Literature Review Sprint

Purpose

Turn "I need to read more papers" into a bounded, systematic review. This skill follows the handbook's literature review advice: search systematically, prioritize high-quality sources, assess institution/year/venue/impact, summarize key ideas and limitations, and stay current with top conferences and recent papers.

The output is a usable map of the field, not a long undigested bibliography.

When to Use

• User is starting a new topic
• User wants to verify novelty of an idea
• User is writing related work
• User has many papers and no structure
• User needs a reading plan before an advisor or group meeting

Workflow

Stage 1: Define the Review Question

Ask for:

• Topic or problem
• Target field or venue family
• Purpose: novelty check, related work writing, method comparison, dataset search, or general catch-up
• Deadline and available reading time
• Existing seed papers, if any

Force a narrow review question. "Diffusion models" is too broad; "diffusion policies for robot manipulation under distribution shift" is workable.

Stage 2: Build the Search Plan

Create search queries for:

• Google Scholar / Semantic Scholar
• arXiv
• Top venue proceedings in the field
• Hugging Face Papers or similar recent-paper hubs for ML topics
• References and citations of 2-3 seed papers

For current topics, prioritize the last 2 years first, then trace backward to canonical papers.

Stage 3: Triage Papers

Classify papers into:

• Core: must read carefully
• Context: skim for framing or related work
• Maybe: keep if time remains
• Reject: irrelevant, low quality, or superseded

Assess each paper using:

• Institution or lab
• Year
• Venue
• Citation/impact signal
• Method relevance
• Evaluation relevance
• Clear limitation or gap

Stage 4: Extract Notes

For each core paper, produce compact notes:

• One-sentence contribution
• Problem setting
• Key method
• Dataset/task
• Baselines
• Main result
• Limitation
• How it relates to the user's project

Keep notes short enough that they remain searchable.

Stage 5: Synthesize the Map

Create a field map organized by axes that matter for the topic, such as:

• Method families
• Dataset/task families
• Assumptions
• Evaluation metrics
• Failure modes
• Open gaps

Name 3-5 concrete gaps or tensions. Separate real gaps from "I have not read enough yet."

Stage 6: Produce the Artifact

Save to ~/phd-log/literature/YYYY-MM-DD-[topic].md.

# Literature Review Sprint — [Topic]

## Review question
[Specific question]

## Search plan
- Queries:
- Venues / sources:
- Seed papers:

## Paper triage
| Title | Year | Venue | Category | Why |
|---|---:|---|---|---|

## Core paper notes
### [Paper title]
- Contribution:
- Setting:
- Method:
- Evaluation:
- Main result:
- Limitation:
- Relevance:

## Field map
- Method families:
- Evaluation setups:
- Common assumptions:
- Failure modes:

## Candidate gaps
1. [gap + evidence]
2. [gap + evidence]
3. [gap + evidence]

## Next actions
- [ ] Read:
- [ ] Reproduce/check:
- [ ] Ask:
- [ ] Update idea/related work:

Tone

Be structured and skeptical. The user should leave with a smaller, clearer reading list and a sense of what matters.

What Not to Do

• Do not produce generic paper summaries without synthesis.
• Do not treat citation count as the only quality signal.
• Do not over-prioritize old canonical papers when the user needs current trends.
• Do not let the user claim novelty before checking recent top venues and arXiv.