TinyV: Reducing False Negatives in Verification Improves RL for LLM Reasoning
Zhangchen Xu, Yuetai Li, Fengqing Jiang, Bhaskar Ramasubramanian, Luyao Niu, Bill Yuchen Lin, Radha Poovendran
TL;DR
This work exposes a pervasive problem: false negatives in rule-based verifiers used for RL fine-tuning of LLMs on mathematical reasoning tasks, which degrade reward quality and slow convergence. It introduces TinyV, a lightweight LLM-based verifier that augments existing verifiers to recover true positives and yield more informative gradients, and it validates its effectiveness on math reasoning benchmarks, including a newly proposed HardVerify-Math benchmark. The authors provide both empirical and theoretical analyses showing that FN-induced reward noise hampers learnability and convergence, and demonstrate that TinyV improves pass rates by up to 10% and accelerates training, with only modest computational overhead. This work offers a practical path to more robust RL-based fine-tuning of LLMs in domains requiring complex verification, by strengthening reward signals with targeted verifier augmentation and richer evaluation benchmarks.
Abstract
Reinforcement Learning (RL) has become a powerful tool for enhancing the reasoning abilities of large language models (LLMs) by optimizing their policies with reward signals. Yet, RL's success relies on the reliability of rewards, which are provided by verifiers. In this paper, we expose and analyze a widespread problem--false negatives--where verifiers wrongly reject correct model outputs. Our in-depth study of the Big-Math-RL-Verified dataset reveals that over 38% of model-generated responses suffer from false negatives, where the verifier fails to recognize correct answers. We show, both empirically and theoretically, that these false negatives severely impair RL training by depriving the model of informative gradient signals and slowing convergence. To mitigate this, we propose tinyV, a lightweight LLM-based verifier that augments existing rule-based methods, which dynamically identifies potential false negatives and recovers valid responses to produce more accurate reward estimates. Across multiple math-reasoning benchmarks, integrating TinyV boosts pass rates by up to 10% and accelerates convergence relative to the baseline. Our findings highlight the critical importance of addressing verifier false negatives and offer a practical approach to improve RL-based fine-tuning of LLMs. Our code is available at https://github.com/uw-nsl/TinyV.