EPIC-Bench - Embodied Compositional Attributes: leaderboard
Metric: Target localization score (0-100) on the embodied compositional tasks (targets described by combined attributes, part-whole relations and interaction context), on EPIC-Bench (6,661 human-annotated image, text and mask tuples from 25 public datasets across 23 fine-grained embodied perception tasks; the model outputs bounding boxes, counts, path points or feasibility judgements that are scored against the masks), zero-shot, averaged over six runs for local models and two for API models; higher is better. Source: arxiv.org. Saturation forecast: Around 2030. 89 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro | 53.9 |
| 2 | Gemini 3.1 Pro (Preview) | 53.04 |
| 3 | Qwen 3 VL 235B A22B (Thinking) | 48.12 |
| 4 | Seed 1.8 | 46.1 |
| 5 | GPT-5.5 (Non-reasoning) | 45.19 |
| 6 | Qwen 3.5 397B A17B (Non-reasoning) | 43.45 |
| 7 | Qwen 3.5 122B A10B (Non-reasoning) | 42.31 |
| 8 | Qwen 3.5 27B (Non-reasoning) | 42.23 |
| 9 | Qwen 3.5 397B A17B | 42.1 |
| 10 | Qwen 3.5 Plus | 41.82 |
| 11 | Qwen 3.6 27B (Non-reasoning) | 41.46 |
| 12 | Qwen 3.5 122B A10B | 41.16 |
| 13 | Qwen 3.6 Plus | 41.07 |
| 14 | Qwen 2.5 VL 72B Instruct | 40.92 |
| 15 | Qwen 3 VL 30B A3B (Thinking) | 40.87 |
Interactive version: theaggregate.ai/benchmark?slug=epic-bench-embodied-compositional-attributes · How It Works · Data refreshed daily, snapshot 2026-10-07.