VideoDR v2 (Agentic): leaderboard
VideoDR (2026) asks multimodal models open-domain questions about a video whose answers need visual anchors from several frames, web search and multi-hop reasoning; an LLM judge checks semantic equivalence with the gold answer. This board is the expanded 500-question v2 set from the paper's second version, under the Agentic paradigm, where the model keeps direct access to the video while it drives its own searches; accuracy (%), higher is better.
Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 53.6 | #91 |
| 2 | GPT-5.4 | 47.4 | #121 |
| 3 | Qwen 3.5 35B A3B | 34.6 | #480 |
| 4 | Gemma 4 31B (IT) | 33.6 | #343 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=videodr-v2-agentic · How It Works · Data refreshed daily, snapshot 2026-11-12.