ThinkRetrieve: Enhancing LRMs with Retrieval-Augmented Reasoning Traces
A new framework called ThinkRetrieve aims to improve test-time scaling for Large Reasoning Models (LRMs) by dynamically retrieving solved examples at each reasoning step. The approach addresses the issue of diminishing returns from sequential test-time scaling, where longer reasoning traces can lead to increased uncertainty and error compounding. ThinkRetrieve injects relevant exemplars from an external corpus directly into the thinking trace, providing guidance on how to reason rather than just what facts are relevant. Experiments across five reasoning models (1.5B–8B parameters) on benchmarks including GSM-8K, MATH-500, AIME 2025, and SciQ demonstrate the framework's effectiveness. The paper is available on arXiv under the identifier 2608.10928.
Key facts
- ThinkRetrieve is a test-time scaling framework for Large Reasoning Models (LRMs).
- It augments reasoning traces with dynamically retrieved solved examples at each step.
- The method addresses diminishing returns from sequential test-time scaling.
- Experiments were conducted on five reasoning models with 1.5B–8B parameters.
- Benchmarks used include GSM-8K, MATH-500, AIME 2025, and SciQ.
- The paper is available on arXiv with identifier 2608.10928.
- The framework injects exemplars into the thinking trace to guide reasoning.
- The approach aims to reduce error compounding and drift from the original problem.
Entities
Institutions
- arXiv