AgentViSS - Interaction Regulation (Direct Vision): leaderboard
AgentViSS (2026) evaluates visual social intelligence: multimodal agents play roles in 240 multi-party scenarios and must read expressions, posture and emotion from images while pursuing role goals, with each task labeled achieved, partial or not achieved by three judge models. This board is the 0 to 100 interaction-regulation score when the agent receives the images directly; higher is better.
Metric: Interaction-regulation task score, the images given directly to the agent; normalized task score (0-100), each role-task instance labeled Achieved 2, Partially Achieved 0.5 or Not Achieved 0 by majority vote of three judges (Gemini 3.1 Pro Preview, GPT-5.4, Qwen3.5-27B), over 2,340 role-task instances in 240 multi-party social scenarios with group images and role portraits; higher is better. Source: arxiv.org. Saturation forecast: Not forecast. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 40.9 | #803 |
| 2 | Claude Sonnet 4.6 | 34.1 | #161 |
| 3 | Qwen 3.5 27B | 33.97 | #359 |
| 4 | Qwen 3.5 122B A10B | 30.64 | #331 |
| 5 | GLM-4.6V | 25.81 | #701 |
| 6 | GPT-5.4 | 25.56 | #123 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=agentviss-interaction-regulation-direct-vision · How It Works · Data refreshed daily, snapshot 2026-11-13.