EgoBench - Dynamic Hard Mode: leaderboard
Egocentric multimodal tool use (EgoBench, Liu et al., 2026): an agent watches a short first-person video and serves a simulated customer in retail, restaurant, food-ordering and kitchen scenarios by calling scenario tools over a database, with thinking disabled, across 1,045 tasks. Here the simulated user reveals the request over several turns but is impatient, adds unrelated chatter to each message and may end the conversation. The score is the joint success rate, the percent of tasks in which the agent issued every required tool call and left the database in the correct final state; higher is better.
Metric: Joint Success Rate (%; every required tool call issued and the final database state correct, impatient simulated user adding unrelated chatter). Source: arxiv.org. Saturation forecast: Around December 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 16.75 | #91 |
| 2 | Qwen 3.5 397B A17B (Non-reasoning) | 14.35 | #242 (Qwen 3.5 397B A17B) |
| 3 | Qwen 3.6 Plus (Non-reasoning) | 13.88 | #217 (Qwen 3.6 Plus) |
| 4 | Kimi K2.5 (Non-reasoning) | 11.2 | #252 (Kimi K2.5) |
| 5 | GLM-5V Turbo | 3.33 | #342 |
| 6 | Qwen 3 VL 235B A22B Instruct | 0.29 | #521 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=egobench-dynamic-hard-mode · How It Works · Data refreshed daily, snapshot 2026-11-12.