Reproduce “ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling” on OpenResearch

About this paper

Vaibhav Singh, Soumya Suvra Ghosal, Sarvesh Gharat, Soumyabrata Pal, Ramasuri Narayanam, Dinesh Manocha

ThinkRetrieve, a framework developed by researchers from IIT Bombay, University of Maryland, College Park, and Adobe Research, enhances large reasoning models by dynamically injecting step-by-step solved examples into their reasoning traces at each intermediate step. This approach consistently improves accuracy and maintains stable performance with increased computational budgets, outperforming traditional sequential test-time scaling on various benchmarks like GSM-8K and MATH-500.

Key insights

• Dynamic, in-trace retrieval of structurally similar solved examples provides effective procedural scaffolding, preventing reasoning degradation and enabling more stable performance scaling than unguided self-reflection.
• Conditioning the retrieval process on the model's evolving intermediate answers, in addition to the initial query, significantly enhances the relevance of retrieved exemplars and overall performance.
• The approach reduces the LRM's predictive entropy and stabilizes per-step confidence, indicating that retrieved examples help anchor the model's beliefs and reduce uncertainty during complex reasoning.

Summary by alphaXiv.