Literature Research

Step 2 — Build query set

Design 5–10 diverse queries covering the topic from different angles:

• Include synonyms and related terms.
• Vary abstraction level: specific (e.g. "GPT-4 code smell detection") and general (e.g. "LLM software quality").
• For arXiv: target categories like cs.SE, cs.PL, cs.LG as appropriate.

Step 3 — Execute searches in parallel

Use /fleet to run multiple searches simultaneously. Each parallel agent should:

1. Fetch papers using the arXiv API (see below).
2. Fetch papers using the Semantic Scholar API (see below).
3. Return results as a structured JSON list.

Step 4 — Deduplicate

Merge all results. Deduplicate by normalising titles (lowercase, strip punctuation) and comparing. Keep the entry with the most metadata (abstract, DOI, URL).

Use a session SQL table to track seen titles:

CREATE TABLE IF NOT EXISTS lit_papers (
    id TEXT PRIMARY KEY,
    title TEXT,
    authors TEXT,
    year INTEGER,
    abstract TEXT,
    url TEXT,
    doi TEXT,
    citations INTEGER DEFAULT 0,
    venue TEXT,
    source TEXT,
    section TEXT
);

Step 5 — Filter

Apply:

• Year range filter.
• Relevance check: does the title/abstract mention the topic keywords?
• For LLM-related surveys: require at least one of: GPT, Copilot, LLM, large language model, CodeLlama, Claude, DeepSeek, Gemini, Codex in title or abstract.

Step 6 — Generate report

Produce a structured Markdown report (see Report Format section below).

API Reference

arXiv API

GET https://export.arxiv.org/api/query
  ?search_query=<QUERY>
  &start=0
  &max_results=<N>
  &sortBy=submittedDate
  &sortOrder=descending

Query syntax:

• all:keyword — search all fields
• ti:keyword — title only
• abs:keyword — abstract only
• cat:cs.SE — restrict to category
• AND, OR, ANDNOT — Boolean operators (surround with +: +AND+, +OR+, +ANDNOT+)
• + between terms = space = OR — to AND two terms, use +AND+ between separate field:term pairs
• Phrase search: URL-encode quotes as %22, e.g. ti:%22code+generation%22

Example: