MM-OptBench - Combinatorial and Logical Models: leaderboard
MM-OptBench (2026) asks multimodal models to turn a text specification plus visual artifacts such as tables, graphs, maps and schedules into an optimization formulation and solver code, which is run and checked against the verified optimum. This board is pass@1 in percent, the mean of five runs, on knapsack, colouring, packing and logical constraint problems; higher is better.
Metric: pass@1 (%) on the pure combinatorial and logical models (120 instances) family, mean of five runs, solver-grounded: the model reads a text specification plus visual artifacts (tables, graphs, maps, schedules) and must return an optimization formulation and executable solver code whose returned objective matches the verified optimum, prompting only with no fine-tuning, retrieval or agentic repair; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 82.3 | #135 |
| 2 | Gemini 3.1 Pro (Preview) | 49.2 | #99 |
| 3 | Qwen-VL-Max (MM-OptBench checkpoint unspecified) | 40.3 | |
| 4 | Claude Sonnet 4.6 | 33.8 | #173 |
| 5 | Qwen 3 VL Plus | 26.3 | #512 |
| 6 | GLM-4.5V | 23.8 | #754 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=mm-optbench-combinatorial-and-logical-models · How It Works · Data refreshed daily, snapshot 2026-11-21.