| name | video-deep-research-agent |
| title | Watching, Reasoning, and Searching: A Video Deep Research Benchmark on Open Web for Agentic Video Reasoning |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2601.06943 |
| keywords | ["Video","Deep Research","Web Search","Reasoning","Agent","Multi-hop","Benchmark"] |
| description | Implements video deep research for multi-hop reasoning combining video analysis, web search, and evidence synthesis. Evaluates workflow vs agentic paradigms with 100-sample benchmark across 6 semantic domains, revealing goal drift and long-horizon consistency as core bottlenecks. |
Overview
This skill implements the VideoDR methodology for evaluating video-grounded deep research agents. The paper presents the first benchmark combining video analysis with multi-hop web search and reasoning. Two execution paradigms are compared: Workflow (two-stage with explicit visual extraction) and Agentic (end-to-end with raw video access).
Core Task Definition
Given a video V, question Q, and search tool S, produce factual answer A where:
- Models must extract multi-frame visual anchors from video
- Translate anchors into web search queries
- Perform multi-hop retrieval and verification
- Answer must depend on BOTH video and web evidence
Key Findings
Paradigm Performance (Agentic vs Workflow):
- Gemini-3-pro-preview: 76% vs 69% (+7% Agentic)
- GPT-5.2: 69% vs 69% (tied)
- GPT-4o: 43% vs 42% (+1% Agentic)
- MiniCPM-V 4.5: 16% vs 25% (-9% Agentic penalty)
Critical Insight: Agentic is NOT consistently superior. Its advantage depends on model's ability to maintain initial video anchors across long search chains without re-watching.
Bottlenecks Discovered
- Goal Drift (Agentic exclusive): Search queries deviate from initial visual anchors when models lack strong state retention
- Long-Horizon Consistency: Multi-round search without video re-watch causes drift, especially on videos >10min
- Numerical Reliability: All models show persistent weakness on numerical answers (6-12 errors), independent of paradigm
Workflow Paradigm (Two-Stage)
Stage 1 - Visual Extraction:
Input: Raw video + question → Output: Structured text of cross-frame visual anchors
Stage 2 - Web Search & Reasoning:
Input: Anchor text + question → Tools: search, think → Output: Answer
Advantages:
- Repeated access to visual summary prevents goal drift
- Better for spatial tasks (Geography: Gemini 70% Workflow vs 50% Agentic)
- Stable on long videos (Gemini 70% Workflow vs 50% Agentic on >10min)
- Supports weak models (MiniCPM-V 4.5: 25% vs 16%)
When to use: Long videos, spatial reasoning, weak perception models, consistent target maintenance
Agentic Paradigm (End-to-End)
Input: Raw video + question → Single loop with search, think tools → Output: Answer
Advantages:
- Preserves full visual details
