菜单
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.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill video-deep-research-agent命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| 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. |
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).
Given a video V, question Q, and search tool S, produce factual answer A where:
Paradigm Performance (Agentic vs Workflow):
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.
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:
When to use: Long videos, spatial reasoning, weak perception models, consistent target maintenance
Input: Raw video + question → Single loop with search, think tools → Output: Answer
Advantages:
Disadvantages:
When to use: Short videos, fine-grained visual reasoning, strong multimodal models
Annotation Process (3 steps):
Quality Verification (2 stages):
Difficulty Stratification (by human success rate):
By Difficulty Level: All models degrade consistently from Low→Mid→High. Agentic gains emerge on Mid/High for strong models (Gemini, GPT-5.2) but cause backslash on High for weak models (MiniCPM, GPT-4o).
By Video Duration:
By Domain:
Table from paper shows tool count does NOT predict accuracy:
Model Think Search Runtime(s) Accuracy
Gemini-Agentic 2.89 2.52 449 76%
GPT-5.2-Agentic 3.06 2.74 1376 69%
Qwen3-Agentic 1.80 1.21 367 37%
MiniCPM-Agentic 1.97 2.07 139 16%
Key insight: Effectiveness is converting searches into high-quality evidence chains, not tool frequency.
Categorical Error (dominant failure mode): 5-36 cases per model
Numerical Error (persistent weakness): 6-12 cases per model
Reasoning Error (nearly absent): 0-1 cases
Choose paradigm based on:
Video Duration:
10 min: Workflow strongly preferred unless using state-strong models (Gemini-class)
Task Type:
Model Capability:
Evidence Requirements:
基于 SOC 职业分类
