Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
The paper introduces Influence-Guided PPO (I-PPO), a method that selectively filters PPO rollouts for LLM post-training based on a gradient-based influence score that identifies episodes misaligned with a validation gradient. By removing noisy or unfaithful reasoning episodes, I-PPO achieves superior performance compared to standard SFT and PPO baselines. The method also acts as an intrinsic early stopping mechanism, improving training efficiency and reducing unfaithful chain-of-thought reasoning.
Noisy rollouts are dragging down your PPO fine-tuning: selectively training on high-influence trajectories boosts performance and cuts training time.
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.