Skip to content

Repository files navigation

ARISE: Agentic Rubric-guided Iterative Survey Engine for Automated Scholarly Paper Generation

ARISE is an Agentic Rubric-guided Iterative Survey Engine for automated survey generation. It integrates citation-aware writing agents with a rubric-based evaluation loop to produce high-quality, self-improving survey papers.

🌐 Project Highlights

  • Agentic Writing System: Decomposes survey writing into agent-driven phases for retrieval, drafting, editing, and refinement.
  • Rubric-guided Agentic Peer-review: Uses structured reviewer feedback (from GPT-4.1, Gemini 2.5, Claude 3.7) to guide stepwise improvements.
  • Iterative Refinement: Refinement is driven purely by agents' feedback.
  • PDF Comparison Suite: Multiple output rounds show how the refinement process improves draft quality in stages.

📁 Folder Structure

.
├── ARISE_Source_Code/                 # Full pipeline, experiment, and evaluation source code and scripts. 
├── agent_prompt                        # All prompt definitions used by ARISE agents
├── full_detailed_review_rubric         # The complete rubric (7 categories, 20+ sub-criteria)
├── SMALL_model_sample_by4.1mini/       # Sample surveys and outputs generated by the small (GPT-4.1 mini) model
├── ARISE_sample_surveys_ouput_gpt4.1/ # Sample surveys and outputs generated by GPT-4.1 (main model)
├── experiment/                         # All experiment results, including comparison tables and figures
├── human_evaluation/                   # Results and tables from blinded human expert evaluation
├── Dataset/                            # Baseline and comparison papers, review results for all systems
│   ├── autosurvey/                         # Outputs from the Autosurvey baseline
│   ├── dataset_review_result/
│   │   ├── review_autosurvey/               # Review logs and chunked review for Autosurvey
│   │   ├── review_published/                # Review logs for human-written published surveys
│   │   ├── review_surveyforge/              # Review logs for SurveyForge
│   │   └── review_surveyx/                  # Review logs for SurveyX
│   ├── published/                           # Human-written (published) survey papers (PDF or LaTeX)
│   ├── surveyforge/                         # Outputs from SurveyForge baseline
│   └── surveyx/                             # Outputs from SurveyX baseline
├── README                                  # This file
└── .gitignore

📑 Folder Details

gpt4.1 sample paper/ All sample surveys and outputs were generated using GPT-4.1 as the writing model. For each survey topic (e.g., "Disease detection across modalities"), you will find:

partial_outlines.md Markdown outline snippets used by agents during intermediate survey drafting.

completed_research_paper.md and completed_research_paper.tex Final, post-refinement survey draft in both Markdown and LaTeX formats.

survey_paper.tex / survey_paper_outline_one_by_one.md Initial and intermediate versions of the survey draft and outline.

citation.csv, summaries_dedup.csv, references.bib, topics.csv All source citations, deduplicated summary data, BibTeX references, and topic lists for the survey.

review_output/ Automated review logs and scores for every refinement round:

round{n}_section_feedback.csv: Section-by-section feedback from reviewer agents.

round{n}_category_scores.csv: Rubric-aligned category scores (e.g., originality, clarity, coverage).

round{n}_overall_evaluation.csv: Overall evaluation and high-level judgment from each reviewer agent.

score_summary.csv: Tracker of overall score progression across rounds.

total_score.txt: Final cumulative score for the paper (target ≥ 90%).

pdf_output/ Compiled PDFs and logs for every round of refinement:

final_paper_round{n}.pdf: Compiled survey PDF for each refinement round (e.g., round0 is initial, roundN is final).

final_paper_round{n}.tex: Raw LaTeX file for each round.

.aux, .bbl, .blg, .out, .log: Auxiliary files from LaTeX build for transparency and full reproducibility.

SMALL_model_sample_by4.1mini/ All surveys and outputs were generated by running the full system pipeline using GPT-4.1 mini as the agent. Useful for demonstrating robustness to smaller models.

experiment/ Contains all comparison results, ablation experiment data, and generated figures referenced in the paper and supplement.

Dataset/ Includes all baseline and comparison papers:

Human-written published surveys

Papers generated by baseline systems (e.g., Autosurvey, SurveyForge, SurveyX)

Subfolders contain review results and reviewer logs for each baseline.

🔬 Evaluation & Reproducibility

Human Evaluation: See human_evaluation/ for blinded expert reviewer results (4 reviewers, 5 topics, before/after RL refinement).

Ablation Studies: See SMALL_model_sample_by4.1mini/ and corresponding tables/plots in experiment/ for robustness with smaller models.

Baseline Reviews: All datasets and reviews for human and automated baselines are organized under Dataset/.

experiment_result.py: Full experimental results, reviewer scores, and system outputs are available in this repository to support transparency and reproducibility.

Source code

Environment Setup

1. Install Python Requirements

pip install -r requirements.txt

2. Install LaTeX

You must have a working LaTeX toolchain (pdflatex, bibtex).

sudo apt install texlive-full

Configuration

Updated the API_KEY for LLMS in .env file with the following content:

SERPER_API_KEY=YOUR_SERPER_API_KEY PAPER_COUNT=15 # Number of papers per topic JOURNAL=acm # Target journal OPENAI_API_KEY=YOUR_OPENAI_API_KEY GEMINI_API_KEY=YOUR_GEMINI_API_KEY MODEL=gpt-4.1 # (e.g., gpt-4.1, gpt-4.1-mini, gemini-1.5-pro, etc.) TEMPERATURE=0.7 TOPIC= # Topic will be set in the first script or batch run ANTHROPIC_API_KEY=YOUR_ANTHROPIC_API_KEY

Usage

One-click: Full Pipeline

Run everything from scratch:

python run_all.py

Phase-by-phase

You can also run each phase individually depending on your needs (e.g., search, summarization, LaTeX build, BibTeX conversion, etc.). See the individual scripts in ARISE_Source_Code/ for details and configuration options.

Notes

First Stage (0userinput.py): Requires human input for topic selection, ensuring only approved topics proceed through the pipeline.

CrewAI Environment: Make sure all CrewAI and LLM provider packages are installed and correctly configured.

Script Flexibility: All scripts can be run independently for debugging, custom runs, or partial pipeline execution.

Custom Topics & Models: To switch topics, model backends, or paper counts, update the .env file as needed.

Questions or issues? Please see the README or open an issue in this repository.

About

ARISE: Agentic Rubric-guided Iterative Survey Engine for Automated Scholarly Paper Generation

Resources

Stars

5 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages