Back to AI Research

AI Research

ThinkRetrieve: Retrieval-Augmented Reasoning Traces... | AI Research

Key Takeaways

  • ThinkRetrieve is a framework designed to improve the performance of Large Reasoning Models (LRMs) by injecting relevant, solved examples into their reasoning...
  • 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.
  • ThinkRetrieve is a framework designed to improve the performance of Large Reasoning Models (LRMs) by injecting relevant, solved examples into their reasoning process.
Paper AbstractExpand

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.

ThinkRetrieve is a framework designed to improve the performance of Large Reasoning Models (LRMs) by injecting relevant, solved examples into their reasoning process. While LRMs typically improve by "thinking longer" through extended chain-of-thought generation, this approach often leads to errors, repetitive cycles, and drift from the correct solution. ThinkRetrieve addresses this by providing models with procedural guidance from an external library of solved problems at each step of their reasoning.

How ThinkRetrieve Works

The framework operates by interleaving the model's standard reasoning process with dynamic retrieval. When a model reaches a reasoning step, it is prompted to generate an intermediate answer. This answer, combined with the original problem, is used to search an external corpus of solved examples. The most relevant example—containing both the question and a step-by-step solution—is then inserted directly into the model's reasoning trace. This allows the model to use the retrieved example as a scaffold to verify its own logic and course-correct before proceeding. Unlike traditional Retrieval-Augmented Generation (RAG), which retrieves information once at the start, ThinkRetrieve performs this retrieval iteratively throughout the thinking process.

Experimental Results

Researchers evaluated ThinkRetrieve across five reasoning models (ranging from 1.5B to 8B parameters) using four benchmarks: GSM-8K, MATH-500, AIME 2025, and SciQ. The results show that while standard sequential test-time scaling often suffers from diminishing or negative returns as the "thinking budget" increases, ThinkRetrieve maintains stable or improving accuracy. On the AIME 2025 benchmark, the framework achieved relative accuracy gains of up to 60% compared to standard scaling methods.

Why This Matters

The findings suggest that the quality of a model's reasoning trace is more important than the sheer length of its "thinking" time. By providing procedural examples, the model is less likely to experience the error compounding and uncertainty that typically plague long reasoning chains. Franklin analysis of the provided data indicates that ThinkRetrieve effectively converts additional compute into more reliable reasoning by grounding the model's internal state in concrete, successful problem-solving strategies.

Considerations

The effectiveness of ThinkRetrieve relies on an external corpus of high-quality, solved examples. The authors note that the framework uses a synthetic corpus of approximately 309,609 examples, which were filtered to ensure no overlap with the evaluation benchmarks. Because the retrieved examples count toward the model's total token budget, the performance gains are attributed to the quality of the retrieved guidance rather than simply providing the model with more compute time.

Comments (0)

No comments yet

Be the first to share your thoughts!