VideoFDB - Perception - Semantic Grounding: leaderboard
The semantic grounding rubric of VideoFDB (Mazumdar et al., 2026) rates whether the response content fits the visual and affective event in the clip. 105 held-out test clips of seven nonverbal conversational dynamics, such as pauses, backchannels, laughter and interruptions, are streamed in real time to a full-duplex speech agent that also sees the user, at 1 FPS for most agents, and a gpt-4o judge scores each response from 0 to 5; higher is better.
Metric: gpt-4o judge score (0-5; the 105 held-out test clips of the seven nonverbal conversational dynamics scored for perception (237 clips of 11 dynamics in the full benchmark), streamed in real time; the judge reads the transcribed, time-stamped agent response with the clip annotations; response content against the visual-affective event, on the dynamics this rubric covers; user audio and video streamed, video at 1 FPS (Mini-Omni2: one frame per user speech segment)). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniCPM-o 4.5 | 3.63 | |
| 2 | Gemini 2.5 Native Audio | 3.37 | |
| 3 | Gemini 3.1 Flash Live | 3.16 | |
| 4 | gpt-realtime | 3.02 | |
| 5 | gpt-realtime-mini | 2.9 | |
| 6 | VITA-1.5 | 2.53 | #3043 |
| 7 | Mini-Omni2 | 1.54 |
Interactive version: theaggregate.ai/benchmark?slug=videofdb-perception-semantic-grounding · How It Works · Data refreshed daily, snapshot 2026-11-12.