Princeton University & PKU & UIUC & University of Chicago & ByteDance Seed
🎯 Mission: Building next-generation reasoning capabilities through innovative LLM post-training algorithms focusing on data selection, reinforcement learning, and inference scaling.
Trajectory-aware reward models that provide dense supervision for both offline data selection and online policy optimization in long-CoT reasoning.
Innovative approach where coders and unit testers evolve together through reinforcement learning, creating more robust coding capabilities.
Revolutionary hierarchical reasoning framework that uses thought templates to guide complex problem-solving, achieving SOTA performance with higher efficiency.
Our ReasonFlux-Zero/F1 models are built upon insights from our preliminary work on thought templates—specifically, Buffer of Thoughts (NeurIPS 2024 Spotlight) and SuperCorrect (ICLR 2025). These works introduce high-level, efficient intermediate reasoning patterns that guide and structure the thinking process of large language models.
- [2025/6/23] 🎉 We introduce ReasonFlux-PRM, a family of trajectory-aware process reward models (PRMs) for long CoT reasoning in LLMs. ReasonFlux-PRM is able to support both offline and online reward supervision, by selecting high-quality training data for model distillation, providing dense process-level rewards for policy optimization during reinforcement learning, and enabling reward-guided test-time scaling. Our trained PRMs including ReasonFlux-PRM-7B and ReasonFlux-PRM-1.5B are now available on HuggingFace-GenX. We also release a 7B advanced thinking and reasoning model ReasonFlux-PRM-Qwen-2.5-7B supervised via our PRM.
- [2025/6/04] 🎉 We release our Co-Evolving RL optimized coding LLMs, ReasonFlux-Coder-7B and ReasonFlux-Coder-14B, which outperform similarly sized Qwen Coders and DeepSeek Coders, and naturally fit into common test-time scaling and agentic coding pipelines. We also release our Long-CoT model ReasonFlux-Coder-4B, outperforming Qwen3-4B while achieving 64.8% efficiency in unit test generation.
- [2025/3/24] 🎉We release ReasonFlux-F1-32B, ReasonFlux-F1-14B, ReasonFlux-F1-7B, a series of SOTA-level reasoning LLMs by leveraging the template-augmented reasoning trajectories collected from our ReasonFlux-Zero. For the training and evaluation scripts, please refer to reasonflux-f1/README.md for detail.
- [2025/2/11]🎉We propose ReasonFlux-Zero, a hierarchical LLM reasoning framework that significantly enhances complex reasoning capabilities, outperforming SOTA models like o1-preview and DeepSeek-V3 on challenging MATH and AIME benchmarks.
| Model | Size | Capabilities | Use Cases | Download |
|---|---|---|---|---|
| ReasonFlux-PRM | 7B | • Trajectory-aware scoring • Online/Offline supervision • Dense process rewards |
PRM: Data selection, RL training, Test-time scaling | 🤗 7B |
| ReasonFlux-PRM | 1.5B | • Lightweight scoring • Efficient inference • Edge deployment |
PRM: Resource-constrained applications | 🤗 1.5B |
| ReasonFlux-PRM-Qwen-2.5 | 7B | • Long CoT reasoning • Solving complex tasks and problems |
Tuned Reasoning Model: Math and Science Reasoning | 🤗 7B |
| Model | Size | Specialization | Performance | Download |
|---|---|---|---|---|
| ReasonFlux-Coder | 14B | • Co-evolutionary RL • Advanced coding • Unit test generation |
Outperforms Qwen & DeepSeek Coders | 🤗 14B |
| ReasonFlux-Coder | 7B | • Balanced performance • Efficient inference • Production ready |
Excellent coding capabilities | 🤗 7B |
| ReasonFlux-Coder | 4B | • Long-CoT reasoning • Compact size • Unit test focused |
64.8% efficiency in unit test generation | 🤗 4B |
| Model | Size | Key Features | Best For | Download |
|---|---|---|---|---|
| ReasonFlux-F1 | 7B/14B/32B | • Template-augmented trajectories • Efficient training • Multiple sizes |
General reasoning tasks | 🤗 Models |
| ReasonFlux-Zero | 32B | • Hierarchical reasoning • Template library • Foundation model |
Research & development | 🤗 Model |
| Model | AIME2024@pass1 | AIME2025@pass1 | MATH500@pass1 | GPQA@pass1 |
|---|---|---|---|---|
| QwQ-32B-Preview | 46.7 | 37.2 | 90.6 | 65.2 |
| LIMO-32B | 56.3 | 44.5 | 94.8 | 58.1 |
| s1-32B | 56.7 | 49.3 | 93.0 | 59.6 |
| OpenThinker-32B | 66.0 | 53.3 | 94.8 | 60.1 |
| R1-Distill-32B | 70.0 | 46.7 | 92.0 | 59.6 |
| ReasonFlux-Zero-32B | 56.7 | 37.2 | 91.2 | 61.2 |
| ReasonFlux-F1-32B | 76.7 | 53.3 | 96.0 | 67.2 |
We observe that in the downstream offline data selection + SFT setting, ReasonFlux-PRM-7B surpasses the performance of the high-quality, human-curated s1k dataset. We further visualize the score distributions over 1,000 trajectory-response pairs generated by Deepseek-R1 and Gemini. The clearly separated distributions indicate that ReasonFlux-PRM-7B effectively differentiates the quality of responses from different models, offering a robust and reliable reward signal for high-quality data selection.
Under the online settings, ReasonFlux-PRM-7B also surpasses other PRM and rule-based baselines during the GRPO policy optimization.
@article{yang2025reasonflux,
title={ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates},
author={Yang, Ling and Yu, Zhaochen and Cui, Bin and Wang, Mengdi},
journal={arXiv preprint arXiv:2502.06772},
year={2025}
}
@article{wang2025reasonfluxcoder,
title={Co-Evolving LLM Coder and Unit Tester via Reinforcement Learning},
author={Wang, Yinjie and Yang, Ling and Tian, Ye and Shen, Ke and Wang, Mengdi},
journal={arXiv preprint arXiv:2506.03136},
year={2025}
}
@article{zou2025reasonfluxprm,
title={ReasonFlux-PRM: Trajectory-Aware PRMs for Long Chain-of-Thought Reasoning in LLMs},
author={Zou, Jiaru and Yang, Ling and Gu, Jingwen and Qiu, Jiahao and Shen, Ke and He, Jingrui and Wang, Mengdi},
journal={arXiv preprint arXiv:2506.18896},
year={2025}
}
