New framework SaFeR-Steer boosts LLM safety in multi-turn dialogues

Researchers have introduced SaFeR-Steer, a novel framework designed to enhance the safety and helpfulness of multi-turn Large Language Models (LLMs). This progressive alignment approach utilizes synthetic bootstrapping and a tutor-in-the-loop reinforcement learning technique to train models under adaptive attacks, addressing the mismatch between single-turn training data and real-world multi-turn deployments. The framework also incorporates a Trajectory-Consistent Summative Reward (TCSR) to penalize any low-quality turn within a dialogue. Experiments show significant improvements in safety and helpfulness across various benchmarks when applied to Qwen2.5-VL models. AI

IMPACT This research introduces a method to improve LLM safety in multi-turn conversations, potentially leading to more robust and trustworthy AI assistants.

RANK_REASON The cluster contains an academic paper detailing a new framework and dataset for improving LLM safety. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Haolong Hu
• Qwen2.5-VL-3B
• Qwen2.5-VL-7B
• SaFeR-Steer
• STEER

• paper
• safety
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →