ACRL: Adaptive Control of Training-Inference Discrepancy for Stable Reinforcement Learning

Reinforcement Learning (RL) training for Large Language Models (LLMs) often suffers from instability due to the discrepancy between training and inference. This training-inference discrepancy stems from two primary factors: an architectural separation between training and inference engines, and the use of low-precision quantization in inference versus higher-precision computation in training. To address training instability issues caused by high training-inference discrepancy, we present the principles and methods for its adaptive control. We propose Adaptive Control Reinforcement Learning (ACRL), which adaptively maintains the training-inference discrepancy within a reasonable range to ensure stable RL training. Beyond stabilization, ACRL inherently increases policy entropy, thereby enhancing exploration and improving accuracy. The experimental results show that when the inference engine utilizes FP8 quantization, ACRL consistently maintains the training-inference discrepancy within a reasonable range and stabilizes RL training. Furthermore, ACRL not only matches the accuracy of the BF16 baseline but also outperforms importance sampling (IS) fixes.

Rethinking Token-Level Credit Assignment in RLVR: A Polarity-Entropy Analysis

Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved the reasoning ability of Large Language Models (LLMs). However, its sparse outcome-based rewards pose a fundamental credit assignment problem. We analyze this problem through the joint lens of reward polarity and token entropy. Our diagnostic tool, the Four Quadrant Decomposition, isolates token updates by polarity and entropy, and controlled ablations show that reasoning improvements concentrate in the high-entropy quadrants. To justify this observation theoretically, we adapt Conditional Mutual Information to the autoregressive RLVR setting and prove that the credit a token can carry is upper-bounded by its entropy. This view yields testable predictions that reasoning gains arise primarily from high-entropy tokens, with unique roles for positive and negative updates. A gradient analysis of GRPO further reveals how uniform reward broadcast dilutes signal at high-entropy positions while over-crediting deterministic tokens. Grounded in these insights, we propose Entropy-Aware Policy Optimization (EAPO) that modulates token-level learning signals accordingly. Extensive experiments demonstrate that EAPO outperforms strong baselines across two model families.