analiza aiqbench
This research addresses a critical bottleneck in LLM safety by proposing GFlowNet fine-tuning to overcome mode collapse in automated red-teaming. The method successfully generates diverse, effective attack prompts that transfer well across various target models, leading to more robust safety tuning. By improving the diversity of adversarial examples, this work directly enhances the resilience of large language models against harmful inputs.
Learning diverse attacks on large language models for robust red-teaming and safety tuning
arXiv:2405.18540v3 Announce Type: replace Abstract: Red-teaming, or identifying prompts that elicit harmful responses, is a critical step in ensuring the safe and responsible deployment of large language models (LLMs). Developing effective protection against many modes of attack prompts requires discovering diverse attacks. Automated red-teaming typically uses reinforcement learning to fine-tune an attacker language model to generate prompts that elicit undesirable responses from a target LLM, as measured, for example, by an auxiliary toxicity classifier. We show that even with explicit regularization to favor novelty and diversity, existing approaches suffer from mode collapse or fail to generate effective attacks. As a flexible and probabilistically principled alternative, we propose to use GFlowNet fine-tuning, followed by a secondary smoothing phase, to train the attacker model to generate diverse and effective attack prompts. We find that the attacks generated by our method are effective against a wide range of target LLMs, both with and without safety tuning, and transfer well between target LLMs. Finally, we demonstrate that models safety-tuned using a dataset of red-teaming prompts generated by our method are robust to attacks from other RL-based red-teaming approaches.
Authors:Seanie Lee, Minsu Kim, Lynn Cherif, David Dobre, Juho Lee, Sung Ju Hwang, Kenji Kawaguchi, Gauthier Gidel, Yoshua Bengio, Esmeralda S. Whitammer, Moksh Jain
Abstract:Red-teaming, or identifying prompts that elicit harmful responses, is a critical step in ensuring the safe and responsible deployment of large language models (LLMs). Developing effective protection against many modes of attack prompts requires discovering diverse attacks. Automated red-teaming typically uses reinforcement learning to fine-tune an attacker language model to generate prompts that elicit undesirable responses from a target LLM, as measured, for example, by an auxiliary toxicity classifier. We show that even with explicit regularization to favor novelty and diversity, existing approaches suffer from mode collapse or fail to generate effective attacks. As a flexible and probabilistically principled alternative, we propose to use GFlowNet fine-tuning, followed by a secondary smoothing phase, to train the attacker model to generate diverse and effective attack prompts. We find that the attacks generated by our method are effective against a wide range of target LLMs, both with and without safety tuning, and transfer well between target LLMs. Finally, we demonstrate that models safety-tuned using a dataset of red-teaming prompts generated by our method are robust to attacks from other RL-based red-teaming approaches.
Computation and Language (cs.CL); Cryptography and Security (cs.CR); Machine Learning (cs.LG)
arXiv:2405.18540 [cs.CL]
(or arXiv:2405.18540v3 [cs.CL] for this version)
Text integral sub licența deschisă a publicației · 222 cuvinte
Subiecte conexe
Links #3: 2026/06 Part 1
LessWrong · 15 iun. 2026
What We Learned Trying to Catch AI Liars: An Aletheia's Quest Retrospective
EleutherAI · 25 aug. 2026
Quality-Diversity Red-Teaming: Automated Generation of High-Quality and Diverse Attackers for Large Language Models
arXiv · 5 aug. 2026
Exposing Long-Tail Safety Failures in Large Language Models through Efficient Diverse Response Sampling
arXiv · 21 iul. 2026