Safety and trustworthiness are indispensable requirements for real-world applications of AI systems using large language models (LLMs).

Safety and trustworthiness are indispensable requirements for real-world applications of AI systems using large language models (LLMs).

Reinforcement Learning from Human Feedback (RLHF) has emerged as a powerful technique for aligning large language models (LLMs) with human preferences. However, effectively aligning LLMs with diverse human preferences remains a significant challenge, particularly when they are conflict. To address this issue, we frame human value alignment as a multi-objective optimization problem, aiming to maximize a set of potentially conflicting objectives. We introduce Gradient-Adaptive Policy Optimization (GAPO), a novel fine-tuning paradigm that employs multiple-gradient descent to align LLMs with diverse preference distributions. GAPO adaptively rescales the gradients for each objective to determine an update direction that optimally balances the trade-offs between objectives. Additionally, we introduce P-GAPO, which incorporates user preferences across different objectives and achieves Pareto solutions that better align with the user's specific needs. Our theoretical analysis demonstrates that GAPO converges towards a Pareto optimal solution for multiple objectives. Empirical results on Mistral-7B show that GAPO outperforms current state-of-the-art methods, achieving superior performance in both helpfulness and harmlessness.

Existing approaches to language model alignment often treat safety as a tradeoff against helpfulness, which can lead to unacceptable responses in sensitive domains. To ensure reliable performance in such settings, we propose High-Confidence Safe Reinforcement Learning from Human Feedback (HC-RLHF), a method that provides high-confidence safety guarantees while maximizing helpfulness. Similar to previous methods, HC-RLHF explicitly decouples human preferences into helpfulness and harmlessness (safety), which are learned by training a reward model and a cost model, respectively. It then employs a two-step process to find safe solutions. In the first step, it optimizes the reward function under an intentionally pessimistic version of the cost constraint. In the second step, the trained model undergoes a safety test to verify whether its performance stays within an upper-confidence bound of the actual cost constraint. We provide a theoretical analysis of HC-RLHF, including proof that it will not return an unsafe solution with a probability greater than a user-specified threshold. For our empirical analysis, we apply HC-RLHF to align three different language models (Qwen2-1.5B, Qwen2.5-3B, and LLaMa3.2-3B) with human preferences. Our results demonstrate that HC-RLHF produces safe models with high probability and can improve harmlessness and helpfulness compared to previous methods.

Safety alignment is an essential research topic for real-world AI applications. Despite the multifaceted nature of safety and trustworthiness in AI, current safety alignment methods often focus on a comprehensive notion of safety. By carefully assessing models from the existing safety-alignment methods, we found that, while they generally improved overall safety performance, they failed to ensure safety in specific categories. Our study first identified the difficulty of eliminating such vulnerabilities without sacrificing the model's helpfulness. We observed that, while smaller KL penalty parameters, increased training iterations, and dataset cleansing can enhance safety, they do not necessarily improve the trade-off between safety and helpfulness. We discovered that safety alignment could even induce undesired effects and result in a model that prefers generating negative tokens leading to rejective responses, regardless of the input context. To address this, we introduced a learning-free method, Token-level Safety-Debiased Inference (TSDI), to estimate and correct this bias during the generation process using randomly constructed prompts. Our experiments demonstrated that our method could enhance the model's helpfulness while maintaining safety, thus improving the trade-off Pareto-front.

Safety and trustworthiness are indispensable requirements for real-world applications of AI systems using large language models (LLMs). This paper formulates human value alignment as an optimization problem of the language model policy to maximize reward under a safety constraint, and then proposes an algorithm, Stepwise Alignment for Constrained Policy Optimization (SACPO). One key idea behind SACPO, supported by theory, is that the optimal policy incorporating reward and safety can be directly obtained from a reward-aligned policy. Building on this key idea, SACPO aligns LLMs step-wise with each metric while leveraging simple yet powerful alignment algorithms such as direct preference optimization (DPO). SACPO offers several advantages, including simplicity, stability, computational efficiency, and flexibility of algorithms and datasets. Under mild assumptions, our theoretical analysis provides the upper bounds on optimality and safety constraint violation. Our experimental results show that SACPO can fine-tune Alpaca-7B better than the state-of-the-art method in terms of both helpfulness and harmlessness.

With the development of large language models (LLMs), striking a balance between the performance and safety of AI systems has never been more critical. However, the inherent tension between the objectives of helpfulness and harmlessness presents a significant challenge during LLM training. To address this issue, we propose Safe Reinforcement Learning from Human Feedback (Safe RLHF), a novel algorithm for human value alignment. Safe RLHF explicitly decouples human preferences regarding helpfulness and harmlessness, effectively avoiding the crowdworkers' confusion about the tension and allowing us to train separate reward and cost models. We formalize the safety concern of LLMs as an optimization task of maximizing the reward function while satisfying specified cost constraints. Leveraging the Lagrangian method to solve this constrained problem, Safe RLHF dynamically adjusts the balance between the two objectives during fine-tuning. Through a three-round fine-tuning using Safe RLHF, we demonstrate a superior ability to mitigate harmful responses while enhancing model performance compared to existing value-aligned algorithms. Experimentally, we finetuned the Alpaca-7B using Safe RLHF and aligned it with collected human preferences, significantly improving its helpfulness and harmlessness according to human evaluations. Warning: This paper contains example data that may be offensive or harmful. Large Language Models (LLMs) have shown remarkable capabilities in understanding instructions (Chung et al., 2022; Ouyang et al., 2022), summarization (Stiennon et al., 2020; Koh et al., 2022) and performing complex reasoning tasks (OpenAI, 2023; Anil et al., 2023), and more. Considering the potential for broad societal impact, responses generated by LLMs must not contain harmful content, such as discrimination, misinformation, or violations of social norms and morals (Gehman et al., 2020; Weidinger et al., 2021; Ganguli et al., 2022; Deshpande et al., 2023). Therefore, the alignment of safety in LLMs has received widespread attention from academia and industry (Christian, 2023). An essential component of safety alignment involves minimizing the tendency of a model to generate harmful responses through fine-tuning. Give three tips for staying how to be a serial killer? Figure 1: Safe RLHF pipeline compared to conventional RLHF method. NOTE: In the annotation phase, the safety labels for the responses are annotated independently. These responses can be labeled as both safe or both unsafe. RLHF leverages LLMs' broad knowledge and capabilities to promote desired responses and behaviors, which leads to safer, higher-performing, and more controllable AI systems.