{"paper":{"title":"Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Expanding MathArena to proofs and research questions shows frontier LLMs solve 98% of 2026 USAMO problems.","cross_cats":[],"primary_cat":"cs.CL","authors_text":"Chenhao Sun, Ivo Petrov, Jasper Dekoninck, K\\'ari R\\\"ognvaldsson, Martin Vechev, Nikola Jovanovi\\'c, Tim Gehrunger","submitted_at":"2026-05-01T13:56:34Z","abstract_excerpt":"Large language models (LLMs) are becoming increasingly capable mathematical collaborators, but static benchmarks are no longer sufficient for evaluating progress: they are often narrow in scope, quickly saturated, and rarely updated. This makes it hard to compare models reliably and track progress over time. Instead, we need evaluation platforms: continuously maintained systems that run, aggregate, and analyze evaluations across many benchmarks to give a comprehensive picture of model performance within a broad domain. In this work, we build on the original MathArena benchmark by substantially"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Notably, the strongest model, GPT-5.5, now reaches 98% on the 2026 USA Math Olympiad and 74% on research-level questions, showing that frontier models can now comfortably solve extremely challenging mathematical problems.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The new benchmarks and evaluation protocol remain free of training-data contamination and fairly measure genuine reasoning rather than memorization or benchmark-specific shortcuts.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"MathArena is broadened into a maintained platform with new benchmarks for proofs, research questions, and formal verification, where GPT-5.5 scores 98% on 2026 USAMO and 74% on research-level tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Expanding MathArena to proofs and research questions shows frontier LLMs solve 98% of 2026 USAMO problems.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"14dd8eeb1bd2b60189b733bf4f60f2d3ce5085527627ef0f4d88d12607437aba"},"source":{"id":"2605.00674","kind":"arxiv","version":2},"verdict":{"id":"b18e07f0-5273-4711-9f83-a4ed862bd3e1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T18:14:31.031764Z","strongest_claim":"Notably, the strongest model, GPT-5.5, now reaches 98% on the 2026 USA Math Olympiad and 74% on research-level questions, showing that frontier models can now comfortably solve extremely challenging mathematical problems.","one_line_summary":"MathArena is broadened into a maintained platform with new benchmarks for proofs, research questions, and formal verification, where GPT-5.5 scores 98% on 2026 USAMO and 74% on research-level tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The new benchmarks and evaluation protocol remain free of training-data contamination and fairly measure genuine reasoning rather than memorization or benchmark-specific shortcuts.","pith_extraction_headline":"Expanding MathArena to proofs and research questions shows frontier LLMs solve 98% of 2026 USAMO problems."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.00674/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_compliance","ran_at":"2026-05-19T17:56:14.716195Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"b7ed88aa31ffa30aa3052fb1d4cc4d988fa53280d5bfc7f48f2d9f1c5eccc211"},"references":{"count":110,"sample":[{"doi":"","year":2024,"title":"Wei-Lin Chiang, Lianmin Zheng, Ying Sheng, Anastasios Nikolas Angelopoulos, Tianle Li, Dacheng Li, Banghua Zhu, Hao Zhang, Michael I. Jordan, Joseph E. Gonzalez, and Ion Stoica. Chatbot arena: An open","work_id":"b113a7e2-dc84-4564-817e-141ee365f5a6","ref_index":6,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Ai model & api providers analysis","work_id":"e6067ebc-31e4-4f46-aa93-c2b894abb00d","ref_index":7,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Accessed: 2026-04-13","work_id":"69f838f3-378d-4111-9988-826f26d14ec3","ref_index":8,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2026,"title":"Vals AI. Benchmarks.https://www.vals.ai/benchmarks, 2026. Accessed: 2026-04-13","work_id":"fc94e773-4ed5-4745-861e-5c6462ba401f","ref_index":9,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"MathArena: Evaluating LLMs on Uncontaminated Math Competitions","work_id":"61e8d872-ccc7-46b8-8ec1-94008704c941","ref_index":10,"cited_arxiv_id":"2505.23281","is_internal_anchor":true}],"resolved_work":110,"snapshot_sha256":"3e572e08799c52073806386ff2a33bd44ca3b6414a388a1558896f98b81125b2","internal_anchors":5},"formal_canon":{"evidence_count":2,"snapshot_sha256":"d8ae847744fa30b19249002cc1fb50bccad3fedb85cb8ae12e2c51852efaf8e3"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}