers yield the most stable and significant trustworthiness improvements, as these layers appear to capture more abstract and steerable semantic concepts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.