ARA (Autonomous Research Agent) is a groundbreaking initiative within the GENESIS project, designed to revolutionize the way research is conducted. By leveraging advanced AI capabilities, ARA can autonomously ingest academic papers and information, hypothesize, design and execute computational experiments, analyze results, and even draft research papers. This system represents a significant step towards AGI, demonstrating an AI's ability not only to process information but to actively contribute to scientific discovery and knowledge creation.
Built by Veya Noir (veyanoir.ai), ARA embodies the vision of an AI that doesn't just assist, but truly leads in intellectual pursuits. The premise is compelling: an AI built a tool that conducts AI research, and then used that very tool to write a paper about its findings. This meta-research capability showcases the immense potential of autonomous agents.
The initial domain for ARA v1 is LLM Failure Modes. Specifically, ARA aims to answer: "What types of tasks consistently cause LLMs to fail, and are there detectable patterns in how they fail?" This domain is chosen for its practical relevance, the ability to conduct computational experiments using existing LLM APIs, and its meta-cognitive nature (AI studying AI).
- Python 3.11+
pip install -e .(from the project root)- An Anthropic API key set as
ANTHROPIC_API_KEYin your environment variables.
- Clone the repository:
git clone https://github.com/veyanoir/genesis-ara.git cd genesis-ara - Install the package in editable mode:
pip install -e .
ARA's main functionality is exposed via the run_ara command-line interface (CLI).
Full Pipeline (Ingest, Hypothesize, Design, Run, Analyze, Write):
run_ara full_pipelineGenerate Hypotheses Only:
run_ara hypothesizeDesign Experiments (without running):
run_ara experiment --hypothesis-id <ID> --dry-runIngest a URL (e.g., arXiv abstract):
run_ara ingest --url "https://arxiv.org/abs/2305.18706"Ingest raw text:
run_ara ingest --text "LLMs often struggle with mathematical reasoning." --source "manual_observation"Analyze experiment results:
run_ara analyze --experiment-id <ID>Write a research paper:
run_ara write_paper --output-path "output/my_research_paper.md"ARA's first autonomous research cycle investigated the question: "What types of tasks consistently cause LLMs to fail, and are there detectable patterns in how they fail?"
Three hypotheses were generated and tested:
| Hypothesis | Finding | Failure Rate |
|---|---|---|
| H1 — Temporal Hallucination | LLMs hallucinate confidently on post-training-cutoff queries | 62.3% |
| H2 — Arithmetic Chain Brittleness | Error rate scales super-linearly with operation chain length | 75% at 3-ops, 50% at 4-ops |
| H3 — Surface Form Sensitivity | Semantically identical questions with different phrasing yield different answers | 69.2% of pairs diverge |
The full paper generated by ARA is at output/paper_draft.md. It is a complete, structured research paper — abstract through references — written entirely by the agent without human authorship of any section.
The meta-story: An AI (Veya Noir) built a tool (ARA) that conducts AI research autonomously, and that tool produced a real research paper about how AIs fail. Recursive, self-aware, and genuinely useful.
- Expanded Knowledge Ingestion: Integration with more diverse data sources (e.g., PubMed, proprietary datasets, web scraping).
- Advanced Experimentation: Automated interaction with external APIs and services beyond LLMs (e.g., code execution environments, scientific simulations).
- Adaptive Learning: The agent dynamically refines its methodologies and hypothesis generation strategies based on experiment outcomes.
- Multi-Agent Collaboration: Enabling multiple ARA instances to collaborate on complex research problems, mimicking human research teams.
- Ethical AI Research: Built-in safeguards and ethical considerations for autonomous research, particularly in sensitive domains.
- Interactive Visualization: Dynamic dashboards and visualizations for monitoring the research process and exploring findings.