EPIC-Bench - Embodied Compositional Attributes: leaderboard

EPIC-Bench (2026) is a mask-grounded benchmark of embodied visual perception: 6,661 annotated image, text and mask samples across 23 tasks in target localization, navigation and manipulation, scored by localization, counting, path and feasibility metrics. This board scores grounding targets described by combined attributes, part-whole relations and interaction context, on a 0-100 scale; higher is better.

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 March 2028. 76 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3 Pro53.9#145
2Gemini 3.1 Pro (Preview)53.04#99
3Qwen 3 VL 235B A22B (Thinking)48.12#459
4Seed 1.846.1#252
5GPT-5.5Non-reasoning45.19#54
6Qwen 3.5 397B A17BNon-reasoning43.45#264
7Qwen 3.5 122B A10BNon-reasoning42.31#346
8Doubao-Seed-1.6-vision42.27#361
9Qwen 3.5 27BNon-reasoning42.23#375
10Qwen 3.5 Plus41.82#242
11Qwen 3.6 27BNon-reasoning41.46#351
12Qwen 3.6 Plus41.07#238
13Qwen 2.5 VL 72B Instruct40.92#841
14Qwen 3 VL 30B A3B (Thinking)40.87#719
15Qwen 3.6 35B A3BNon-reasoning40.01#481

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=epic-bench-embodied-compositional-attributes · How It Works · Data refreshed daily, snapshot 2026-11-21.