![[2602.16793] Escaping the Cognitive Well: Efficient Competition Math with Off-the-Shelf Models](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16793] Escaping the Cognitive Well: Efficient Competition Math with Off-the-Shelf Models
Summary
The paper presents a novel inference pipeline that leverages off-the-shelf models to solve International Mathematical Olympiad problems efficiently and cost-effectively, outperforming existing methods.
Why It Matters
This research addresses the high costs associated with advanced mathematical problem-solving using AI. By demonstrating that off-the-shelf models can achieve state-of-the-art performance at a fraction of the cost, it opens up new possibilities for accessible AI applications in competitive mathematics and beyond.
Key Takeaways
- Introduces a cost-effective inference pipeline for math problem-solving.
- Achieves superior performance on IMO-style problems compared to existing methods.
- Utilizes insights into grader failures to improve solution accuracy.
- Demonstrates the potential of general-purpose models in specialized applications.
- Reduces average problem-solving costs significantly, making advanced AI more accessible.
Computer Science > Machine Learning arXiv:2602.16793 (cs) [Submitted on 18 Feb 2026] Title:Escaping the Cognitive Well: Efficient Competition Math with Off-the-Shelf Models Authors:Xingyu Dang, Rohit Agarwal, Rodrigo Porto, Anirudh Goyal, Liam H Fowl, Sanjeev Arora View a PDF of the paper titled Escaping the Cognitive Well: Efficient Competition Math with Off-the-Shelf Models, by Xingyu Dang and 5 other authors View PDF HTML (experimental) Abstract:In the past year, custom and unreleased math reasoning models reached gold medal performance on the International Mathematical Olympiad (IMO). Similar performance was then reported using large-scale inference on publicly available models but at prohibitive costs (e.g., 3000 USD per problem). In this work, we present an inference pipeline that attains best-in-class performance on IMO-style math problems at an average inference cost orders of magnitude below competing methods while using only general-purpose off-the-shelf models. Our method relies on insights about grader failure in solver-grader pipelines, which we call the Cognitive Well (iterative refinement converging to a wrong solution that the solver as well as the pipeline's internal grader consider to be basically correct). Our pipeline addresses these failure modes through conjecture extraction, wherein candidate lemmas are isolated from generated solutions and independently verified alongside their negations in a fresh environment (context detachment). On IMO-ProofBench...
