Salesforce/Elastic-Reasoning

Elastic Reasoning

🚀 Scalable Chain of Thoughts via Elastic Reasoning 🌟

Introduction

We propose Elastic Reasoning, a novel framework for scalable chain of thoughts that explicitly separates reasoning into two phases—thinking and solution—with independently allocated budgets. At test time, Elastic Reasoning prioritize that completeness of solution segments, significantly improving reliability under tight resource constraints. To train models that are robust to truncated thinking, we introduce a lightweight budget-constrained rollout strategy, integrated into GRPO, which teaches the model to reason adaptively when the thinking process is cut short and generalizes effectively to unseen budget constraints without additional training.

Main Takeaways

1. ✂️ Thinking + Solution are explicitly separated with independent budgets — boosting reliability under tight compute constraints.
2. 🧠 Budget-Constrained Rollout: We train models to handle truncated reasoning using GRPO.
3. 📈 Flexible scalability: Robust performance across diverse inference budgets on reasoning benchmarks like AIME and LiveCodeBench.
4. ⚙️ Better performance with fewer tokens: Our trained model generates outputs that are 30% shorter while maintaining (or even improving) accuracy.

Models

E1-Math-1.5B

E1-Code-14B

E1-Math-7B

Performance (Avg@16)

Model	Tokens	Acc (%)	Tokens	Acc (%)	Tokens	Acc (%)	Tokens	Acc (%)	Tokens	Acc (%)
DeepSscaleR-1.5B	10050	41.0	1488	5.2	1904	9.6	2809	15.8	3700	22.7
E1-Math-1.5B	6825	35.0	1340	13.5	1799	17.5	2650	24.8	3377	27.9
Skywork-OR1-Math-7B	13803	68.3	1534	1.0	2047	2.1	3051	7.7	4023	14.0
E1-Math-7B	11768	69.6	1381	16.9	1841	21.3	2799	26.0	3742	32.9

Github

Training and evaluation code

Citation

@article{xu2025scalable,
  title={Scalable Chain of Thoughts via Elastic Reasoning},
  author={Xu, Yuhui and Dong, Hanze and Wang, Lei and Sahoo, Doyen and Li, Junnan and Xiong, Caiming},
  journal={arXiv preprint arXiv:2505.05315},
  year={2025}
}

Ethical Considerations

This release is for research purposes only in support of an academic paper. Our models, datasets, and code are not specifically designed or evaluated for all downstream purposes. We strongly recommend users evaluate and address potential concerns related to accuracy, safety, and fairness before deploying this model. We encourage users to consider the common limitations of AI, comply with applicable laws, and leverage best practices when selecting use cases, particularly for high-risk scenarios where errors or misuse could significantly impact people’s lives, rights, or safety. For further guidance on use cases, refer to our AUP and AI AUP.