| Rank | Model | Score |
|---|---|---|
| 1 | lfm2-5-vl-3b | 58.3 |
1 phaseActive
Multi-image understanding benchmark spanning 12 task types and 10 multi-image relations, with unanswerable-pair robustness checks.
Quick answer: MuirBench (Multi-Image Understanding and Reasoning Benchmark) tests a model's ability to reason across multiple images spanning 12 task types and 10 multi-image relations (e.g., diff-spotting, ordering, complementary views), and pairs each question with an "unanswerable" variant to check for shortcut guessing. LFM2.5-VL-3B scores 58.3% as of August 2026.
What it tests: Reasoning that requires integrating information across multiple images — comparison, ordering, complementary-view synthesis, and difference-spotting — rather than answering from a single image.
Why it matters: Real-world multimodal use cases (comparing product photos, reasoning across video frames, cross-referencing document pages) require multi-image reasoning, which single-image benchmarks don't test.
Known limitations: As a relatively new benchmark (2024), independent replication and long-term saturation trends are less established than for older single-image VQA benchmarks.
MuirBench was built to isolate multi-image reasoning as a distinct capability from single-image visual question answering. It spans 12 task types (including comparison, difference-spotting, ordering, complementary reasoning, and multi-view synthesis) across 10 categories of image relations. To guard against models exploiting single-image shortcuts, each question is paired with an "unanswerable" variant — a near-identical image set for which the correct answer is that the question cannot be answered — requiring genuine cross-image reasoning rather than pattern matching.
| Field | Value |
|---|---|
| Task category | Reasoning (multi-image) |
| Metric | % accuracy |
| Saturation | Low |
| Created by | Wang et al. |
| Source paper | Wang et al. 2024 |
| GitHub | muirbench/MuirBench |
| Dataset | MUIRBENCH/MUIRBENCH on HuggingFace |
Models answer multiple-choice questions that require reasoning across a set of images; score is the percentage of questions (including unanswerable-pair variants) answered correctly.
| Rank | Model | Score | Source | Date |
|---|---|---|---|---|
| 1 | LFM2.5-VL-3B | 58.3% | LFM2.5-VL-3B launch blog | 2026-08 |
Score sourced from Liquid AI's LFM2.5-VL-3B announcement, August 2026.
No Benchgen results yet — be the first to run MuirBench.
| Benchmark | What it tests | Saturation |
|---|---|---|
| MuirBench | Multi-image reasoning across 12 task types | Low |
| BLINK | Core visual perception (mostly single/paired-image) | Low |
Benchgen lets you evaluate your own vision-language model's multi-image reasoning and track it across model versions.