ThinkRetrieve Retrieves Worked Examples Mid-Reasoning, Boosting Qwen3-1.7B’s AIME 2025 Accuracy by 13.4 Points

A research team released [ThinkRetrieve](https://arxiv.org/abs/2608.10928) on August 11, seeking to address a common failure mode of “letting the model think longer.” Sequential test-time scaling typically adds a prompt such as “think again” after the model finishes a reasoning segment. As the reasoning trajectory grows longer, however, incorrect assumptions may be repeatedly amplified. Instead, at each `</think>` boundary, ThinkRetrieve asks the model to produce a tentative answer, encodes the original problem and tentative answer as vectors, retrieves the most similar problem and its step-by-step solution from a solution database, and inserts them into the next reasoning segment.

The experiments used a retrieval layer comprising E5-Large, FAISS, and approximately 309,600 filtered NuminaMath examples. The researchers tested DeepSeek-R1-Distill-Qwen-1.5B and four Qwen3/Qwen3.5 models ranging from 1.7B to 8B parameters. All example text counted toward the same 22,528-token budget, so the gains did not come from additional context capacity. On AIME 2025, Qwen3-1.7B’s best accuracy rose from 22.2% with sequential scaling to 35.6%. On GSM-8K with a long token budget, the DeepSeek 1.5B baseline fell from 83% to 52%, while ThinkRetrieve maintained approximately 84%. The authors had previously disclosed the diminishing returns of extended reasoning and the direction of this method in a [University of Maryland PhD defense abstract](https://talks.cs.umd.edu/talks/4595).

The key distinction is that ThinkRetrieve retrieves “how to solve” rather than merely supplementing “what is known.” This turns external memory into a reasoning-control signal, making the approach particularly attractive for local small models or fixed-weight deployments. From an engineering perspective, it also allows the retriever, example database, and model to be updated independently.

The limitations are equally clear: the main results are concentrated on mathematics and multiple-choice tasks, while the quality, coverage, and licensing of the solution database directly constrain the domains in which the method can be deployed. Similar problems may also leak solution strategies. Although the authors performed decontamination using exact matching and a 0.90 cosine-similarity threshold, the highest retained similarity still reached 0.898. Future work should examine whether the approach continues to provide benefits for programming or open-ended research tasks, as well as the latency and cost of repeated vector retrieval and long prompts in production services.