On Designing Effective RL Reward at Training Time for LLM Reasoning

Reward models have been increasingly critical for improving the reasoning capability of LLMs. Existing research has shown that a well-trained reward model can substantially improve model performances at inference time via search. However, the potential of reward models during RL training time still remains largely under-explored. It is currently unclear whether these reward models can provide additional training signals to enhance the reasoning capabilities of LLMs in RL training that uses sparse success rewards, which verify the correctness of solutions. In this work, we evaluate popular reward models for RL training, including the Outcome-supervised Reward Model (ORM) and the Process-supervised Reward Model (PRM), and train a collection of LLMs for math problems using RL by combining these learned rewards with success rewards. Surprisingly, even though these learned reward models have strong inference-time performances, they may NOT help or even hurt RL training, producing worse performances than LLMs trained with the success reward only. Our analysis reveals that an LLM can receive high rewards from some of these reward models by repeating correct but unnecessary reasoning steps, leading to a severe reward hacking issue. Therefore, we introduce two novel reward refinement techniques, including Clipping and Delta. The key idea is to ensure the accumulative reward of any reasoning trajectory is upper-bounded to keep a learned reward model effective without being exploited. We evaluate our techniques with multiple reward models over a set of 1.5B and 7B LLMs on MATH and GSM8K benchmarks and demonstrate that with a carefully designed reward function, RL training without any additional supervised tuning can improve all the evaluated LLMs, including the state-of-the-art 7B LLM Qwen2.5-Math-7B-Instruct on MATH and GSM8K benchmarks.

Is DPO Superior to PPO for LLM Alignment? A Comprehensive Study

Reinforcement Learning from Human Feedback (RLHF) is currently the most widely used method to align large language models (LLMs) with human preferences. Existing RLHF methods can be roughly categorized as either reward-based or reward-free. Novel applications such as ChatGPT and Claude leverage reward-based methods that first learn a reward model and apply actor-critic algorithms, such as Proximal Policy Optimization (PPO). However, in academic benchmarks, state-of-the-art results are often achieved via reward-free methods, such as Direct Preference Optimization (DPO). Is DPO truly superior to PPO? Why does PPO perform poorly on these benchmarks? In this paper, we first conduct both theoretical and empirical studies on the algorithmic properties of DPO and show that DPO may have fundamental limitations. Moreover, we also comprehensively examine PPO and reveal the key factors for the best performances of PPO in fine-tuning LLMs. Finally, we benchmark DPO and PPO across various a collection of RLHF testbeds, ranging from dialogue to code generation. Experiment results demonstrate that PPO is able to surpass other alignment methods in all cases and achieve state-of-the-art results in challenging code competitions.

Learning Zero-Shot Cooperation with Humans, Assuming Humans Are Biased

There is a recent trend of applying multi-agent reinforcement learning (MARL) to train an agent that can cooperate with humans in a zero-shot fashion without using any human data. The typical workflow is to first repeatedly run self-play (SP) to build a policy pool and then train the final adaptive policy against this pool. A crucial limitation of this framework is that every policy in the pool is optimized w.r.t. the environment reward function, which implicitly assumes that the testing partners of the adaptive policy will be precisely optimizing the same reward function as well. However, human objectives are often substantially biased according to their own preferences, which can differ greatly from the environment reward. We propose a more general framework, Hidden-Utility Self-Play (HSP), which explicitly models human biases as hidden reward functions in the self-play objective. By approximating the reward space as linear functions, HSP adopts an effective technique to generate an augmented policy pool with biased policies. We evaluate HSP on the Overcooked benchmark. Empirical results show that our HSP method produces higher rewards than baselines when cooperating with learned human models, manually scripted policies, and real humans. The HSP policy is also rated as the most assistive policy based on human feedback.

Multi-Agent Vulnerability Discovery for Autonomous Driving with Hazard Arbitration Reward

Discovering hazardous scenarios is crucial in testing and further improving driving policies. However, conducting efficient driving policy testing faces two key challenges. On the one hand, the probability of naturally encountering hazardous scenarios is low when testing a well-trained autonomous driving strategy. Thus, discovering these scenarios by purely real-world road testing is extremely costly. On the other hand, a proper determination of accident responsibility is necessary for this task. Collecting scenarios with wrong-attributed responsibilities will lead to an overly conservative autonomous driving strategy. To be more specific, we aim to discover hazardous scenarios that are autonomous-vehicle responsible (AV-responsible), i.e., the vulnerabilities of the under-test driving policy. To this end, this work proposes a Safety Test framework by finding Av-Responsible Scenarios (STARS) based on multi-agent reinforcement learning. STARS guides other traffic participants to produce Av-Responsible Scenarios and make the under-test driving policy misbehave via introducing Hazard Arbitration Reward (HAR). HAR enables our framework to discover diverse, complex, and AV-responsible hazardous scenarios. Experimental results against four different driving policies in three environments demonstrate that STARS can effectively discover AV-responsible hazardous scenarios. These scenarios indeed correspond to the vulnerabilities of the under-test driving policies, thus are meaningful for their further improvements.