Literature Review (arXiv)

Overview

Turn a research topic into a synthesized, cited review: search arXiv, read a focused set of key papers in chunks, and organize findings by theme. Built on the arxiv-toolkit (arxiv-mcp tools, or the arxiv CLI as a fallback).

Core principle: narrow and synthesize. Read a right-sized set of papers well, organized by theme — not a paper-by-paper dump, and not 80 tool calls.

When to use

• "Review the literature on X" / "survey of X" / "related work for X"
• "What's the state of the art / recent advances in X?"
• Orienting in a new area, or drafting a related-work section.

Prerequisite: the arxiv-mcp server is registered (the five tools below), or the arxiv CLI is installed.

Workflow

1. Scope first — do not skip. Before any search, settle: the key terms, the likely arXiv categories (e.g. cs.CL, cs.LG — see references/arxiv-query-syntax.md), the time window, and the depth. If the request is broad, ask ONE scoping question: a quick orientation (~5 papers), a standard review (~8–12), or exhaustive? Diving in without this wastes dozens of calls.

2. Discover. arxiv_search with field scoping (category, title, abstract) and the right sort — relevance for foundational work, submittedDate for the recent frontier. Run 2–3 complementary searches (broad terms; a title-scoped pass for surveys; a recent-sorted pass). If total is large or hints fire, narrow by category/date — never blind-paginate with `start`.

3. Triage. arxiv_get_metadata(ids, bibtex: true) on the candidates. Rank by fit + recency. Pick a deep-read set sized to the scope (default 5–8); list the rest as "also relevant."

4. Read — targeted, not exhaustive. For each chosen paper, read the abstract, then only the sections that matter (usually method + results) via arxiv_read_paper(section=…), or maxChars + nextCursor to walk a long section. After each paper, distill to 3–5 notes (contribution, method, result, limitation) and drop the raw text. Don't read every section of every paper.

5. Synthesize by theme. Group approaches into themes, contrast them, and surface agreements, trends, and open gaps. The output is a synthesis, not an annotated list.

6. Cite. Emit a BibTeX bibliography from the metadata already gathered. Offer arxiv_download for PDFs.

Tools

CLI equivalents: arxiv search / get / read / recent / download (see the toolkit README).

Output

Overview → thematic sections with inline citations (Author et al., YEAR, arXiv:ID) → a short comparison (a table when speedups/metrics invite one) → open problems → a BibTeX references block. Scale the length to the scope from step 1.

Common mistakes

• Over-fetching. Reading 15+ papers in full is dozens of rate-limited calls — the toolkit deliberately spaces requests (~3 s per host), so that plan is minutes of waiting and it crowds out the synthesis. Right-size the deep-read set; read targeted sections.
• Blind deep-paging. A large result set means narrow the query (category/date), not walk start.
• Paper-by-paper dump. Organize by theme, not "Paper 1… Paper 2…".
• No citations. Gather BibTeX as you go (arxiv_get_metadata(bibtex: true)), not at the end from memory.
• Skipping scope. Clarify depth/breadth before searching.