PyraMathBench - Task Decomposition: leaderboard
A text-only subtask of PyraMathBench (2026), a hierarchical math benchmark derived from word problems: the task decomposition subtask (4,812 items: determine the steps needed to solve a word problem). Items are scored 0-100 with partial credit for small numeric error, zero-shot chain of thought, median of three samples; higher is better.
Metric: Score (0-100) on the task decomposition subtask (4,812 items: determine the steps needed to solve a word problem), score (0-100) per item: numeric answers score 100 within 1e-4 and fall linearly to 0 at 1 percent relative error, expressions are matched with Math-Verify, text answers by MathBERT cosine similarity of at least 0.9 and choices by exact match; zero-shot chain of thought, median of three samples at temperature 0.7; higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.6 | 97.8 | #102 |
| 2 | DeepSeek R1 | 75.6 | #473 |
| 3 | GPT-4o (2024-11-20) | 66.6 | #866 |
| 4 | GPT-4o Mini (2024-07-18) | 61.2 | #1510 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pyramathbench-task-decomposition · How It Works · Data refreshed daily, snapshot 2026-11-13.