VinePPO: Unlocking RL Potential For LLM Reasoning Through Refined Credit Assignment

2024-10-02

arXiv.org

AI Summary

Reinforcement learning (RL) fine-tuning of large language models (LLMs) for multi-step reasoning tasks suffers from poor performance due to severe credit assignment bias in conventional value networks, leading to inefficient policy updates. We identify that existing value networks exhibit near-random attribution capability in reasoning-intensive settings. Method: We propose VinePPO—a value-network-free, lightweight RL framework—replacing the biased value network with unbiased Monte Carlo policy evaluation driven by environment interaction, and integrating trajectory-level reward attribution with an improved PPO algorithm. Contribution/Results: On MATH and GSM8K, VinePPO significantly outperforms standard PPO and RL-free baselines. It reduces gradient update counts by up to 9× and training time by up to 3.0×, achieving more stable, efficient, and interpretable RL fine-tuning.

Technology Category

• Machine Learning: Reinforcement Learning
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Large language models (LLMs) are increasingly applied to complex reasoning tasks that require executing several complex steps before receiving any reward. Properly assigning credit to these steps is essential for enhancing model performance. Proximal Policy Optimization (PPO), a state-of-the-art reinforcement learning (RL) algorithm used for LLM finetuning, employs value networks to tackle credit assignment. However, value networks face challenges in predicting the expected cumulative rewards accurately in complex reasoning tasks, often leading to high-variance updates and suboptimal performance. In this work, we systematically evaluate the efficacy of value networks and reveal their significant shortcomings in reasoning-heavy LLM tasks, showing that they barely outperform a random baseline when comparing alternative steps. To address this, we propose VinePPO, a straightforward approach that leverages the flexibility of language environments to compute unbiased Monte Carlo-based estimates, bypassing the need for large value networks. Our method consistently outperforms PPO and other RL-free baselines across MATH and GSM8K datasets with fewer gradient updates (up to 9x), less wall-clock time (up to 3.0x). These results emphasize the importance of accurate credit assignment in RL finetuning of LLM and demonstrate VinePPO's potential as a superior alternative.

Problem

• Evaluating value networks' efficacy in LLM credit assignment
• Improving credit assignment for RL training of LLMs
• Proposing VinePPO for better Monte Carlo return estimates

Innovation

• VinePPO replaces value networks with Monte Carlo estimates
• Unbiased credit assignment improves RL training efficiency
• Achieves higher accuracy with less training time