Yifan Zeng

Yifan Zeng, Yiran Wu, Xiao Zhang et al.

Despite extensive pre-training in moral alignment to prevent generating harmful information, large language models (LLMs) remain vulnerable to jailbreak attacks. In this paper, we propose AutoDefense, a multi-agent defense framework that filters harmful responses from LLMs. With the response-filtering mechanism, our framework is robust against different jailbreak attack prompts, and can be used to defend different victim models. AutoDefense assigns different roles to LLM agents and employs them to complete the defense task collaboratively. The division in tasks enhances the overall instruction-following of LLMs and enables the integration of other defense components as tools. With AutoDefense, small open-source LMs can serve as agents and defend larger models against jailbreak attacks. Our experiments show that AutoDefense can effectively defense against different jailbreak attacks, while maintaining the performance at normal user request. For example, we reduce the attack success rate on GPT-3.5 from 55.74% to 7.95% using LLaMA-2-13b with a 3-agent system. Our code and data are publicly available at https://github.com/XHMY/AutoDefense.

Hui Yuan, Yifan Zeng, Yue Wu et al.

Reinforcement Learning from Human Feedback (RLHF) has become the predominant approach for language model (LM) alignment. At its core, RLHF uses a margin-based loss for preference optimization, specifying ideal LM behavior only by the difference between preferred and dispreferred responses. In this paper, we identify a common pitfall of margin-based methods -- the under-specification of ideal LM behavior on preferred and dispreferred responses individually, which leads to two unintended consequences as the margin increases: (1) The probability of dispreferred (e.g., unsafe) responses may increase, resulting in potential safety alignment failures. (2) The probability of preferred responses may decrease, even when those responses are ideal. We demystify the reasons behind these problematic behaviors: margin-based losses couple the change in the preferred probability to the gradient of the dispreferred one, and vice versa, often preventing the preferred probability from increasing while the dispreferred one decreases, and thus causing a synchronized increase or decrease in both probabilities. We term this effect, inherent in margin-based objectives, gradient entanglement. Formally, we derive conditions for general margin-based alignment objectives under which gradient entanglement becomes concerning: the inner product of the gradients of preferred and dispreferred log-probabilities is large relative to the individual gradient norms. We theoretically investigate why such inner products can be large when aligning language models and empirically validate our findings. Empirical implications of our framework extend to explaining important differences in the training dynamics of various preference optimization algorithms, and suggesting potential algorithm designs to mitigate the under-specification issue of margin-based methods and thereby improving language model alignment.