> ## Content Index > Fetch the complete content index at: https://www.notatechguy.com/llms.txt > Use this file to discover other available public pages before exploring further. # Euclid-Omni: AI for Olympiad geometry with far less compute - URL: https://www.notatechguy.com/euclid-omni-ai-for-olympiad-geometry-with-far-less-compute/ - Published: 2026-08-18T18:26:42.000Z - Updated: 2026-08-18T18:26:42.000Z - Description: Euclid-Omni couples LLMs and vision models with a formal geometry solver to match state-of-the-art on Olympiad proofs using less compute. - Author: Marcello Babbili - Tags: Technology & AI, AI Models A research paper posted to arXiv this week describes a framework called Euclid-Omni that claims to match state-of-the-art systems on Olympiad-level geometry proofs while using orders of magnitude less compute and training data [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). The system pairs large language models with a formal symbolic solver, an approach AI researchers have pursued for decades. Whether this particular combination of neural pattern-matching and mathematical rigor holds together at competition difficulty is what the next few months will test. **My read:** The neuro-symbolic pitch is the right one. Pure language models hallucinate geometric facts because they pattern-match, not derive. Pure symbolic solvers are rigorous but can't read a diagram or parse a natural-language problem. Welding them together is the obvious move, but execution is where these projects usually fall apart. The claim about orders-of-magnitude less compute is the one I'm most skeptical of, because the paper doesn't name which systems it's comparing against or give exact numbers. If Euclidea, the symbolic solver at the heart of this, genuinely generates clean deductive chains that an LLM can verify, that's a real contribution. But "competitive with state-of-the-art" is a carefully chosen phrase. It doesn't say "beats." I'd want to see independent benchmarks before believing the efficiency claim. ## Why geometry breaks language models Euclidean geometry is a three-part problem. You have to look at a diagram and understand what it shows. You have to reason from axioms, step by step, to reach a conclusion. And you have to do algebra along the way, calculating angles, lengths, and areas [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). Most AI systems handle one or two of these well but stumble on the third. Language models are good at reading problem text and producing plausible-sounding proofs, but they invent facts about the diagram. Symbolic solvers are rigorous but can't parse a natural-language problem statement or interpret a visual. The authors, from the University of Toronto, UC Berkeley, and ETH Zürich [P²](https://arxiv.org/html/2608.14585?ref=notatechguy.com), say existing methods usually cover only some of these skills and falter on competition-level problems [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). ## The two-engine design Euclid-Omni connects a formal geometry engine to large language models (LLMs) and vision-language models (VLMs) [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). The neural models handle the parts they are good at: reading the problem, understanding the diagram, and translating between formal and natural language. The formal system handles the parts that require rigor: deductive inference and algebraic computation. At the centre is a symbolic solver called Euclidea, which automatically generates reasoning steps through formal deduction and algebra [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). The framework tackles both calculation-style problems (find the angle) and proving-style problems (show that these triangles are congruent), in both formal and natural languages, up to Olympiad-level difficulty [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). ## Synthetic data as the multiplier The team built a data-generation pipeline that creates symbolic geometry problems and their solutions, renders diagrams, and translates everything into natural language [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). This pipeline generates large, varied datasets for training LLMs and VLMs across many reasoning settings [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). The authors report that VLMs trained on their synthetic data do better on calculation tasks [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). LLMs paired with Euclidea match state-of-the-art systems on Olympiad-level proving problems, while using far less compute and training data, by orders of magnitude [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). The code and scripts are publicly available on GitHub [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). ## What to do about it If you build automated tutoring tools or educational AI, this is worth a close look. A platform that helps students prepare for math competitions needs exactly what Euclid-Omni claims: the ability to read a geometry problem, understand the diagram, produce a step-by-step solution, and do so without inventing intermediate facts. A tutoring company that currently relies on human graders for geometry proofs could test whether the open-source code can generate reliable worked solutions for textbook problems, where the answer is known and can be checked. The synthetic data pipeline is separately useful: if you train models on mathematical content, a system that generates diverse geometry problems with verified solutions could fill a gap in your training data. The practical step this week: clone the repository, run the Euclidea solver on a few known Olympiad problems, and see whether the deductive chains it produces are actually correct. That will tell you more than any abstract claim. ## What we don't know yet The paper is a preprint on arXiv, and none of the experimental claims have been independently peer-reviewed [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). The abstract does not specify which state-of-the-art systems were used as benchmarks, and it provides no quantitative scores or exact compute comparisons [S¹](https://arxiv.org/abs/2608.14585?ref=notatechguy.com). The phrase "competitive with state-of-the-art" could mean anything from "marginally behind" to "neck and neck." The claim about orders-of-magnitude less compute is self-reported and lacks the kind of controlled comparison that would make it convincing. The paper is also under review at ICLR 2026, according to its OpenReview listing [P³](https://openreview.net/forum?id=1GQv7jhmtV&ref=notatechguy.com), which means it has not yet passed peer review. The next signal: ICLR 2026 review decisions, which will reveal whether independent reviewers found the claims and experiments convincing. We'll check this framework against the review scores and reviewer comments when they appear. Subscribe and we'll send you the ICLR review results when they land. --- *Sources: [S1 — Euclid-Omni : A Unified Neuro-Symbolic Framework for Plane Geometry](https://arxiv.org/abs/2608.14585?ref=notatechguy.com) · [P2 — Euclid-Omni: A Unified Neuro-Symbolic Framework for Plane Geometry](https://arxiv.org/html/2608.14585?ref=notatechguy.com) · [P3 — Euclid-Omni: A Unified Neuro-Symbolic Framework for Geometry Problem S](https://openreview.net/forum?id=1GQv7jhmtV&ref=notatechguy.com) · [P4 — EUCLID-OMNI: A](https://openreview.net/pdf?id=1GQv7jhmtV&ref=notatechguy.com) · [P5 — tianyi-lab/NeSyS](https://github.com/tianyi-lab/NeSyS?ref=notatechguy.com)* ## Related reading - [AI uncertainty measures unified in new risk decomposition paper](https://www.notatechguy.com/ai-uncertainty-measures-unified-in-new-risk-decomposition-paper/) — our technology desk, 2026-07-18 - [Vulnerability scanner divergence explained in new arXiv framework](https://www.notatechguy.com/vulnerability-scanner-divergence-explained-in-new-arxiv-framework/) — our technology desk, 2026-07-17 - [LLM framework automates adversary emulation at 84% success](https://www.notatechguy.com/llm-framework-automates-adversary-emulation-at-84-success/) — our technology desk, 2026-07-17 --- *Generated from an audited evidence pack with primary-source research. Social-media items are discussion signals, not verified facts. Nothing here is financial, legal or medical advice.*