Our new X account is live! Follow @wizwand_team for updates
WorkDL logo mark

TraPO: A Semi-Supervised Reinforcement Learning Framework for Boosting LLM Reasoning

About

Reinforcement learning with verifiable rewards (RLVR) has proven effective in training large reasoning models (LRMs) by leveraging answer-verifiable signals to guide policy optimization, which, however, suffers from high annotation costs. To alleviate this problem, recent work has explored unsupervised RLVR methods that derive rewards solely from the model's internal consistency, such as through entropy and majority voting. While seemingly promising, these methods often suffer from model collapse in the later stages of training, which may arise from the reinforcement of incorrect reasoning patterns in the absence of external supervision. In this work, we investigate a novel semi-supervised RLVR paradigm that utilizes a small labeled set to guide RLVR training on unlabeled samples. Our key insight is that supervised rewards are essential for stabilizing consistency-based training on unlabeled samples, ensuring that only reasoning patterns verified on labeled instances are incorporated into RL training. Technically, we propose an effective policy optimization algorithm, TraPO, that identifies reliable unlabeled samples by matching their learning trajectory similarity to labeled ones. Building on this, TraPO achieves remarkable data efficiency and strong generalization on six widely used mathematical reasoning benchmarks (AIME24/25, AMC, MATH-500, Minerva, and Olympiad) and three out-of-distribution tasks (ARC-c, GPQA-diamond, and MMLU-pro). With only 1K labeled and 3K unlabeled samples, TraPO reaches 42.6% average accuracy, surpassing the best unsupervised method trained on 45K unlabeled samples (38.3%). Notably, when using 4K labeled and 12K unlabeled samples, TraPO even outperforms the fully supervised model trained on the full 45K labeled samples on all benchmarks, while using only 10% of the labeled data. The code is available via https://github.com/ShenzhiYang2000/TRAPO.

Shenzhi Yang, Guangcheng Zhu, Xing Zheng, Yingfan MA, Zhongqi Chen, Bowen Song, Weiqiang Wang, Junbo Zhao, Gang Chen, Haobo Wang• 2025

Related benchmarks

TaskDatasetResultRank
Mathematical ReasoningMATH 500
pass@184.6
153
Mathematical ReasoningMinerva
Pass@139.3
55
Mathematical ReasoningIn-Distribution Reasoning Performance Suite (AIME, AMC, MATH-500, Minerva, Olympiad)
AIME 2024 Score24.3
30
ReasoningOut-of-Domain Reasoning Suite
ARC-c Score84.6
29
Mathematical ReasoningAIME 2025
Avg@3217.1
27
Mathematical ReasoningCompetition-level Math Benchmarks AIME24, AIME25, AMC23, MATH500, Olympiad, Minerva
AIME 24 Score27.9
21
Science ReasoningGPQA Diamond
Pass@143.9
21
Mathematical ReasoningAMC
Avg@3260
21
Academic ReasoningMMLU-Pro
Pass@150.7
15
Mathematical ReasoningIn-Distribution Avg
Average Score45.6
15
Showing 10 of 13 rows

Other info

GitHub

Follow for update