ThinkRetrieve: Dynamically Injecting Retrieved Examples into Reasoning Traces Boosts AIME by 60%
_reachsumit · x · 2026-08-12
To address error compounding and drift in long reasoning chains, the paper introduces ThinkRetrieve. This framework dynamically retrieves solved examples at each reasoning step and injects them directly into the trace to guide the model. Experiments show consistent accuracy improvements, with relative gains of up to 60% on AIME 2025.
More from Research
- Anil Seth pushes back on LLM 'pain vector' paper: parsimony says no conscious experience — anilkseth · 2026-09-21
- Anil Seth notes LLM 'pain representations' lack bodily language, unlike human pain — anilkseth · 2026-09-21
- Drammund Adds References on Free Energy Principle vs. Enactivism Debate — mjdramstead · 2026-09-21
- Alignment post argues 'the corrigibility basin of attraction' is a misleading gloss — JacquesThibs · 2026-09-21
- Researcher explains how to build applied multimodal projects: own datasets, metrics, models — mariyaivasileva · 2026-09-21
- As AI reshapes science, scientists' scarce skill shifts to asking the right questions — caglarml · 2026-09-21