Skip to main content

Mathematics

MathConstruct

MathConstruct: 121 math-olympiad constructive-proof problems parameterised into 455 instances, graded by deterministic checkers into binary correct/incorrect; released per-(model x instance) results for 19 (CoT) + 10 (code) LLMs.

455items
20subjects
90%observed
MITlicense
mathematicsdomain
reasoningdomain
textmodality

Response matrix

Fit to width. Hover for subject & item; click a cell for details.

MathConstruct response matrix: AI models (rows) against items (columns)
Correct (1)Incorrect (0)Unobserved

Scale: 1 = correct · 0 = incorrect

Subjects

  1. 1gpt-5--high0.811
  2. 2o3-mini--high0.5956
  3. 3deepseek-chat0.5
  4. 4o3-mini0.4527
  5. 5o10.4462
  6. 6deepseek-ai/DeepSeek-R10.3253
  7. 7gemini-2.0-flash-thinking-exp0.2703
  8. 8o1-mini0.2549
  9. 9deepseek-reasoner0.25
  10. 10deepseek-ai/DeepSeek-R1-Distill-Llama-70B0.1868
  11. 11gemini-2.0-flash-exp0.1516
  12. 12Qwen/QwQ-32B-Preview0.1341
  13. 13claude-3-5-sonnet-latest0.1022
  14. 14Qwen/Qwen2.5-72B-Instruct-Turbo0.0967
  15. 15gpt-4o0.078
  16. 16meta-llama/Meta-Llama-3.1-405B-Instruct-Turbo0.0725
  17. 17claude-3-5-haiku-latest0.0681
  18. 18meta-llama/Llama-3.3-70B-Instruct-Turbo0.0626
  19. 19gpt-4o-mini0.0571
  20. 20meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo-128K0.0022