Embodied-BenchForge - Embodied QA: leaderboard
Trajectory-grounded EQA accuracy on all 220 Interactive Embodied Track tasks. This is distinct from task success and the joint successful-and-correct SRL metric.
Metric: Accuracy (%, 220 interactive tasks). Source: arxiv.org. Saturation forecast: Around July 2028. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 (High) | 70 | #51 (Claude Opus 4.7) |
| 2 | Gemini 3 Flash (Preview) (High) | 59.55 | #90 (Gemini 3 Flash (Preview)) |
| 3 | Qwen 3.5 4B (Thinking) | 55.91 | #514 (Qwen 3.5 4B) |
| 4 | Claude Haiku 4.5 (Non-reasoning) | 55.91 | #287 (Claude Haiku 4.5) |
| 5 | Qwen 3.6 35B A3B (Thinking) | 50.91 | #238 (Qwen 3.6 35B A3B) |
| 6 | Qwen 3.5 2B (Non-reasoning) | 46.82 | #1114 (Qwen 3.5 2B) |
| 7 | Claude Sonnet 4.6 (High) | 42.27 | #94 (Claude Sonnet 4.6) |
| 8 | Qwen 3.5 0.8B (Non-reasoning) | 22.73 | #1410 (Qwen 3.5 0.8B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=embodied-benchforge-embodied-qa · How It Works · Data refreshed daily, snapshot 2026-11-05.