OpenRS-GRPO

Model Summary

This repository hosts model for the Open RS project, accompanying the paper Reinforcement Learning for Reasoning in Small LLMs: What Works and What Doesn’t. The project explores enhancing reasoning capabilities in small large language models (LLMs) using reinforcement learning (RL) under resource-constrained conditions.

We focus on a 1.5-billion-parameter model, DeepSeek-R1-Distill-Qwen-1.5B, trained on 4 NVIDIA A40 GPUs (48 GB VRAM each) within 24 hours. By adapting the Group Relative Policy Optimization (GRPO) algorithm and leveraging a curated, compact mathematical reasoning dataset, we conducted three experiments to assess performance and behavior. Key findings include:

• Significant reasoning improvements, e.g., AMC23 accuracy rising from 63% to 80% and AIME24 reaching 46.7%, outperforming o1-preview.
• Efficient training with just 7,000 samples at a cost of $42, compared to thousands of dollars for baseline models.
• Challenges like optimization instability and length constraints with extended training.

These results showcase RL-based fine-tuning as a cost-effective approach for small LLMs, making reasoning capabilities accessible in resource-limited settings. We open-source our code, models, and datasets to support further research.

For more details, please refer our github.

Evaluation

Performance Highlights

• Open-RS1: 53.0% avg. score
• Open-RS2: 55.7% avg. score, 80.0% on AMC23
• Open-RS3: 56.3% avg. score, 46.7% on AIME24 (outperforms o1-preview at 44.6%)
• Competitive MATH-500 scores; Minerva lags behind 7B models.

Cost Efficiency

Our approach uses 7,000 samples (42,000 total outputs) and costs ~$42 on 4x A40 GPUs in 24 hours, compared to:

• 7B models: Qwen2.5-7B-SimpleRL ($1,633), Eurus-2-7B-PRIME ($1,088)
• 1.5B models: DeepScaleR-1.5B-Preview ($3,629), Still-3-1.5B-Preview ($2,268)

Citation

If this project aids your work, please cite it as:

@misc{dang2025reinforcementlearningreasoningsmall,
      title={Reinforcement Learning for Reasoning in Small LLMs: What Works and What Doesn't}, 
      author={Quy-Anh Dang and Chris Ngo},
      year={2025},
      eprint={2503.16219},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2503.16219}, 
}