LongAct (Pure VLM): leaderboard
LongAct (Zhu et al., Chinese Academy of Sciences, 2026) long-horizon household benchmark: from a free-form instruction for a routine of about nine goals in a multi-room AI2-THOR house, the agent must navigate and manipulate objects (pick and place, open and close, toggle, slice) with RGB-D and segmentation observations and up to 16,000 steps. Here the vision-language model acts directly, without an agent scaffold. The score is goal completion in percent, the share of checklist objects left in the required state, on the detailed-instruction split; higher is better.
Metric: Goal completion (%; detailed-instruction split, the VLM chooses low-level actions directly, no agent scaffold). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 2 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen3-VL-32B | 6.14 | |
| 2 | Qwen3-VL-8B | 0.74 |
Interactive version: theaggregate.ai/benchmark?slug=longact-pure-vlm · How It Works · Data refreshed daily, snapshot 2026-11-12.