# BrokenArXiv ![Category](https://img.shields.io/badge/Category-Math-1f6feb) ![Metric](https://img.shields.io/badge/Metric-Judge--Based-2da44e) ![Saturation](https://img.shields.io/badge/Saturation-low-d29922) ![Created](https://img.shields.io/badge/Created-2026-555) [![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b)](https://arxiv.org/abs/2605.00674) [![GitHub](https://img.shields.io/badge/GitHub-eth--sri%2Fmatharena-181717?logo=github&logoColor=white)](https://github.com/eth-sri/matharena) [![Leaderboard](https://img.shields.io/badge/Leaderboard-MathArena-0969da)](https://matharena.ai/) > **Quick answer:** BrokenArXiv is a companion track to MathArena's ArXivMath, built by ETH SRI. Instead of solving a final-answer question, it tests whether a model can correctly identify invalid, flawed, or broken mathematical reasoning and proofs in freshly published arXiv papers, scored by a judge model. ## At a Glance **What it tests:** Critical mathematical review — can a model spot where a piece of published mathematical reasoning is actually wrong, rather than just producing its own answer? **Why it matters:** Verifying a proof or derivation is a distinct (and arguably harder) skill than generating one; this measures a model's ability to critically audit mathematical content, relevant for any agentic use case involving reviewing or fact-checking technical work. **Known limitations:** Uses freshly-sourced papers on a rolling basis (no fixed task count published) and relies on judge-based scoring rather than a fully automatic check, so evaluation consistency depends on the judge's reliability. ## What BrokenArXiv Measures BrokenArXiv is part of the MathArena platform (the same ETH SRI project behind ArXivMath), built specifically to test critical mathematical reasoning rather than generative problem-solving. Where ArXivMath asks a model to solve a fresh, automatically-gradable final-answer question derived from a new arXiv paper, BrokenArXiv instead presents reasoning or proof steps from freshly published papers — some intentionally containing flaws — and asks the model to correctly identify whether the reasoning is valid or broken. Because this requires a judgment call rather than a single checkable numeric answer, BrokenArXiv is scored using a judge model rather than fully automatic grading. A high score indicates a model can reliably distinguish sound mathematical reasoning from subtly flawed reasoning — a skill distinct from, and arguably harder than, generating correct proofs from scratch. ## Benchmark Specifications | Field | Value | |---|---| | Task category | Math | | Metric | Judge-based scoring / accuracy | | Number of tasks | Rolling monthly collection; no fixed aggregate count published | | Saturation | Low | | Created by | MathArena / ETH SRI | | Source paper | [MathArena team 2026](https://arxiv.org/abs/2605.00674) | | GitHub | [eth-sri/matharena](https://github.com/eth-sri/matharena) | | Leaderboard | [matharena.ai](https://matharena.ai/) | ## How BrokenArXiv Is Scored A judge model evaluates whether the model under test correctly identified valid vs. flawed reasoning in each sampled excerpt from a freshly published arXiv math paper. Scores are reported as an accuracy percentage. ## State-of-the-Art Results | Rank | Model | Score | Source | Date | |---|---|---|---|---| | 1 | [Hy4 Preview](/models/hungyuan/hy4-preview) | 54.6% | [Tencent Hunyuan model card](https://huggingface.co/tencent/Hy4-preview) | 2026-08 | *Scores sourced from published technical reports and model cards. Results depend on harness, prompt format, and effort settings — see each source for methodology.* ## BrokenArXiv on Benchgen > No Benchgen results yet — [be the first to run BrokenArXiv](/benchmarks/matharena-eth-sri/brokenarxiv). ## BrokenArXiv vs Other Benchmarks | Benchmark | What it tests | Tasks | Saturation | |---|---|---|---| | **BrokenArXiv** | Identifying flawed reasoning/proofs in new arXiv papers | Rolling | Low | | [ArXivMath](/benchmarks/matharena-eth-sri/arxivmath) | Solving fresh final-answer math from new arXiv papers | Rolling | Low | | [MathArena Apex 2025](/benchmarks/matharena-eth-sri/matharena-apex-2025) | Competition mathematics | — | Medium | Use BrokenArXiv specifically to evaluate a model's critical-review capability on mathematics, distinct from its generative problem-solving ability measured by ArXivMath. ## Run BrokenArXiv on Your Model Benchgen lets teams run BrokenArXiv against their own model versions, compare results across runs, and catch regressions in critical mathematical reasoning — rather than relying on a single vendor-reported number. ## Related Models & Benchmarks - Top performers: [Hy4 Preview](/models/hungyuan/hy4-preview) - Related benchmarks: [ArXivMath](/benchmarks/matharena-eth-sri/arxivmath) · [MathArena Apex 2025](/benchmarks/matharena-eth-sri/matharena-apex-2025) ## Frequently Asked Questions

What is BrokenArXiv?

BrokenArXiv is a MathArena track that tests whether a model can identify invalid or flawed mathematical reasoning and proofs in freshly published arXiv papers, scored by a judge model.

What does a good BrokenArXiv score look like?

As of August 2026, frontier models score in the 42–78% range; scores above 54% represent solid critical mathematical review capability.

Who created BrokenArXiv?

BrokenArXiv is part of MathArena, created by researchers at ETH Zurich's SRI lab and INSAIT as a companion track to ArXivMath.

BrokenArXiv — Results

BrokenArXiv — Evaluations