AutoScholarLoop

English | 中文

AutoScholarLoop is an open-source AUTO Research framework for building auditable, multi-agent research loops. It helps users turn an initial research direction, recent papers, reference notes, and optional code into a staged research process that can generate ideas, run execution loops, draft papers, review claims, and package submission candidates.

The project is developed for research automation scenarios at AI Group, CAS CNIC (Computer Network Information Center, Chinese Academy of Sciences).

AI Group

Description

AutoScholarLoop simulates a small research group rather than a single chatbot. The system separates research work into role-based stages:

1. S00 Field Archive Group builds the field map, paper cards, method map, dataset/baseline map, and evidence bank.
2. S01 Professor Decision Group runs multi-round idea generation, critique, novelty probing, ranking, and direction selection.
3. S02 PhD Execution Group plans baselines, implementation, experiments, failure analysis, and professor review memos.
4. S03 Writing Group turns evidence into a manuscript plan, claim-evidence table, draft, figures, and review-driven revisions.
5. S04 Quality Control Group audits novelty, citations, reproducibility, unsupported claims, and final release readiness.

The core design is a nested loop:

S00 evidence preparation
  -> S01 professor decision loop
  -> S02 execution-review loop
  -> S03 writing-review loop
  -> S04 quality gate
  -> submission_candidate / revise / pivot / kill

Every stage writes Markdown checkpoints and structured artifacts so users can inspect how an idea was created, why it was selected, what evidence supports the paper, and where quality gates passed or failed.

Features

• Multi-stage AUTO Research loop with explicit checkpoints.
• Deterministic local provider for offline demos and tests. It is not a real model-backed research run.
• OpenAI-compatible provider adapter for real model APIs.
• Local, Semantic Scholar, and OpenAlex literature adapters.
• Dry-run and local shell execution backends. The Web console defaults to local shell execution: model-generated code is written under code/ and then run on the current machine.
• Format-aware paper writing for acm, ieee, springer_lncs, and chinese_thesis.
• Markdown and LaTeX manuscript export.
• Optional PDF compilation if --compile-pdf is enabled and a LaTeX toolchain plus required venue class files are installed.
• Vue Web console for first-run model configuration, paper upload, live loop progress, and checkpoint preview.
• CLI and Web API use the same underlying research pipeline.

Installation

cd AutoScholarLoop
pip install -e ".[api,web,dev]"

For the Web frontend:

cd web
npm install

The frontend currently uses Vite 2 for compatibility with older Node environments. Node 18+ is recommended for future frontend upgrades.

CLI Quick Start

autoscholarloop run `
  --seed "I want to study retrieval-augmented agents for scientific writing." `
  --loop-mode fast `
  --paper-format ieee `
  --workspace runs/demo

Useful options:

autoscholarloop run `
  --seed "your research idea" `
  --reference "paper title, URL, local path, or note" `
  --num-ideas 5 `
  --loop-mode standard `
  --paper-format acm `
  --literature semanticscholar `
  --execution-backend dry-run `
  --review-ensemble 5 `
  --compile-pdf `
  --workspace runs/demo

The legacy command alias new-ai-scientist is kept for compatibility.

By default, local provider mode is a deterministic demo. For real research runs, configure an OpenAI-compatible model provider and API key. The system always writes paper/main.tex; PDF generation is attempted only when explicitly enabled and when the local LaTeX environment supports the selected format. The compiler prefers latexmk, then falls back to xelatex or pdflatex with multi-pass compilation.

If your network environment injects a system proxy that breaks httpx connections, set:

$env:AUTOSCHOLARLOOP_HTTP_TRUST_ENV='0'

Example for DeepSeek-compatible APIs:

$env:OPENAI_API_KEY='your_api_key'
$env:AUTOSCHOLARLOOP_HTTP_TRUST_ENV='0'
autoscholarloop run `
  --seed "your research idea" `
  --provider openai-compatible `
  --model deepseek-chat `
  --base-url https://api.deepseek.com/v1 `
  --workspace runs/deepseek_demo

Web Console

Start the Python API:

autoscholarloop web --host 127.0.0.1 --port 8000

Start the Vue frontend:

cd web
npm run dev

The Web console supports:

• first-run large model API configuration;
• provider presets for Local demo, DeepSeek, OpenAI, and custom OpenAI-compatible APIs;
• optional system proxy bypass for environments where httpx fails through inherited proxy settings;
• research direction and target venue input;
• PDF, Markdown, text, and BibTeX upload;
• loop mode and backend selection;
• manuscript format selection;
• live S00-S04 progress visualization;
• checkpoint preview for field maps, ideas, execution reports, paper plans, claim evidence, final gate, and final draft.

Generated Workspace

Each run creates an auditable workspace:

run/
  source_papers/
  inputs/
  artifacts/
  logs/
  code/
  00_field_context/
  01_decision/
  02_execution/
  03_writing/
  04_quality/
  paper/
  release/

Important outputs include:

• code/experiments/run_experiment.py
• code/methods/proposed_method.py
• code/experiments/result.json
• 00_field_context/field_map.md
• 00_field_context/paper_cards.md
• 01_decision/IDEA_REPORT.md
• 01_decision/chosen_direction.md
• 02_execution/RESULTS_ANALYSIS.md
• 02_execution/CLAIMS_FROM_RESULTS.md
• 02_execution/EXPERIMENT_AUDIT.md
• 03_writing/PAPER_PLAN.md
• 03_writing/claim_evidence_table.md
• 04_quality/CITATION_AUDIT.md
• 04_quality/final_gate.md
• 04_quality/compile_report.md
• paper/final_draft.md
• paper/main.tex
• paper/main.pdf if local LaTeX compilation succeeds
• release/README.md

Paper Formats

Supported manuscript targets:

• acm: ACM-style conference or journal article draft.
• ieee: IEEE conference or journal article draft.
• springer_lncs: Springer LNCS proceedings-style draft.
• chinese_thesis: generic Chinese thesis-style manuscript.

Official venue class files and bibliography rules still need to be checked before real submission. The generated manuscript is a research draft and audit package, not a guarantee of venue compliance.

Repository Layout

docs/                         Design, roadmap, workflow, and version notes
src/open_research_agent/       Python research loop package
web/                          Vue Web console
configs/                       Example pipeline configs
templates/                     Research workspace templates
examples/                      Example inputs
tests/                         Smoke tests

Development

Run tests:

$env:PYTHONPATH='src'
python -m pytest tests -q

Build the Web frontend:

cd web
npm run build

Project Status

AutoScholarLoop currently provides a runnable research-loop scaffold with auditable outputs. It is designed for iterative extension. It does not yet guarantee true scientific novelty, correct citations, valid experiments, or submission-ready papers without human supervision.

Use it as a research automation assistant, not as a replacement for scientific judgment, domain expertise, peer review, or responsible authorship.

Acknowledgements

AutoScholarLoop is inspired by and learns from several open research automation and paper-writing projects, including:

• AI-Scientist-main, for demonstrating an end-to-end AI scientist workflow.
• academic-paper-writer-main, for paper format and manuscript-generation workflow ideas.
• Codex and ClaudeCode style coding-agent workflows, for skill-oriented automation patterns.
• Open literature infrastructure such as Semantic Scholar and OpenAlex, for retrieval and bibliography-oriented research workflows.

This repository is an independent implementation. Architecture, code, stage contracts, and Web UI are written for AutoScholarLoop.

Organization

Developed for:

CAS CNIC
Computer Network Information Center, Chinese Academy of Sciences

License & Responsible Use

This project is licensed under The AI Scientist Source Code License, a derivative of the Responsible AI License.

Mandatory Disclosure: By using this code, you are legally bound to clearly and prominently disclose the use of AI in any resulting scientific manuscripts or papers.

We recommend the following attribution in your paper's Abstract or Methods section:

"This manuscript was autonomously generated using AutoScholarLoop, an AI-assisted research-loop system inspired by The AI Scientist."

Users are responsible for verifying all claims, citations, experiments, authorship requirements, venue policies, and disclosure obligations before submitting any generated manuscript.