AgentSkillsCN

tao-deepresearch

深入进行AI/ML研究。在分析arXiv论文、开展文献综述、寻找有影响力的研究论文、识别研究趋势,或探索各类实现方案时使用此功能。支持LaTeX源码分析、网络搜索论文,以及GitHub代码发现。

SKILL.md
--- frontmatter
name: tao-deepresearch
description: Deep research for AI/ML research. Use when analyzing arxiv papers, conducting literature reviews, finding influential papers, identifying research trends, or exploring implementations. Handles LaTeX source analysis, web search for papers, and GitHub code discovery.
argument-hint: [topic]
allowed-tools: Read, Grep, Glob, WebSearch, WebFetch, Bash(gh *)

Tao Deep Research

Conduct comprehensive AI/ML research by analyzing provided materials and searching for related literature.

Arguments

  • $ARGUMENTS[0] - Research topic or question
  • --depth - Research depth: quick (overview), medium (detailed), thorough (comprehensive)
    • Default: thorough
  • --focus - Research focus: methods (algorithms/techniques), experiments (benchmarks/results), theory (foundations/proofs)
    • Default: methods

When no arguments are provided, use depth=thorough and focus=methods.

Input Types

The user may provide:

  1. LaTeX source from arxiv papers (analyze the "Related Works" section as starting point)
  2. Code for implementation context
  3. Topic description for open-ended research

Research Workflow

Step 1: Analyze Provided Materials

If LaTeX source is provided:

  • Parse the "Related Works" or "Related Work" section
  • Extract cited papers, authors, and key themes
  • Identify the research area and subfields
  • Note any code repositories mentioned

If code is provided:

  • Identify the methods/models implemented
  • Check for paper references in comments or README

Step 2: Search for Related Literature

Use web search to find:

  • Recent papers on the identified topics (prioritize last 2 years)
  • Highly-cited foundational papers
  • Survey papers for comprehensive coverage
  • Arxiv preprints for cutting-edge work

Search queries should include:

  • Key method names and algorithms
  • Problem domain keywords
  • Author names from seminal works

Step 3: Fetch and Analyze Papers

For important papers found:

  • Fetch arxiv abstracts and metadata
  • Identify citation counts and influence (when available)
  • Note publication venues (NeurIPS, ICML, ICLR, ACL, CVPR, etc.)
  • Track chronological development

Step 4: Search for Implementations

Search GitHub for:

  • Official implementations of key papers
  • Popular reimplementations with high stars
  • Relevant libraries and frameworks
  • Benchmark repositories

Step 5: Synthesize Findings

Analyze the collected information to identify:

  • Evolution of methods over time
  • Current state-of-the-art approaches
  • Emerging trends and directions
  • Open problems and challenges

Output Format

Generate a structured research report with these sections:

Key Papers

List the most important papers with:

  • Title, authors, year, venue
  • Arxiv link (if available)
  • Brief description of contribution
  • Relative influence/importance

Format:

code
**[Paper Title]** (Year)
Authors: [Names]
Venue: [Conference/Journal] | [arxiv:XXXX.XXXXX](https://arxiv.org/abs/XXXX.XXXXX)
Contribution: [1-2 sentence summary]

Methods/Approaches Overview

  • Categorize methods by type/approach
  • Explain key techniques and innovations
  • Compare strengths and limitations
  • Note computational requirements

Timeline/Trends Analysis

  • Chronological development of the field
  • Key breakthroughs and when they occurred
  • Current dominant approaches
  • Emerging directions and recent shifts

Open Problems/Future Directions

  • Unsolved challenges in the field
  • Limitations of current methods
  • Promising research directions
  • Potential applications

Code/Implementation References

  • Official repositories with links
  • Popular frameworks and libraries
  • Benchmark datasets and evaluation code
  • Format: [Repo Name](GitHub URL) - Brief description (stars if notable)

Depth Levels

Quick (5-10 papers):

  • Focus on most influential recent papers
  • Brief overview of methods
  • Key trends only

Medium (15-25 papers):

  • Comprehensive coverage of recent work
  • Detailed methods comparison
  • Full trend analysis

Thorough (30+ papers):

  • Exhaustive literature review
  • Historical context and evolution
  • Deep technical analysis
  • Multiple subcategories explored

Focus Areas

Methods: Prioritize algorithmic innovations, architectural designs, training techniques Experiments: Prioritize benchmarks, datasets, evaluation metrics, empirical results Theory: Prioritize mathematical foundations, convergence proofs, theoretical guarantees

Example Usage

code
/tao-deepresearch "GRPO training for reasoning models" --depth thorough --focus methods
code
/tao-deepresearch "efficient attention mechanisms" --depth quick

When LaTeX is provided:

code
User: [pastes LaTeX source]
/tao-deepresearch --depth medium --focus experiments

Research Best Practices

  1. Recency: Prioritize papers from the last 2 years for current state-of-the-art
  2. Influence: Weight highly-cited papers and papers from top venues
  3. Diversity: Include different approaches and perspectives
  4. Verification: Cross-reference claims across multiple sources
  5. Completeness: Cover both foundational work and recent advances