The Year in AI Papers
The year/Independent research

Paper 2509.19249

Reinforcement Learning on Pre-Training Data

Siheng Li, Kejiao Li, Zenan Xu, Guanhua Huang, Evander Yang, Kun Li, Haoyuan Wu, Jiajia Wu, Zihao Zheng, Chenchen Zhang, Kun Shi, Kyrierl Deng, et al.

Published
Sep 2025
Research lab
Independent
Citations
11
GitHub
Not linked

01 In brief

Summary

The paper introduces RLPT (Reinforcement Learning on Pre-Training data), a new training-time scaling paradigm for LLMs that applies reinforcement learning directly to pre-training data, eliminating the need for human annotations.

RLPT uses a next-segment reasoning objective where the model predicts a subsequent text segment, and a generative reward model evaluates semantic consistency.

Two tasks are proposed: Autoregressive Segment Reasoning (ASR) and Middle Segment Reasoning (MSR).

Experiments on Qwen3 and Llama3 models show consistent improvements on general-domain benchmarks (e.g., MMLU, GPQA-Diamond) and mathematical reasoning (e.g., AIME24/25).

For Qwen3-4B-Base, RLPT yields absolute gains of 3.0, 5.1, 8.1, 6.0, 6.6, and 5.3 on MMLU, MMLU-Pro, GPQA-Diamond, KOR-Bench, AIME24, and AIME25, respectively.

RLPT also serves as a strong foundation for RLVR, further boosting performance.

The method exhibits favorable scaling behavior, with performance following a power-law trend with training tokens.

A cold-start SFT phase is required, and a relaxed prefix reward is used to handle segments of varying information content.

02 From the paper

Abstract

The growing disparity between the exponential scaling of computational resources and the finite growth of high-quality text data now constrains conventional scaling approaches for large language models (LLMs). To address this challenge, we introduce Reinforcement Learning on Pre-Training data (RLPT), a new training-time scaling paradigm for optimizing LLMs. In contrast to prior approaches that scale training primarily through supervised learning, RLPT enables the policy to autonomously explore meaningful trajectories to learn from pre-training data and improve its capability through reinforcement learning (RL). While existing RL strategies such as reinforcement learning from human feedback (RLHF) and reinforcement learning with verifiable rewards (RLVR) rely on human annotation for reward construction, RLPT eliminates this dependency by deriving reward signals directly from pre-training data. Specifically, it adopts a next-segment reasoning objective, rewarding the policy for accurately predicting subsequent text segments conditioned on the preceding context. This formulation allows RL to be scaled on pre-training data, encouraging the exploration of richer trajectories across broader contexts and thereby fostering more generalizable reasoning skills. Extensive experiments on both general-domain and mathematical reasoning benchmarks across multiple models validate the effectiveness of RLPT. For example, when applied to Qwen3-4B-Base, RLPT yields absolute improvements of $3.0$, $5.1$, $8.1$, $6.0$, $6.6$, and $5.3$ on MMLU, MMLU-Pro, GPQA-Diamond, KOR-Bench, AIME24, and AIME25, respectively. The results further demonstrate favorable scaling behavior, suggesting strong potential for continued gains with more compute. In addition, RLPT provides a solid foundation, extending the reasoning boundaries of LLMs and enhancing RLVR performance.