EMem-Bench - Interaction Failure: leaderboard
The Interaction Failure family of EMem-Bench (Liang et al., 2026): after a long history of household activity in AI2-THOR and ProcTHOR homes, the agent must avoid an object that an earlier failed action showed to be unusable and choose a working alternative. Success rate in percent over 263 episodes, with the full history given as text and at most eight action steps; higher is better.
Metric: Success rate (%; the 263 Interaction Failure episodes, avoiding a target that an earlier failed action showed to be unusable; 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 | 60.5 | #159 |
| 2 | Gemini 2.5 Pro | 55.1 | #262 |
| 3 | Qwen 3.6 27B | 54.4 | #339 |
| 4 | GPT-5.4 | 51 | #121 |
| 5 | Qwen 3 VL 32B Instruct | 45.3 | #582 |
| 6 | InternVL3-38B | 29.3 | #905 |
| 7 | Qwen 3 VL 4B Instruct | 20.9 | #1281 |
| 8 | GPT-5.4 Mini | 19 | #389 |
| 9 | Qwen 3 VL 8B Instruct | 11.8 | #982 |
| 10 | Mistral Small 3.1 | 11.8 | #1699 |
| 11 | InternVL3-8B | 5.3 | #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-interaction-failure · How It Works · Data refreshed daily, snapshot 2026-11-12.