PCL-Reasoner-V1.5: Advancing Math Reasoning with Offline Reinforcement Learning

TL;DR

PCL-Reasoner-V1.5 is a large language model specialized in mathematical reasoning, utilizing a novel offline reinforcement learning approach to achieve state-of-the-art accuracy on math benchmarks.

Contribution

The paper introduces an offline RL method for training large language models, improving stability and efficiency over traditional online RL techniques.

Findings

Achieved 90.9% accuracy on AIME 2024

Attained 85.6% accuracy on AIME 2025

Demonstrated offline RL as a stable training paradigm for reasoning models

Abstract

We present PCL-Reasoner-V1.5, a 32-billion-parameter large language model (LLM) for mathematical reasoning. The model is built upon Qwen2.5-32B and refined via supervised fine-tuning (SFT) followed by reinforcement learning (RL). A central innovation is our proposed offline RL method, which provides superior training stability and efficiency over standard online RL methods such as GRPO. Our model achieves state-of-the-art performance among models post-trained on Qwen2.5-32B, attaining average accuracies of 90.9% on AIME 2024 and 85.6% on AIME 2025. Our work demonstrates offline RL as a stable and efficient paradigm for advancing reasoning in LLMs. All experiments were conducted on Huawei Ascend 910C NPUs.