![[2506.05316] Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2506.05316] Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay
Summary
This paper presents techniques to enhance data efficiency in reinforcement learning fine-tuning of large language models (LLMs) through difficulty-targeted online data selection and rollout replay.
Why It Matters
Improving data efficiency in LLM fine-tuning is crucial as it reduces resource consumption and enhances model performance. This research addresses a significant gap in existing methodologies, potentially leading to more sustainable AI practices and faster model training.
Key Takeaways
- Introduces adaptive difficulty for online data selection in LLM fine-tuning.
- Demonstrates a 23% to 62% reduction in fine-tuning time without sacrificing performance.
- Proposes a rollout replay mechanism to optimize computational resources.
Computer Science > Machine Learning arXiv:2506.05316 (cs) [Submitted on 5 Jun 2025 (v1), last revised 16 Feb 2026 (this version, v4)] Title:Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay Authors:Yifan Sun, Jingyan Shen, Yibin Wang, Tianyu Chen, Zhendong Wang, Mingyuan Zhou, Huan Zhang View a PDF of the paper titled Improving Data Efficiency for LLM Reinforcement Fine-tuning Through Difficulty-targeted Online Data Selection and Rollout Replay, by Yifan Sun and 6 other authors View PDF HTML (experimental) Abstract:Reinforcement learning (RL) has become an effective approach for fine-tuning large language models (LLMs), particularly to enhance their reasoning capabilities. However, RL fine-tuning remains highly resource-intensive, and existing work has largely overlooked the problem of data efficiency. In this paper, we propose two techniques to improve data efficiency in LLM RL fine-tuning: difficulty-targeted online data selection and rollout replay. We introduce the notion of adaptive difficulty to guide online data selection, prioritizing questions of moderate difficulty that are more likely to yield informative learning signals. To estimate adaptive difficulty efficiently, we develop an attention-based framework that requires rollouts for only a small reference set of questions. The adaptive difficulty of the remaining questions is then estimated based on their similarity to this set. To fu...
