Adaptive Negative Reinforcement Boosts LLM Reasoning

A new arXiv preprint (2605.07137) introduces Adaptive Negative Sample Reinforcement (A-NSR), an extension of Negative Sample Reinforcement (NSR) for improving reasoning in Large Language Models (LLMs). NSR penalizes incorrect reasoning steps rather than rewarding correct ones, matching or exceeding complex frameworks like PPO and GRPO across the Pass@k spectrum. However, current NSR applies a fixed penalty uniformly. A-NSR uses time-dependent scheduling: early training focuses on error correction to stabilize the model, then shifts to more subtle adjustments. The paper proposes dynamically balancing correction and diversity in reinforcement learning with verifiable rewards (RLVR).