EMem-Bench - Passive Observation: leaderboard
The Passive Observation family of EMem-Bench (Liang et al., 2026): after a long history of household activity in AI2-THOR and ProcTHOR homes, the agent must find an object that appeared earlier among many others in an object-dense scene. Success rate in percent over 1,036 episodes, with the full history given as text and at most eight action steps; higher is better.
Metric: Success rate (%; the 1036 Passive Observation episodes, finding an object seen earlier in an object-dense scene; 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 | 67 | #159 |
| 2 | Gemini 2.5 Pro | 63.5 | #262 |
| 3 | Qwen 3.6 27B | 61 | #339 |
| 4 | GPT-5.4 | 60.4 | #121 |
| 5 | Mistral Small 3.1 | 58.5 | #1699 |
| 6 | GPT-5.4 Mini | 55.1 | #389 |
| 7 | InternVL3-38B | 47.9 | #905 |
| 8 | Qwen 3 VL 32B Instruct | 47.1 | #582 |
| 9 | Qwen 3 VL 4B Instruct | 46.4 | #1281 |
| 10 | Qwen 3 VL 8B Instruct | 41.4 | #982 |
| 11 | InternVL3-8B | 20.5 | #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-passive-observation · How It Works · Data refreshed daily, snapshot 2026-11-12.