CGRS: Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression

(Coming Soon)

---

🚀 Overview

CGRS (Certainty-Guided Reflection Suppression) is a lightweight, training-free decoding method designed to make Large Reasoning Language Models (LRLMs) think more efficiently.

By dynamically suppressing unnecessary reflection behaviors (e.g., tokens like “Wait”, “But”, “Alternatively”) when the model is confident in its current reasoning trajectory, CGRS effectively mitigates the overthinking problem in LRLMs — reducing redundant reasoning steps, token usage, and inference cost, without sacrificing accuracy.

---

🧩 Key Idea

Large reasoning models like DeepSeek-R1-Distill series, QwQ-32B, and Qwen3 family often reflect excessively while reasoning — revisiting the same ideas even after reaching correct intermediate answers.

CGRS introduces a certainty-guided decoding controller that:

1. Estimates certainty at checkpoints by probing tentative answers and computing token-level entropy.
2. Suppresses reflection triggers probabilistically when the model exhibits high certainty.

This dynamic suppression ensures reasoning stays concise and accurate — thinking just enough.

---

✨ Highlights

• 🔹 Training-free & model-agnostic Works with any autoregressive LRLM; no fine-tuning or architectural changes required.

• 🔹 Certainty-aware reflection control Uses internal confidence signals to guide when to suppress redundant reasoning.

• 🔹 Efficiency with accuracy Achieves 18.5%–41.9% average token reduction across four benchmarks (AIME24, AMC23, MATH500, GPQA-D) while preserving performance.

• 🔹 Compatible with major LRLMs Evaluated on DeepSeek-R1-Distill, Qwen3 (4B–32B), and QwQ-32B models.

---

📊 Experimental Summary

CGRS consistently achieves the best balance between length reduction and accuracy retention among efficient reasoning methods evaluated at the time of release.

---

🛠️ Coming Soon

We are actively preparing the code release and evaluation scripts for public use.

Stay tuned for updates in this repository.

---

📚 Citation

If you find this work helpful, please cite our paper:

@misc{huang2025efficientreasoninglargereasoning,
      title={Efficient Reasoning for Large Reasoning Language Models via Certainty-Guided Reflection Suppression}, 
      author={Jiameng Huang and Baijiong Lin and Guhao Feng and Jierun Chen and Di He and Lu Hou},
      year={2025},
      eprint={2508.05337},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.05337}, 
}

---

🧑‍💻 Authors

• Jiameng Huang, Peking University
• Baijiong Lin, HKUST (Guangzhou)
• Guhao Feng, Peking University
• Jierun Chen, Huawei Technologies Co., Ltd.
• Di He, Peking University
• Lu Hou, Huawei Technologies Co., Ltd.

---

📅 Status

⏳ Repository under active preparation.

Official code and documentation will be released after internal verification. Follow the repository or star ⭐ it to get notified upon release.