The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning

Xinyu Zhu, Mengzhou Xia, Zhepei Wei, Wei-Lin Chen, Danqi Chen, Yu Meng

arXiv:2506.01347 · 2026-07-27 공개 · arXiv · PDF

mathematical-reasoning llm-reasoning rlvr llama qwen pass-k policy-gradients gradient-analysis

Abstract

Reinforcement learning with verifiable rewards (RLVR) is a promising approach for training language models (LMs) on reasoning tasks that elicit emergent long chains of thought (CoTs). Unlike supervised learning, it updates the model using both correct and incorrect samples via policy gradients. To better understand its mechanism, we decompose the learning signal into reinforcing correct responses and penalizing incorrect ones, referred to as Positive and Negative Sample Reinforcement (PSR and NSR), respectively. We train Qwen2.5-Math-7B, Qwen3-4B and Llama-3.1-8B-Instruct on a mathematical reasoning dataset and uncover a surprising result: training with only negative samples -- without reinforcing correct responses -- can be highly effective: it consistently improves performance over the base model across the entire Pass@$k$ spectrum $k$ up to 256), often matching or surpassing PPO and GRPO. In contrast, reinforcing only correct responses improves Pass@1 but degrades performance at higher $k$, due to reduced diversity. These inference-scaling trends highlight that solely penalizing incorrect responses may contribute more to performance than previously recognized. Through gradient analysis, we show that NSR works by suppressing incorrect generations and redistributing probability mass toward other plausible candidates, guided by the model's prior beliefs. It refines the model's existing knowledge rather than introducing entirely new behaviors. Building on this insight, we propose a simple variant of the RL objective that upweights NSR, and show that it consistently improves overall Pass@$k$ performance on MATH, AIME 2025, and AMC23. Our code is available at https://github.com/TianHongZXY/RLVR-Decomposed.

한국어 요약

한 줄 요약

음성 샘플 강화(NSR)만을 사용해도 LLM의 추론 성능을 크게 향상시킬 수 있음을 밝힌 연구.

핵심 기여도

핵심 아이디어

기존 RLVR(Verification-based RL)은 정답과 오답 모두를 사용해 모델을 업데이트하지만, 그 내부 메커니즘은 명확하지 않았다. 본 연구는 RLVR을 PSR(Positive Sample Reinforcement)과 NSR(Negative Sample Reinforcement)로 분해하여 각각의 영향을 분석했다. 연구 결과, NSR만을 사용해 학습하더라도 모델 성능을 크게 향상시킬 수 있음을 밝혔다. 이는 NSR가 오답 생성을 억제하고, 모델의 사전 확률에 기반한 타당한 대안을 재분배함으로써 기존 지식을 정제하는 방식으로 작동하기 때문이다. 반면, PSR는 정답 경로를 강조하면서 대안 솔루션을 억제해 추론 다양성을 감소시키는 경향이 있다. 이는 Pass@k(k ≥ 16) 성능에 부정적인 영향을 미친다.

기술적 접근법

주요 결과

의의 및 한계

실용적 활용