CHASE: Adversarial Red-Blue Teaming for Improving LLM Safety using Reinforcement Learning
Quick Answer
CHASE introduces a co-evolutionary framework for LLM safety, reducing mean StrongREJECT scores by 43.2% with 0% false refusals on benign prompts.
Quick Take
It utilizes to train both attackers and defenders, enhancing resilience against adaptive black-box adversaries.
Key Points
- CHASE employs a closed-loop red-blue teaming approach for safety.
- Achieves 43.2% reduction in mean StrongREJECT scores on benchmark tests.
- Utilizes Group Relative Policy Optimization for training both attackers and defenders.
- Maintains 0% false refusals on benign prompts during evaluations.
- Demonstrates template-free RL exploration for broader attack resilience.
Paper Resources
Source Excerpt
arXiv:2606. 05523v1 Announce Type: new Abstract: Despite advances in safety alignment, prompt-rewriting attacks such as persona modulation, fictional framing and persuasion-based reformulation, can bypass safety filters even on frontier models. Existing defenses either rely on non-scalable human curation or white-box optimisation that overfits to specific model internals, leaving aligned models brittle against the very class of adaptive black-box adversaries they will face in deployment. …
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from arXiv cs.CL
See more →RF-Agent: A Practical Framework for Building Language Agents for RFIC Design
RF-Agent introduces a novel framework for RF circuit design using , creating a unique RF-domain reasoning dataset with over 11,000 samples. The study reveals that domain-specific supervised fine-tuning and semantic retrieval strategies significantly enhance RF reasoning performance, particularly for smaller models.