VideoNexus: Video Perception Beyond the Reasoning Context
VideoNexus scored 88.6% on LVBench at 44% of the measured API cost of Gemini’s built-in agentic processing.
The training-free framework gives perception sub-agents separate contexts and sends compact, time-linked evidence back to a planner. The planner chooses where to inspect, what details to resolve and how densely to sample. Using Gemini 3.8 Flash throughout, it beat static ingestion by 3.6 percentage points and built-in agentic processing by 1.1 points. With Kimi K3, the authors report an 8.5-point gain over the official Video-GAIA ReAct harness in their evaluation.
HF Daily Papers' note
The training-free framework gives perception sub-agents separate contexts and sends compact, time-linked evidence back to a planner. The planner chooses where to inspect, what details to resolve and how densely to sample. Using Gemini 3.8 Flash throughout, it beat static ingestion by 3.6 percentage points and built-in agentic processing by 1.1 points. With Kimi K3, the authors report an 8.5-point gain over the official Video-GAIA ReAct harness in their evaluation.
HF Daily Papers' note
score 5