BlogLLM Benchmark Scores as Published, Side by Side
MathVerse: Results as Published
Every result the catalog holds for MathVerse: 9 models from 4 developers, read from 1 page. Each number is as printed by the company named beside it; nothing is ranked or rescored. Models are listed by developer, newest first.
Area Multimodal UnderstandingMaintainer Renrui Zhang and co-authors (CUHK MMLab, Shanghai AI Laboratory, and UCLA)Introduced 2024Data version 2026103001, as of 2026-10-30
MathVerse tests whether a multimodal model actually uses the diagram in a visual math problem. It collects 2,612 multi-subject problems with diagrams, and people rewrite each problem into six versions that move information between the text and the figure, giving 15K test samples. The project page reports accuracy on the testmini subset. The authors also score a written solution by extracting reasoning steps and marking each step, rather than only marking the final answer true or false.
MathVerse (mini)
| Model | Value | Printed by | Setup as printed |
|---|---|---|---|
| Gemma 3 12BGoogle, released 2025-03-12 | 29.8% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit |
| Qwen3-VL-32B-InstructAlibaba, released 2025-10-21 | 64.2% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit; Context: 4096 |
| 64.2% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit; Context: 32768 | |
| Qwen3-VL-32B-ThinkingAlibaba, released 2025-10-21 | 78.2% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit; Context: 40960 |
| 78.3% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit; Context: 4096 | |
| Qwen3-VL-8B-InstructAlibaba, released 2025-10-15 | 54.5% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit; Context: 4096 |
| 57.4% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit; Context: 32768 | |
| Qwen3-VL-8B-ThinkingAlibaba, released 2025-10-15 | 67.3% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit; Context: 4096 |
| 73.3% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit; Context: 40960 | |
| Kimi-VL-A3B-InstructMoonshot AI, released 2025-04-10 | 41.7% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit |
| Kimi-VL-A3B-ThinkingMoonshot AI, released 2025-04-10 | 61% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: thinking; Harness: VLMEvalKit |
| Phi-4-reasoning-vision-15BMicrosoft, released 2026-03-04 | 43.8% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: force nothink; Harness: VLMEvalKit |
| 53.1% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Effort: force thinking; Harness: VLMEvalKit | |
| 44.9% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit | |
| Phi-4-multimodal-instructMicrosoft, released 2025-02-26 | 32.4% | Microsoft, Microsoft Phi-4-reasoning-vision-15B model cardchecked 2026-10-15 | Harness: VLMEvalKit |
Every number here is read from the same catalog file as the benchmark table, data version 2026103001, and is the number the table shows for that model and test. Where one page printed a model under several setups, each value is listed. The catalog is public in armanobosyan/sugra-llm-benchmarks; a wrong number can be reported from the table, with Report an error under the value, or as an issue in that repository.