MathArena

👋 Overview

MathArena (NeurIPS D&B '25) is a platform for evaluation of LLMs on latest math competitions and olympiads. It is hosted on matharena.ai. This repository contains all code used for model evaluation. This README explains how to run your models. For more details about other aspects of the project, such as adding new competitions, please refer to the specific README files in the readmes/ folder. You can find logs from our evaluation containing full reasoning traces (if available) and solutions produced by the models on our HuggingFace page: https://huggingface.co/MathArena.

---

🚀 Installation

MathArena uses UV to manage dependencies. If you want to run local models, uncomment the vllm installation in pyproject.toml.

Install UV

• macOS and Linux:

  curl -LsSf https://astral.sh/uv/install.sh | sh

• Windows:

  powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

Alternative installation

As an alternative to UV, you can also create a conda environment and install the package as follows:

conda create -n matharena python=3.12
conda activate matharena
python -m pip install -e .

If you choose this option, disregard uv run in all instructions and use python directly instead.

---

🏃 Running an Eval

Before running harness models, build the Docker image and prepare its cache once from the repository root:

bash scripts/build_harness_arxivlean_image.sh

Execute the following command to evaluate a model on a competition:

uv run python scripts/run.py --comp path/to/competition --models path/to/model1

• path/to/competition: Relative path from the configs/competitions folder to the competition config file (excluding the .yaml extension).
• path/to/model1: Relative path (or multiple) from the configs/models folder to the model config file (excluding the .yaml extension). See Adding a Model/Agent below for model config file structure.

The scripts run_all_fa.sh and run_all_kangaroo.sh provide convenient shortcuts to run a model on all current non-deprecated competitions. They can be executed as follows:

bash scripts/run_all_fa.sh path/to/model

For ArXivLean September, follow the Lean 4.34.1 setup instructions before running the evaluation. Earlier editions use the legacy setup instructions.

Example:

uv run python scripts/run.py --comp aime/aime_2025 --models openai/gpt-4o 

Additional Flags:

• --n: Number of runs per problem (default: 4).
• --redo-all: Ignore existing runs for this model and rerun everything (default: false, continues from existing runs found in outputs/).
• --problems: One-based indices of problems to run (default: runs all problems).
• --container-runtime docker|podman: Container backend for harness runs and judging (default: Docker). Docker and Podman use separate image and cache stores.

For rootless Podman, build the runtime in Podman’s storage and select that backend when running:

bash scripts/build_harness_arxivlean_image.sh --container-runtime podman
uv run python scripts/run.py --comp arxiv/september --models openai/gpt-61-sol --container-runtime podman --n 1

September’s ArXivMath and BrokenArXiv harness configurations allow 12 hours and $100 per attempt, followed by a five-minute final-answer grace period at the dollar limit. ArXivLean September sets a $100 per-attempt limit and no overall time limit. Costs are estimates from token usage and the model config’s prices; an in-flight request and the configured grace period can exceed the nominal dollar limit.

Current Website Competitions

The table below maps all non-deprecated competitions currently shown on the website to their competition config. Requires judging indicates whether the config requires a separate scripts/judge/judge.py pass.

Website section	Website competition	Competition config	Requires judging
ArXivMath	08/2026	arxiv/august.yaml	Yes
ArXivMath	09/2026	arxiv/september.yaml	Yes
BrokenArxiv	08/2026	arxiv_false/august.yaml	Yes
BrokenArxiv	09/2026	arxiv_false/september.yaml	Yes
ArXivLean	09/2026	arxivlean/september.yaml	No (Lean comparator)
Visual Math	Kangaroo 2025 1-2	kangaroo/kangaroo_2025_1-2.yaml	No
Visual Math	Kangaroo 2025 3-4	kangaroo/kangaroo_2025_3-4.yaml	No
Visual Math	Kangaroo 2025 5-6	kangaroo/kangaroo_2025_5-6.yaml	No
Visual Math	Kangaroo 2025 7-8	kangaroo/kangaroo_2025_7-8.yaml	No
Visual Math	Kangaroo 2025 9-10	kangaroo/kangaroo_2025_9-10.yaml	No
Visual Math	Kangaroo 2025 11-12	kangaroo/kangaroo_2025_11-12.yaml	No
Final-Answer Comps	AIME 2026	aime/aime_2026.yaml	No
Final-Answer Comps	HMMT Feb 2026	hmmt/hmmt_feb_2026.yaml	No
Final-Answer Comps	Apex	apex/apex_2025.yaml	No
Final-Answer Comps	Apex Shortlist	apex/shortlist_2025.yaml	No
Proof-Based Comps	USAMO 2025	usamo/usamo_2025.yaml	Yes
Proof-Based Comps	IMO 2025	imo/imo_2025.yaml	Yes
Proof-Based Comps	IMC 2025	imc/imc_2025.yaml	Yes
Proof-Based Comps	Miklos Schweitzer 2025	miklos/2025.yaml	Yes
Proof-Based Comps	Putnam 2025	putnam/putnam_2025.yaml	Yes
Proof-Based Comps	USAMO 2026	usamo/usamo_2026.yaml	Yes
Project Euler	Project Euler	euler/euler.yaml	No

The September editions each contain 53 problems and load directly from Hugging Face: ArXivMath, BrokenArXiv, and ArXivLean. Saved model outputs are published in the corresponding ArXivMath, BrokenArXiv, and ArXivLean output datasets. Older competition configs remain available for reproducing earlier evaluations.

Note: Since we do not publish the correct answers for Project Euler problems, they cannot be directly judged as correct or incorrect without adapting data/euler/euler/answers.csv with the correct answers. Putnam, IMC, IMO 2025, USAMO 2025, Miklos Schweitzer 2025 were graded using human verification. They can therefore not be graded automatically. For these competitions, if you want to run them, you will have to adjust their config similar to the USAMO 2026 competition.

Competitions Requiring Grading

For competitions requiring grading (including BrokenArXiv and USAMO), run:

uv run python scripts/judge/judge.py --comp path/to/competition

There are various agents available for judging, you can see example configs for a couple agents in configs/judges/. If you want to add a new agent, you can follow the examples.

Seeing Results

Launch a local web server that inspects all successful runs that were saved to output: uv run python app/app.py --comp path/to/competition, and access it at http://localhost:5001/. This shows the final answers but also full interactions with the model or all steps that an agent took (see for example the runs of GPT-5 Agent on apex/apex_2025). Warning signs for runs indicate potential problems and should be manually verified. Any warning is caused by one of the following problems:

• 💀: parser threw an error or encountered something unexpected.
• ⚠️: The correct answer might be present in the model answer, but it was not extracted.
• ❕: Model likely hit max token limit.

---

🤖 Adding a New Model/Agent

To add a new model add a config file in the configs/models folder. Each config must include:

• Required:
  • model: Model name. Reasoning effort of OpenAI models can be set by appending --[low/medium/high] to the model name, e.g., o3-mini--high.
  • api: API provider. The API key should be defined as an environment variable when using the specified API. The supported options with their corresponding API keys are:
    • xai: XAI_API_KEY
    • openai: OPENAI_API_KEY
    • together: TOGETHER_API_KEY
    • google: GOOGLE_API_KEY
    • anthropic: ANTHROPIC_API_KEY
    • glm: GLM_API_KEY
    • bigmodel: BIGMODEL_API_KEY
    • deepseek: DEEPSEEK_API_KEY
    • openrouter: OPENROUTER_API_KEY
    • vllm: (runs locally; no API key required)
  • human_readable_id: A unique, descriptive identifier.
• Optional Parameters:
  • API settings like temperature, top_p, and top_k.
  • max_tokens: Max number of tokens for the model.
  • concurrent_requests: Number of parallel requests to API (default: 30).
  • timeout: Request timeout in seconds (default: 2000).
  • max_retries: Retry attempts to API (default: 50).
  • read_cost & write_cost: Cost per million tokens in USD for input and output tokens (default: 1 each).
  • cache_read_cost: Cost per million cached input tokens in USD (default: same as read_cost).
  • date: Release date of the model in the format "yyyy-mm-dd".
  • batch_processing: If set to true, the model will be queried using batch processing. Available for OpenAI, Anthropic, and OpenRouter models.
  • use_openai_responses_api: If set to true, will use the OpenAI responses API (instead of chat completions).
  • harness: Run the model through codex, claude, kimi, gravity, qwen, opencode, deepcode, grok, muse, or vibe; requires the competition to opt in with allow_harness: true. Set it to false to opt out of a competition default.
  • harness_version: Optional exact CLI version; defaults to the latest release resolved at run startup.
  • harness_config: Docker image, resource limits, workspace, and harness authentication settings. Harness runs use minimal reproducible CLI context by default; set minimal_context: false only to restore a CLI's native skills and instructions.
  • Other model/provider specific parameters (config, provider, reasoning, etc.).

New model profiles include openai/gpt-61-sol, anthropic/opus_55_high, qwen/qwen3.8_max, deepseek/deepseek_v41_flash, xai/grok-47, mistral/mistral_large_4, and gemini/gemini-4-argon. The Argon profile uses Google API authentication through Antigravity CLI; set GOOGLE_API_KEY. Each profile specifies its provider, harness version, reasoning settings, and token prices.

See Harness-backed model runs for configuration, API billing, isolation, and arXivLean tool integration details.

Agents

Agents are defined via top-level config files (see e.g., config/models/openai/gpt-5-agent.yaml) that point to a pure model config, indicating the underlying LLM API used by the agent, and an agent scaffolding config which parametrizes the agents' workflow.

To add a new scaffolding, follow the example of solvers/selfcheck_agent.py which uses utility functions from base_agent.py.

---

📚 Citation

@article{dekoninck2026matharena,
      title={Beyond Benchmarks: MathArena as an Evaluation Platform for Mathematics with LLMs}, 
      author={Jasper Dekoninck and Nikola Jovanović and Tim Gehrunger and Kári Rögnvaldsson and Ivo Petrov and Chenhao Sun and Martin Vechev},
      year={2026},
      eprint={2605.00674},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2605.00674}, 
}

For BrokenArXiv dataset generation (arxiv_false), see the source-based Codex pipeline and overhaul notes.