AI scientist:deep research - chunhualiao/public-docs GitHub Wiki

AI scientist

what are best open-source automatic paper writing github repos using Large language models? I want to automate scientific paper writing, in the domain of computer science (including AI)

Below is a curated, up-to-date shortlist of open-source repos that actually move the needle for automating scientific writing in CS/AI—from end-to-end “AI scientist” systems to literature-review generators, citation-aware drafting, and LaTeX-native workflows. I grouped them by what you’re trying to automate and called out why each is worth your time.

1) End-to-end “generate experiments → write paper” agents

AI-Scientist-v2 (Sakana AI) — end-to-end pipeline that ideates topics, runs experiments via agentic tree search, analyzes results, and writes the manuscript (PDF/LaTeX) with a citation phase. Designed for workshop-level outputs; comes with caution notes about sandboxing. Apache-2.0. ([GitHub]1)

When to use: you want maximum automation (including running code) and are comfortable curating/validating the results. Great for ML experimentation where you can sandbox compute.

2) Long-form reports / sections with grounded citations

STORM (Stanford OVAL) — a modular “knowledge curation” system that researches a topic on the web, auto-outlines, then writes a full-length, Wikipedia-style article with citations; actively maintained, MIT-licensed, with a Python package (knowledge-storm) and a demo. Excellent as a scaffold for Related Work / Background. ([GitHub]2)

GPT-Researcher — a mature multi-agent “deep research” tool that plans, crawls, aggregates sources, and outputs long research reports with citations (PDF/DOCX/Markdown). Apache-2.0; active releases and strong adoption. Useful as a drop-in “research + draft” microservice. ([GitHub]3)

When to use: you want a high-quality, traceable narrative (survey/position/background) grounded in live web sources and PDFs.

3) Literature-review & survey paper generators

LitLLM (Mila/ServiceNow) — a toolkit that turns an abstract/plan into a structured literature review using hybrid retrieval + plan-based generation; ships a web app and a paper describing the method. Apache-2.0. ([GitHub]4, [arXiv]5)

SurveyX (IAAR-Shanghai) — pipeline for automated academic survey writing (LaTeX/PDF outputs). The open-source version runs offline over your local references; the full online system adds search/multimodal parsing. Good if you already maintain a curated reference folder. ([GitHub]6)

SurveyForge (ACL-2025) — a newer system that improves outline quality, memory-driven generation, and multi-dimensional evaluation (SurveyBench) for AI-generated surveys. Code and demo papers available. ([GitHub]7)

InteractiveSurvey (2025) — interactive, personalized survey generation with PDF export (Markdown/LaTeX); Dockerized, supports GPU; designed for iterative refinement. ([GitHub]8)

When to use: you want to automate “Related Work/Survey” sections with controllable structure and citations, starting from an abstract + a seed bibliography.

4) Drafting while you type with accurate citations

ScholarCopilot (COLM-2025) — trained specifically for academic writing with context-aware next-sentence/section suggestions and real-time citation insertion (retrieval integrated into generation). Ships models, data, demo, and training code. MIT license. ([GitHub]9)

When to use: you prefer an editor-like experience that auto-suggests text and citations as you write sections.

5) LaTeX-first editing, polishing, and integration

Manubot AI Editor — GitHub-Actions-ready tool to run automatic, AI-assisted revisions over Manubot manuscripts (paragraph-level; OpenAI/Anthropic via LangChain). It opens PRs with diffs; great for enforceable, reviewable polishing. BSD-3-Clause. ([GitHub]10)

paper-qa (Future-House) — high-accuracy, citation-first RAG over your PDFs to answer questions with page-grounded citations. Ideal to verify claims before they enter your manuscript, or to draft factual paragraphs with sources attached. Very active, v5+ integrates LiteLLM. ([GitHub]11)

LeafLLM (Overleaf tools) — a lightweight way to bring LLM completions into Overleaf via a browser extension (OSS repo + Chrome extension). Good for in-place edits; keep version history on. ([GitHub]12, [Chrome Web Store]13)

When to use: you’re committed to LaTeX and want reliable revision loops or “ask-my-PDFs” grounding while writing.

Fast chooser (what to try first)

  • Need a full pipeline (experiments → paper): AI-Scientist-v2. Then use Manubot AI Editor to polish. ([GitHub]1)
  • Need a strong, cited draft for Background/Related Work: STORM or GPT-Researcher. If you already have a local paper corpus, add paper-qa for grounded paragraphs. ([GitHub]2)
  • Writing as you go with auto-citations: ScholarCopilot. ([GitHub]9)
  • Survey paper automation: start with LitLLM (interactive), or try SurveyX / SurveyForge for more structured survey generation. ([GitHub]4)
  • LaTeX polishing + safe CI: Manubot AI Editor; for on-the-fly Overleaf help, LeafLLM. ([GitHub]10, [Chrome Web Store]13)

How to assemble a pragmatic CS/AI paper-writer stack (OSS-first)

  1. Ingest & curate sources

    • Use STORM or GPT-Researcher to search + outline. Add your lab PDFs into a local corpus. ([GitHub]2)
    • Keep a RAG index for paper-qa so any paragraph you draft can be verified with page-level citations. ([GitHub]11)

  2. Draft the literature sections

    • Feed the outline to LitLLM (or SurveyX/SurveyForge) to produce a clean, sectioned Related Work/Survey. ([GitHub]4)

  3. Write methods/results & weave in citations

    • Draft interactively with ScholarCopilot to get section-level completions and automatically suggested citations/bibtex. ([GitHub]9)

  4. LaTeX polish & CI

    • Maintain the manuscript in Manubot; enable Manubot AI Editor via GitHub Actions to open PRs with paragraph-level revisions (style, clarity). ([GitHub]10)
    • If you live in Overleaf, keep LeafLLM in the loop for quick edits—always review diffs. ([Chrome Web Store]13)

  5. Quality gates

    • Before merging any section, run paper-qa queries like “List every quantitative claim in §3 and show the exact page and figure backing each” to catch hallucinations. ([GitHub]11)

Notes on maturity, licensing, and fit

  • Actively maintained & cited: STORM (MIT; NAACL/EMNLP lineage), GPT-Researcher (Apache-2.0; frequent releases), ScholarCopilot (COLM-2025; MIT), Manubot AI Editor (BSD-3). ([GitHub]2)
  • Survey writers are rapidly evolving; SurveyX and SurveyForge are promising for CS/AI topics but expect some glue work (data prep, model keys, local references). ([GitHub]6)
  • End-to-end agents like AI-Scientist-v2 can produce a full PDF, but you must sandbox execution and review the science—treat outputs as drafts. ([GitHub]1)

A few complementary utilities (optional)

  • Overleaf integration: LeafLLM for quick completions in place. ([Chrome Web Store]13)
  • Citation-first RAG for verification: paper-qa. ([GitHub]11)

Ethical & venue compliance (quick checklist)

  • Keep a changelog of automated edits (PRs from Manubot AI Editor make this easy). ([GitHub]10)
  • Ground all factual claims with line/page citations (paper-qa + your RAG index). ([GitHub]11)
  • Follow your venue’s authorship/AI-use policies; most CS venues now require disclosure for LLM assistance.

Want a ready-to-run starter?

If you’d like, I can generate a minimal scaffold that stitches STORM → LitLLM → ScholarCopilot → Manubot AI Editor with example configs and a Makefile so you can drop in a topic + PDFs and get a first Related Work draft, then iterate toward a full paper.