MECoBench: leaderboard

Metric: Success rate (%) averaged over the parallel setting (every agent may enter every room; 60 total steps shared by the team) and the sequential setting (agents hold disjoint room zones, so objects must be handed over; 2 agents, 80 steps) with two agents, on 96 VirtualHome household rearrangement tasks (2 to 7 subgoals each) in which multimodal agents act from four-view first-person observations; agents exchange one broadcast message each before acting; printed 0-1 and x100 here; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 15 models tracked.

Top models

#ModelScore
1GPT-5.483.4
2Gemma 4 31B (IT)82.8
3Qwen 3.5 27B81.8
4Qwen 3.5 9B77.7
5Gemma 4 26B A4B (IT)75
6Gemini 3.1 Pro (Preview)67.2
7Qwen 3 VL 32B Instruct58.4
8Qwen 3 VL 235B A22B Instruct51.1
9GLM-4.6V36.5
10Qwen 3 VL 8B Instruct26.6
11Llama 4 Scout Instruct14.6

Interactive version: theaggregate.ai/benchmark?slug=mecobench · How It Works · Data refreshed daily, snapshot 2026-09-29.