Learning from the Right Rollouts: Data Attribution for PPO-based LLM Post-Training
This addresses a specific bottleneck in RL post-training for LLMs, offering an incremental improvement by enhancing data selection.
The paper tackles the problem of noisy or unfaithful reasoning in PPO-based LLM post-training by proposing Influence-Guided PPO (I-PPO), which uses data attribution to filter out anti-aligned episodes, resulting in consistent outperformance over SFT and PPO baselines with accelerated training efficiency and reduced unfaithful reasoning.
Traditional RL algorithms like Proximal Policy Optimization (PPO) typically train on the entire rollout buffer, operating under the assumption that all generated episodes provide a beneficial optimization signal. However, these episodes frequently contain noisy or unfaithful reasoning, which can degrade model performance and slow down training. In this paper, we propose \textbf{Influence-Guided PPO (I-PPO)}, a novel framework that integrates data attribution into the RL post-training loop. By calculating an influence score for each episode using a gradient-based approximation, I-PPO identifies and eliminates episodes that are anti-aligned with a validation gradient. Our experiments demonstrate that I-PPO consistently outperforms SFT and PPO baselines. We show that our filtering process acts as an intrinsic early stopping mechanism, accelerating training efficiency while effectively reducing unfaithful CoT reasoning.