AgeSci is a fully autonomous system that generates complete IEEE-format academic research papers from a single topic input. The system orchestrates multiple specialized AI agents through local large language models via Ollama, with each agent responsible for distinct aspects of the research paper creation pipeline: architectural planning, literature research, academic writing, peer review, diagram generation, bibliography management, and LaTeX compilation.
The name "AgeSci" reflects the project's core philosophy: Agents doing Science.
- Overview
- Key Features
- System Architecture
- Installation
- Dependencies
- Usage
- Configuration
- Output Structure
- Agent System
- Citation Pipeline
- Quality Control
- Example Output
- Limitations and Considerations
- Troubleshooting
- Contributing
- License
- Citation
Academic paper writing traditionally demands extensive literature review, meticulous structuring, proper citation management, and strict adherence to formatting standards. AgeSci automates this entire workflow by deploying a coordinated team of specialized AI agents that collaborate to produce publication-ready documents.
Given a research topic as input, the system autonomously performs the following operations:
- Designs a comprehensive paper structure with appropriate sections and subsections
- Extracts domain-specific search terms using LLM-powered keyword analysis
- Queries academic databases (arXiv, Semantic Scholar) for relevant citations
- Writes each section with formal academic tone and proper LaTeX formatting
- Conducts iterative peer review with scoring and revision cycles
- Generates TikZ diagrams for methodology and architecture visualization
- Compiles the final LaTeX document with complete bibliography
- Produces a PDF output ready for review or submission
All citations are sourced from real academic papers through the arXiv and Semantic Scholar APIs, ensuring verifiability and academic integrity.
Unlike generative systems that fabricate references, AgeSci retrieves actual papers from established academic databases:
- arXiv API: Direct access to preprints across physics, mathematics, computer science, and related disciplines
- Semantic Scholar API: Comprehensive paper metadata including citation counts, abstracts, and author information
- LLM-Powered Keyword Extraction: Dynamic extraction of domain-specific search terms for any research topic, enabling the system to work across diverse fields
The system employs seven specialized agents, each with distinct responsibilities:
| Agent | Primary Function |
|---|---|
| Architect | Designs paper structure and section organization |
| Scholar | Writes publication-quality LaTeX content |
| Critic | Evaluates sections against IEEE standards |
| Artist | Generates TikZ diagrams and visualizations |
| Librarian | Manages BibTeX bibliography generation |
| Typesetter | Resolves LaTeX compilation errors |
| Integrator | Ensures document coherence and consistency |
Each section undergoes systematic review:
- Multi-dimensional scoring across technical depth, clarity, citations, formatting, and academic tone
- Iterative revision based on critic feedback with configurable thresholds
- Maximum revision attempts to prevent infinite loops
- Optional grammar checking via language-tool-python
Full compatibility with reasoning-enabled models including Qwen3, DeepSeek-R1, and DeepSeek-V3:
- Real-time streaming of model reasoning processes
- Complete visibility into agent decision-making
- Separate logging of thinking traces for debugging and analysis
Professional terminal output using the Rich library:
- Colored panels with clear visual hierarchy
- Progress indicators for long-running operations
- Structured tables for configuration and results display
- Real-time streaming of model outputs during generation
The generation pipeline follows a sequential flow with feedback loops for quality control:
Input Topic
|
v
+---------------------+
| Architect |
| (Structure Design) |
+---------------------+
|
v
+---------------------+
| Keyword Extractor |
| (Search Term Gen) |
+---------------------+
|
v
+--------------------+--------------------+
| |
v v
+------------------+ +--------------------+
| arXiv API | | Semantic Scholar |
+------------------+ +--------------------+
| |
+--------------------+--------------------+
|
v
+---------------------+
| Scholar |
| (Section Writing) |
+---------------------+
|
v
+---------------------+
| Critic |
| (Peer Review) |
+---------------------+
|
+----------+----------+
| |
[PASS] [REVISE]
| |
| v
| +---------------------+
| | Scholar |
| | (Revision) |
| +---------------------+
| |
+----------+----------+
|
v
+---------------------+
| Artist |
| (Diagram Creation) |
+---------------------+
|
v
+---------------------+
| Librarian |
| (Bibliography) |
+---------------------+
|
v
+---------------------+
| Typesetter |
| (Error Fixing) |
+---------------------+
|
v
+---------------------+
| PDF Compiler |
+---------------------+
|
v
Final Output
Before installation, ensure the following are available on your system:
- Python 3.10 or higher
- Ollama installed and running
- A compatible LLM model pulled in Ollama
- LaTeX distribution (optional but recommended for PDF compilation)
git clone https://github.com/Eeman1113/AgeSci.git
cd AgeSciInstall required packages using pip:
pip install ollama requests rich duckduckgo-search language-tool-pythonAlternatively, use the requirements file if provided:
pip install -r requirements.txtFollow the installation instructions at ollama.ai for your operating system.
After installation, pull a compatible model:
# Recommended for thinking mode support
ollama pull qwen3:latest
# Alternative with vision capabilities
ollama pull qwen3-vl:latest
# For enhanced reasoning tasks
ollama pull deepseek-r1:latestPDF compilation requires a LaTeX distribution. Installation varies by platform:
macOS:
brew install --cask mactexUbuntu/Debian:
sudo apt install texlive-fullFedora/RHEL:
sudo dnf install texlive-scheme-fullWindows:
Download and install MiKTeX from miktex.org or TeX Live from tug.org/texlive.
| Package | Purpose | Installation |
|---|---|---|
| ollama | Local LLM interface | pip install ollama |
| requests | HTTP requests for APIs | pip install requests |
| rich | Terminal formatting | pip install rich |
| duckduckgo-search | Web search fallback | pip install duckduckgo-search |
| Package | Purpose | Installation |
|---|---|---|
| language-tool-python | Grammar checking | pip install language-tool-python |
Note: The grammar tool downloads approximately 1GB of language data on first use.
Run the main script and provide your research topic when prompted:
python research.pyThe system displays current configuration and prompts for input:
Enter your research topic: [Your detailed topic description]
The system performs best with detailed topic descriptions that include:
- The specific problem or research question being addressed
- Relevant methods, techniques, or approaches to consider
- The target domain or application area
- Any constraints or specific angles to explore
Example Input:
Selective Concept Unlearning Without Catastrophic Forgetting: Most current
methods like gradient ascent or ROME-style edits struggle to surgically
remove specific knowledge without degrading general capabilities. A paper
proposing a new technique that better preserves model utility while ensuring
complete removal in LLMs.
During execution, the system displays:
- Current phase and agent activity
- Real-time model thinking (if enabled)
- Search results and citation counts
- Review scores and revision status
- Compilation progress and any errors
System behavior is controlled through the Config class at the top of the main script. Key parameters include:
model: str = "qwen3:latest" # Ollama model identifier
temperature: float = 0.7 # Generation temperature (0.0-1.0)
enable_thinking: bool = True # Enable thinking mode for compatible modelsoutput_dir: str = "research_output" # Directory for generated files
main_tex: str = "main.tex" # Primary LaTeX filenamepdf_compiler: str = "pdflatex" # LaTeX compiler command
compile_timeout: int = 60 # Maximum compilation time in secondsmax_rewrites: int = 3 # Maximum revision attempts per section
min_citations_per_section: int = 3 # Minimum citations required
max_arxiv_results: int = 15 # Maximum papers to retrievemax_retries: int = 3 # API retry attempts on failure
retry_delay: float = 1.0 # Delay between retries in seconds
enable_checkpointing: bool = True # Save intermediate progressenable_grammar_check: bool = True # Enable grammar validation
enable_latex_lint: bool = True # Enable LaTeX syntax checking
verbose_logging: bool = True # Enable detailed logging
log_thinking: bool = True # Log model reasoning to file| Model | Thinking Support | Recommended Use Case |
|---|---|---|
| qwen3:latest | Yes | General paper generation |
| qwen3-vl:latest | Yes | Topics requiring visual reasoning |
| deepseek-r1:latest | Yes | Complex technical topics |
| llama3:latest | No | Fast generation, simpler topics |
| mistral:latest | No | Balanced speed and quality |
All generated files are saved to the configured output directory (default: research_output/):
research_output/
main.tex # Complete LaTeX document
references.bib # BibTeX bibliography file
main.pdf # Compiled PDF (if LaTeX available)
research.log # Progress and status log
thinking.log # Model reasoning traces
verbose.log # Complete input/output log
plan.pkl # Cached paper structure
research.pkl # Cached research results
sections.pkl # Cached section content
main.tex: The complete LaTeX document containing all sections, figures, tables, and formatting. Uses IEEEtran document class with standard academic packages.
references.bib: BibTeX entries for all cited papers. Includes complete metadata (title, authors, journal, year, URL) for each reference.
main.pdf: The compiled PDF document. Generated only if a LaTeX distribution is installed and compilation succeeds.
Checkpoint Files (.pkl): Serialized Python objects containing intermediate results. Enable resumption of interrupted generation and caching of expensive operations.
Log Files: Detailed records of the generation process for debugging and analysis.
The Architect designs the paper structure based on the input topic:
- Generates a formal academic title
- Writes a 150-200 word abstract
- Identifies appropriate keywords
- Defines section organization with descriptions
- Marks sections requiring diagrams
- Estimates word counts for each section
Output format: JSON specification parsed by downstream agents.
The primary content generation agent:
- Writes LaTeX-formatted section content
- Maintains formal academic tone throughout
- Incorporates provided citations appropriately
- Formats equations, tables, and itemized lists
- Follows IEEE style guidelines
The Scholar receives citation context and section descriptions from upstream agents.
Evaluates written content against quality criteria:
- Technical Depth (0-10): Substantive content and analysis
- Clarity (0-10): Organization and readability
- Citations (0-10): Appropriate use of references
- Formatting (0-10): Correct LaTeX syntax
- Academic Tone (0-10): Professional language
Sections scoring below 35/50 or flagged with major issues trigger revision cycles.
Generates TikZ diagrams for visual content:
- Architecture diagrams for methodology sections
- Flowcharts for process descriptions
- Data visualization frameworks
- Uses standard TikZ libraries for compatibility
Manages bibliography generation:
- Formats BibTeX entries from paper metadata
- Validates required fields (title, author, year)
- Handles special characters and escaping
- Detects and manages duplicate entries
Resolves LaTeX compilation issues:
- Identifies syntax errors from compiler output
- Escapes special characters (%, &, $, #, _)
- Fixes malformed environments
- Repairs broken table and figure structures
- Preserves content while correcting markup
Ensures document-wide coherence:
- Verifies consistent terminology usage
- Checks logical flow between sections
- Identifies redundant or contradictory content
- Suggests structural improvements
The system uses LLM-powered analysis to extract search terms:
- The input topic is analyzed for technical terms and concepts
- Domain-specific jargon and method names are identified
- A balanced set of specific and general terms is generated
- Terms are formatted for optimal database query performance
Multiple search strategies maximize relevant results:
Strategy 1: Primary keywords (first 4 extracted terms) Strategy 2: Alternative keyword combinations Strategy 3: Title-derived terms (after structure generation)
Results are deduplicated by normalized title to prevent redundant citations.
Unique keys follow the format: {firstAuthorLastName}{year}{firstTitleWord}
Example: gupta2024model for a 2024 paper by Gupta titled "Model Editing..."
Available citations are provided to the Scholar agent as structured context:
AVAILABLE CITATIONS (use these exact keys):
- \cite{gupta2024model}: Model Editing at Scale leads to... (2024)
- \cite{doan2020a}: A Theoretical Analysis of Catastrophic... (2020)
If a section references an unfound citation key, the Librarian generates a placeholder:
@misc{unknown2024citation,
title = {[Citation needed]},
author = {Unknown},
year = {2024},
note = {Citation key referenced but source not found}
}This ensures compilation succeeds while flagging citations requiring manual verification.
Each section is evaluated across five dimensions:
| Dimension | Weight | Criteria |
|---|---|---|
| Technical Depth | 10 | Substantive analysis, appropriate complexity |
| Clarity | 10 | Logical organization, readable prose |
| Citations | 10 | Relevant references, proper integration |
| Formatting | 10 | Valid LaTeX, consistent style |
| Academic Tone | 10 | Formal language, objective presentation |
- Pass: Total score >= 35 AND no major issues flagged
- Revise: Total score < 35 OR major issues present
Failed sections enter a revision cycle:
- Critic feedback is compiled into specific improvement points
- Scholar rewrites the section addressing each point
- Revised content is re-evaluated
- Process repeats until pass or maximum attempts reached
When enabled, sections are analyzed for:
- Spelling errors
- Grammar violations
- Style inconsistencies
- Punctuation issues
Issues are logged and optionally incorporated into revision feedback.
The repository includes a complete example output in the research_output/ directory:
Topic: Selective Concept Unlearning Without Catastrophic Forgetting in LLMs
Generated Paper Structure:
- Title: "Selective Concept Unlearning in Large Language Models: A Dual-Path Architecture to Prevent Catastrophic Forgetting"
- Sections: Introduction, Related Work, Methodology, Results and Discussion, Conclusion
- Citations: 18 real papers from arXiv (2018-2025)
- Figures: TikZ architecture diagram
- Tables: Quantitative comparison of methods
Citation Examples from Generated Paper:
@article{gupta2024model,
title={Model Editing at Scale leads to Gradual and Catastrophic Forgetting},
author={Akshat Gupta and Anurag Rao and Gopala Anumanchipalli},
journal={arXiv preprint arXiv:2401.07453v4},
year={2024},
url={http://arxiv.org/abs/2401.07453v4}
}
@article{doan2020a,
title={A Theoretical Analysis of Catastrophic Forgetting through the NTK Overlap Matrix},
author={Thang Doan and Mehdi Bennani and Bogdan Mazoure and others},
journal={arXiv preprint arXiv:2010.04003v2},
year={2020},
url={http://arxiv.org/abs/2010.04003v2}
}-
Citation Relevance: While all citations reference real papers, the system cannot guarantee deep semantic relevance to specific claims. Citations are selected based on keyword matching and may require manual verification.
-
Factual Accuracy: Generated content synthesizes patterns from training data and retrieved abstracts. Claims should be verified against primary sources before use.
-
Novel Contributions: The system produces literature synthesis and structured writing but does not conduct original research or generate genuinely novel findings.
-
Domain Coverage: Performance varies by field based on arXiv coverage. Fields with limited preprint culture may yield fewer relevant citations.
-
Compilation Dependencies: PDF generation requires local LaTeX installation. Complex diagrams may require manual adjustment.
-
Model Limitations: Output quality depends on the underlying LLM capabilities. Smaller models may produce less coherent or technically accurate content.
This tool is designed and intended for:
- Research exploration and ideation
- Understanding academic paper structure and conventions
- Generating initial drafts for human refinement
- Educational purposes and writing assistance
- Rapid prototyping of research directions
This tool should not be used to:
- Submit generated papers as original work without disclosure
- Circumvent academic integrity requirements
- Generate papers for fraudulent publication
- Misrepresent AI-generated content as human-authored
All generated content should be thoroughly reviewed, verified against primary sources, and substantially revised before any academic submission. Users bear responsibility for ensuring compliance with relevant academic integrity policies.
Issue: Model not responding or connection refused
Solution: Ensure Ollama is running
$ ollama serve
Verify model availability:
$ ollama list
Issue: No papers found for topic
Possible causes:
- Topic too specific or uses non-standard terminology
- arXiv coverage limited for the domain
- Network connectivity issues
Solutions:
- Use broader, more standard academic terminology
- Check arXiv directly for relevant papers
- Verify network access to api.arxiv.org
Issue: LaTeX compilation fails
Common causes:
- Unescaped special characters in generated content
- Missing LaTeX packages
- Malformed TikZ diagrams
Solutions:
- Review error messages in compilation output
- Check main.tex for obvious syntax errors
- Install missing packages via your TeX distribution
- Disable diagram generation if issues persist
Issue: Grammar tool initialization fails
Solution: The tool downloads ~1GB on first use
$ pip install language-tool-python --upgrade
Alternative: Disable grammar checking
enable_grammar_check: bool = False
Issue: Out of memory during generation
Solutions:
- Use a smaller model (e.g., 7B instead of 70B)
- Reduce max_arxiv_results in configuration
- Close other memory-intensive applications
- Consider using quantized model variants
Issue: Generation stuck or extremely slow
Possible causes:
- Model too large for available hardware
- Network timeouts during API calls
- Infinite revision loops
Solutions:
- Monitor system resources during generation
- Check network connectivity
- Reduce max_rewrites to limit revision cycles
- Use a faster/smaller model
Contributions to AgeSci are welcome. Areas particularly suited for improvement include:
- Additional academic database integrations (IEEE Xplore, PubMed, Google Scholar, ACL Anthology)
- Support for additional document formats (ACM, Springer, Nature, AAAI)
- Enhanced diagram generation with more complex visualizations
- Web-based interface for non-technical users
- Batch processing for multiple topics
- Improved citation relevance scoring using semantic similarity
- Better handling of multi-concept and interdisciplinary topics
- Enhanced grammar and style checking
- More sophisticated revision feedback
- Parallel agent execution for faster generation
- Distributed processing support
- Model-agnostic backend supporting multiple LLM providers
- Improved caching and checkpointing
- Additional usage examples
- Video tutorials
- Troubleshooting guides for specific platforms
Please submit issues for bug reports and feature requests. Pull requests should include appropriate tests and documentation updates.
This project is provided as-is for research and educational purposes. Users must ensure compliance with:
- Terms of service for arXiv API
- Terms of service for Semantic Scholar API
- Academic integrity policies of their institutions
- Applicable laws regarding AI-generated content
If you use AgeSci in your research or projects, please consider citing:
@software{agesci2025,
title = {AgeSci: Autonomous Multi-Agent Research Paper Generation System},
author = {Majumder, Eeman},
year = {2025},
url = {https://github.com/Eeman1113/AgeSci},
note = {Autonomous system for generating IEEE-format academic papers using multi-agent LLM coordination}
}AgeSci builds upon and integrates several open-source projects and public APIs:
- Ollama for local LLM inference
- arXiv API for academic paper access
- Semantic Scholar API for citation metadata
- Rich for terminal formatting
- language-tool-python for grammar checking
For questions, issues, or feature requests, please open an issue on the GitHub repository.