EMem-Bench: leaderboard
EMem-Bench (Liang et al., 2026) tests whether multimodal agents build and use memory over long embodied interaction: after a long history of household activity in simulated homes, the agent must complete a new task that depends on something it saw, tracked, learned from a failure or generalized from earlier corrections. Success rate in percent, the unweighted mean of the four task families, with the full history given as text; higher is better.
Metric: Success rate (%; unweighted mean of the four task-family success rates over 2,554 episodes; full-context condition: the complete interaction history as sanitized text plus the current 500 x 500 RGB view, one primitive action per call with deterministic decoding and at most eight target-task steps in AI2-THOR and ProcTHOR homes; an episode succeeds only if the simulator reaches the target terminal state). Source: arxiv.org. Saturation forecast: Around February 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 64.2 | #159 |
| 2 | Gemini 2.5 Pro | 60.8 | #262 |
| 3 | Qwen 3.6 27B | 58.2 | #339 |
| 4 | GPT-5.4 | 56.8 | #121 |
| 5 | Qwen 3 VL 32B Instruct | 43.9 | #582 |
| 6 | Mistral Small 3.1 | 40.9 | #1699 |
| 7 | GPT-5.4 Mini | 35.3 | #389 |
| 8 | InternVL3-38B | 31.8 | #905 |
| 9 | Qwen 3 VL 4B Instruct | 30.1 | #1281 |
| 10 | Qwen 3 VL 8B Instruct | 25.4 | #982 |
| 11 | InternVL3-8B | 15.5 | #1695 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=emem-bench · How It Works · Data refreshed daily, snapshot 2026-11-12.