ThinkRetrieve: Retrieval-Augmented Reasoning Traces for Test-Time Scaling

2026-08-11

AI Summary

This work addresses the challenge that large reasoning models often fail to improve performance under test-time scaling due to error accumulation and deviation from the correct problem-solving path. To mitigate this, the authors propose ThinkRetrieve, a novel framework that, for the first time, dynamically retrieves step-by-step solutions to similar problems from an external corpus at each reasoning step and injects them into the model’s reasoning trajectory. Rather than merely supplementing factual knowledge, this approach actively guides the reasoning process itself. By integrating retrieval-augmented generation with chain-of-thought reasoning, ThinkRetrieve enables real-time, step-level guidance. Experiments demonstrate that the method significantly outperforms existing test-time scaling approaches on GSM-8K, MATH-500, AIME 2025, and SciQ, achieving a relative accuracy improvement of up to 60% on AIME 2025.

Abstract

Large Reasoning Models (LRMs) improve performance by allocating additional inference-time compute to generate extended chain-of-thought reasoning. However, recent studies reveal that sequential test-time scaling often yields diminishing or even negative returns, as longer traces exhibit increased uncertainty, error compounding, and drift from the original problem. We propose ThinkRetrieve, a test-time scaling framework that augments the reasoning traces of LRMs with dynamically retrieved solved examples at each reasoning step. Given an external corpus of problems paired with step-by-step solutions, ThinkRetrieve retrieves relevant exemplars at each intermediate step and injects them directly into the thinking trace, providing the model with guidance on how to reason rather than merely what facts are relevant. Experiments across five reasoning models (1.5B--8B parameters) on GSM-8K, MATH-500, AIME 2025, and SciQ demonstrate that ThinkRetrieve consistently improves accuracy over standard test-time scaling, with relative gains of up to $60\%$ on AIME 2025.

Problem

• test-time scaling
• reasoning traces
• error compounding
• uncertainty
• reasoning drift

Innovation

• Retrieval-Augmented Reasoning
• Test-Time Scaling
• Chain-of-Thought
• Dynamic Retrieval
• Large Reasoning Models