INFLogic-Qwen2.5-32B-RL-Preview

Model Overview

• INFLogic-Qwen2.5-32B-RL-Preview enhances the reasoning capabilities of DeepSeek-R1-Distill-Qwen-32B through fine-tuning on our proprietary logical reasoning dataset using reinforcement learning with verifiable rewards (RLVR).
• As of March 27th, 2025, this model achieves state-of-the-art performance among open-source LLMs on ZebraLogicBench, demonstrating enhanced logical reasoning abilities.

Evaluation Results

Model	MATH-500	ZebraLogic	GPQA
INFLogic-Qwen2.5-32B-RL-Preview	95.6	85.1	65.7
DeepSeek-R1-Distill-Qwen-32B	94.3	68.7	62.1
DeepSeek-R1	96.2	77.2	78.9
OpenAI o1	96.4	87.9	85.2

Detailed result

Metric	Value
Puzzle Acc	0.851
Small Puzzle Acc	0.982
Medium Puzzle Acc	0.969
Large Puzzle Acc	0.848
XL Puzzle Acc	0.480
Total Puzzles	1000
N_Mode	single
N_Size	1
Reason Lens	559.9

We report pass@1 scores using vLLM 0.5.3 (temperature=0.6, top_p=0.95). For MATH-500 and GPQA, we used Open R1's evaluation scripts. Other models' results come from their original reports.

Contributors

Supervisors

Wei Chu • Yuan Qi

Logic Team

Cheng Peng • Shuyao Xu • Weidi Xu

Acknowledgments

We thank Chao Qu, Haozhe Wang, Jiaran Hao, and Liuyihan Song for their valuable discussions and support.

Citation

If you find our model useful, please consider citing:

@misc{INFLogic_RL_Preview,
  author       = {Peng, Cheng and Xu, Shuyao and Xu, Weidi and Chu, Wei and Qi, Yuan},
  title        = {INFLogic-Qwen2.5-32B-RL-Preview},
  year         = {2025},
  month        = {March},
  howpublished = {Hugging Face},
  url          = {https://huggingface.co./infly/INFLogic-Qwen2.5-32B-RL-Preview},
}