ai-paper-survey（AI 论文调查）

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

- 研究关键词*.md - research-keywords*.md - research-interests*.md in the current working directory.

1. If found, read it and extract:

- Theme list: the major research themes (e.g., "RL optimization", "Agent & Tool Calling") - Keywords: specific terms to search for (e.g., "GRPO", "Nested Learning", "VLA") - Models of interest: specific model names (e.g., "DeepSeek V4", "Qwen3.5")

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

- Research topics (1-5 topics) - Time range (default: last 3 months) - Any specific papers or authors to track

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

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

Semantic query:  "Fundamental advances in {T}, paradigm shift, redefine {T}, {year}"
Keyword query:   "{specific_keywords_from_T} {year_range}"
Contrast query:  "Alternative to {current_paradigm_of_T}, beyond {T}, {year}"

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:

python path/to/paper-impact-analyzer/scripts/analyze.py {arxiv_id_1} {arxiv_id_2} ...

Merge impact data with the content analysis from Phase 3.

Phase 5: Synthesize Report

Generate a structured Markdown report with the following sections:

# {Topic} Paper Survey — {Date Range}

> Survey date: {today}
> Scope: {themes covered}
> Papers screened: {N candidates} → {M selected}

## Classification Framework
{Describe the tier system used}

## Tier 1 (Essence): Redefining the Problem
### Paper 1: {Title}
- **Essential question**: What fundamental assumption does this challenge?
- **Core contribution**: {1 sentence}
- **Key result**: {best number}
- **Impact**: {rating from analyzer} | {venue} | {github stars}
- **Links**: arXiv | GitHub
{... repeat for each Tier 1 paper}

## Tier 2 (Engineering): Doing It Better
| Paper | Contribution | Impact | Links |
|-------|-------------|--------|-------|
{table rows}

## Tier 3 (Patches): Symptom Relief
| Paper | What it fixes | Links |
|-------|--------------|-------|
{table rows}

## Top 3 Recommended Papers
{Ranked list with justification combining content depth + impact signals}

## Trends & Observations
{2-3 paragraphs on emerging patterns}

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

Survey the last 3 months of papers in my research areas

Quick survey: what's new in LLM reasoning and agent tool-calling since January?

Thorough literature review on RL training methods for LLMs, classify by innovation tier