AI Paper Survey Skill

Structured, multi-phase paper survey workflow for AI research.

When to Use

• - "Survey recent papers in [topic]"
• "What's new in agent/LLM/multimodal research?"
• "Find the most important papers from the last N months"
• "Do a literature review on [topic]"
• "Track progress in [research area]"

Prerequisites

• - alphaXiv MCP server must be connected (provides embedding_similarity_search, full_text_papers_search, get_paper_content)
• paper-impact-analyzer skill installed (for impact assessment)
• Research keywords file (optional): a Markdown file listing the user's research interests and keywords

Workflow: 5-Phase Pipeline

Phase 0: Load Research Context

1. 1. Check if a research keywords file exists. Look for files matching patterns:

1. 2. If found, read it and extract:

1. 3. If no keywords file, ask the user for:

1. 4. Determine the time range (default: last 3 months from today).

1. 5. Generate search queries using the template below. For each user theme T, generate:

CODEBLOCK0

Phase 1: Broad Search (Parallel)

Execute search queries in parallel using alphaXiv MCP tools:

• - Use embedding_similarity_search for semantic queries (captures conceptual matches)
• Use full_text_papers_search for keyword queries (captures exact term matches)

Rules:

• - Launch 4-6 parallel searches covering different themes
• Each search returns up to 15 results
• Collect all results into a candidate pool
• Deduplicate by arXiv ID
• Filter by publication date (must be within the specified time range)

Expected output: 30-60 unique candidate papers with titles and abstracts.

Phase 2: Initial Screening (LLM Judgment)

For each candidate paper, classify by the user's framework. Default framework (3-tier):

• - Tier 1 (Essence): "What IS X?" — Redefines the problem itself. Asks fundamental questions about the nature of learning, reasoning, action, perception, etc. These papers have lasting impact because they challenge assumptions.
• Tier 2 (Engineering): "How to do X better?" — Optimizes within existing frameworks. Valuable but doesn't change paradigms. Examples: better MoE routing, improved training recipes, new benchmarks.
• Tier 3 (Patch): "How to mitigate this symptom?" — Short-term fixes. Inference token pruning, fine-tuning tricks, quantization improvements.

Rules:

• - Use ONLY title + abstract for screening (don't read full papers yet)
• Be selective: aim for 8-12 papers across all tiers
• Tier 1 should have 3-5 papers max
• Apply the user's specific keywords to boost relevance

Expected output: Classified paper list with tier assignments.

Phase 3: Deep Reading (Parallel, Top Candidates Only)

For Tier 1 and top Tier 2 papers (4-6 papers max), use get_paper_content to retrieve full analysis.

After reading each paper, immediately extract and cache:

• - Core contribution (1 sentence)
• Method keywords (3-5 terms)
• Best experimental result (1-2 numbers)
• Open-source links (GitHub URL if any)
• Venue acceptance status
• Key limitation

Discard the raw full-text analysis after extraction to manage context window.

Phase 4: Impact Assessment

For each paper in the deep reading set, run the paper-impact-analyzer:

CODEBLOCK1

Merge impact data with the content analysis from Phase 3.

Phase 5: Synthesize Report

Generate a structured Markdown report with the following sections:

CODEBLOCK2

Save the report to {working_directory}/{topic}-paper-survey-{date}.md.

Configuration

Custom Classification Framework

Users can override the default 3-tier framework by specifying their own in the keywords file. The skill will use whatever framework the user provides.

Search Depth Control

Default: Standard.

Example Usage

CODEBLOCK3

CODEBLOCK4

CODEBLOCK5