Open sourceMIT · Python · Claude Code

Literature surveys
that stay alive.

SurveyAtlas turns a research field into a website you can search, slice and question. LLM agents harvest, classify and deep-read every paper, verify every reported number, and rebuild the survey as new work appears.

Open the live NavAtlas Watch the demo Code
3,147papers in NavAtlas
2,694deep-read from full text
9,806benchmark results
97.6%found verbatim in papers

Demo

SurveyAtlas in seventy seconds

What it is, how an atlas is built, the two-command quick start, and a tour of every page of the public NavAtlas: library, year range, authors, paper pages, Ask, map, timeline, leaderboards, surveys and pipeline.

billzhao1030.github.io/SurveyAtlas/navatlas

Why

A survey paper freezes the day it is submitted

Anyone can ask an LLM to summarize a stack of papers. What stays valuable is a checked, structured view of a field: what belongs to it, which routes exist, which numbers can be compared and where the gaps are. SurveyAtlas keeps that view current.

The usual way

A survey paper

  • Out of date after a few months in a field that publishes daily
  • One fixed taxonomy and one reading order
  • Results tables copied by hand, protocols silently mixed
  • Cannot answer the question you actually have
SurveyAtlas

A living survey website

  • New papers harvested, classified and read on a schedule
  • Slice by task, paradigm, year, author, venue, benchmark and more
  • Every number checked against the paper; subsets never ranked with full splits
  • Ask in plain language; answers cite the papers

Features

Everything a survey gives you, and what it cannot

Survey drafting

Live outlines whose sections are queries; ./atlas survey writes evidence packs, LaTeX tables and figures, and a PDF draft.

Export anything

BibTeX, CSV or \cite{} for any selection; official BibTeX from DBLP and CrossRef where it exists.

Updates itself

./atlas cron install --daily --publish harvests, classifies, reads and redeploys the public site every day.

One file per field

Queries, taxonomy, classifier rules, benchmark protocols and landmarks live in a single atlas.py.

Built with Claude Code

No API key: classification, deep reading and Ask run on your Claude Code login. A bundled skill builds new atlases.

Tiny footprint

Python standard library plus requests; a vanilla-JS site with no build step; static export for GitHub Pages.

How it works

From twenty thousand records to a verified survey

Recall first, precision from the classifier, trust from verification. The numbers are NavAtlas's own funnel.

01
20,480records harvested

Harvest

About fifty arXiv queries plus OpenAlex, written for recall rather than precision.

02
10,897after rule prefilters

Merge & prefilter

Duplicates across sources and versions are merged; cheap rules drop what is obviously off-topic.

03
3,147papers kept in scope

Classify with Claude

Headless Claude labels 40 papers per call into the field taxonomy; Opus re-judges the uncertain ones.

04
2,694papers deep-read

Deep read

Each paper's full text becomes structured notes: problem, key idea, method, insight, results, limitations.

9,806

reported results, verified

97.6% of the numbers are found verbatim in the paper text, and an LLM judge flags protocol deviations, so subsets are never ranked with full splits.

Daily

rebuilt and published

Leaderboards, maps and BibTeX are rebuilt on a schedule, and the static site redeploys from a 23 MB snapshot with no API calls.

NavAtlas today

The field, as of the last rebuild

Vision-and-language navigation and embodied navigation: from task-specific policies and large-scale pretraining to fine-tuned foundation models, zero-shot pipelines and agentic navigators.

Core papers per year, by paradigm

Click a bar to open those papers in the live atlas.

Most-used benchmarks

Core papers that evaluate on each.
20,480records harvested
2,212core papers
10,237authors, each searchable
1,849with a resolved venue

Build your own

Your field, as a living survey, in an afternoon

Describe the fieldWrite a short brief. Claude drafts the queries, taxonomy, classifier rules and landmark papers into one atlas.py.
Build itHarvest, classify, resolve venues and build. About two hours for ten thousand candidates; a few minutes per update after that.
Read and publishOptional deep reading fills the leaderboards; ./atlas export and the bundled workflow put it on GitHub Pages.
Or open Claude Code and just say build a literature atlas for LLM agents. The bundled skill knows the playbook.
~/SurveyAtlas

  






  
Landscape

  
How it compares

  

    
Self-hosted & openField taxonomyUpdates itselfVerified leaderboardsQ&A with citationsSurvey drafting

    
      
SurveyAtlas✓✓✓✓✓✓

      
Awesome-lists✓by handby hand———

      
Papers with Code (closed 2025)—partial✓unverified——

      
Connected Papers · Litmaps · ResearchRabbit——✓———

      
Elicit · Consensus · Undermind——✓—✓—

      
PaperQA2 · OpenScholar✓———✓—

      
AutoSurvey · SurveyX · STORM✓—one shot——✓

    






  
Cite

  
If it helps your research

  
@software{surveyatlas2026,
  title  = {SurveyAtlas: Living Literature Surveys for Fast-Moving Research Fields},
  author = {Zhao, Xunyi},
  year   = {2026},
  url    = {https://github.com/billzhao1030/SurveyAtlas}
}