README.md

metadata

license: apache-2.0
pretty_name: TB-MathModel
task_categories:
  - text-generation
  - question-answering
language:
  - en
  - zh
tags:
  - math-modeling
  - terminal-bench
  - agent-benchmark
  - cumcm
  - mcm
  - lean4
  - multi-oracle-evaluation
size_categories:
  - n<1K

Terminal-Bench Math Modeling

TB-MathModel (short: tb-mathmodeling) is a Terminal-Bench-style benchmark for end-to-end mathematical-modeling workflows. It contains 18 tasks: CUMCM and MCM problems A, B, and C from 2023, 2024, and 2025.

Code, verifiers, oracles, scoring scripts, and leaderboard live in the GitHub repository: wxj630/terminal-bench-math-modeling. This dataset mirrors its tasks/ directory so the tasks can be downloaded without cloning.

Paper: arXiv:2611.08237 (cs.CL), doi:10.48550/arXiv.2611.08237

@misc{wu2026terminalstylemathematicalmodelingbench,
      title={Terminal-style Mathematical Modeling Bench: The First Agentic and Verifiable Benchmark for Mathematical Modeling},
      author={Xiaojun Wu and Cehao Yang and Xuhui Jiang and Chengjin Xu and Jia Li and Jian Guo},
      year={2026},
      eprint={2611.08237},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2611.08237},
}

What is in this dataset

tasks/
├── CUMCM/    # 9 tasks: 2023 A/B/C, 2024 A/B/C, 2025 A/B/C
└── MCM/      # 9 tasks: 2023 A/B/C, 2024 A/B/C, 2025 A/B/C

Each task directory is self-contained:

Path	Contents
instruction.md	The original contest statement plus the benchmark's final-answer requirement
environment/	Dockerfile, mirrored official attachments, and output_schema.json
tests/	Verifier: score_config.json, score_result.py, test_outputs.py, expected_result.json
solution/	Oracle reproduction and solve.sh
task.toml	Harbor task metadata: timeouts, resources, scoring scope
scoring_table.md	Human-readable scoring panel for the task

The required artifact for every task is /root/results/<task-slug>_result.json.

Scoring

Real modeling problems rarely have one exact answer, so each task is scored on a single core scalar for the final question — an optimized objective, a recommended policy value, or a binary decision — rather than on every number in a report.

Two reference constructs are used:

• B-Eval (default): raw_panel_score - baseline_panel_score, i.e. how much the submission improves over a baseline method. May be negative.
• BO-Eval: the baseline-to-outstanding normalized score, using explicit baseline_value / outstanding_value / normalization_direction per metric.

The released board in the GitHub repository additionally reports a Multi-O Eval, which averages each task over every O-award human paper that reports the task's scored field, and a Multi-BO index anchored at DeepSeek V4.1 Flash = 1.000.

Leaderboard

Rank	Model	Multi-O Eval	Multi-BO index (Flash = 1.000)
1	claude-code + Fable 5.1	64.62%	1.1637
2	codex + GPT-6 Astra	61.47%	1.1200
3	opencode + DeepSeek V4.1 Flash	59.81%	1.0000
4	codex + GPT-5.5 xhigh	59.75%	1.0061
5	opencode + DeepSeek V4.1 Flash (max)	48.11%	0.8636

Single runs; across 3 seeds the per-model deviation is ±1.3–2.9pp.

Quickstart

harbor run -p tasks/CUMCM/cumcm-2024-a-dragon-dance -a oracle
harbor run -p tasks/MCM/mcm-2024-c-tennis-momentum -a codex -m openai/gpt-5