Scalable Chain of Thoughts via Elastic Reasoning
Yuhui Xu, Hanze Dong, Lei Wang, Doyen Sahoo, Junnan Li, Caiming Xiong·May 08, 2025
Summary
Elastic Reasoning offers a scalable framework for large models, addressing output length control. It separates reasoning into thinking and solution phases, allocating budgets independently to prioritize solution completeness. A budget-constrained rollout strategy enhances solution quality under incomplete reasoning, enabling models to adaptively reason. This method outperforms direct truncation and the S1 approach, generalizing effectively across varying budgets. Elastic Reasoning demonstrates robust performance under strict constraints, with lower training costs and produces more concise reasoning.
Introduction
Background
Overview of challenges in managing output length in large models
Importance of scalable reasoning frameworks in AI systems
Objective
To introduce Elastic Reasoning as a solution for output length control in large models
Highlighting the framework's ability to separate reasoning into thinking and solution phases
Method
Data Collection
Description of data sources and types used in Elastic Reasoning
Data Preprocessing
Explanation of preprocessing steps to prepare data for reasoning
Budget Allocation
Detailed explanation of how budgets are allocated for thinking and solution phases
Discussion on how budgets are used to prioritize solution completeness
Budget-Constrained Rollout Strategy
Description of the strategy and its role in enhancing solution quality under incomplete reasoning
Explanation of how the strategy enables models to adaptively reason within budget constraints
Performance Evaluation
Comparison with Direct Truncation and S1 Approach
Overview of the methods being compared
Results showing Elastic Reasoning's superiority in generalizing across varying budgets
Robustness under Constraints
Discussion on Elastic Reasoning's performance under strict constraints
Cost and Conciseness
Analysis of lower training costs associated with Elastic Reasoning
Examination of how the framework produces more concise reasoning
Conclusion
Summary of Elastic Reasoning's contributions
Future directions and potential applications
Recommendations for further research
Basic info
papers
computation and language
machine learning
artificial intelligence
Advanced features