What it is
The Autonomous Researcher transforms a research hypothesis into a complete experiment, designing methodology, writing code, running training, analyzing results, and producing publication-ready papers. All without human intervention.
Origin
Forked from Matt Shumer's open-source autonomous researcher, then heavily modified for local DGX Spark execution. The original was cloud-dependent; this version runs entirely on personal GPU infrastructure.
The Research Loop
- Receive a research question or hypothesis
- Design experimental methodology
- Generate training code (PyTorch, HuggingFace)
- Execute experiments on DGX Spark's 128GB VRAM
- Analyze results with statistical rigor
- Produce Arxiv-style LaTeX papers
- Suggest follow-up experiments
Multi-Model Reasoning
Uses Gemini 3 Pro for broad research context and Claude Opus for precise code generation. The combination leverages each model's strengths: Gemini's knowledge breadth and Claude's implementation accuracy.
DGX Integration
Full access to NGC containers, CUDA 13.0, and 128GB unified VRAM. Experiments that would take days on cloud instances complete in hours. No API costs, no data privacy concerns.
Real Output
This system produced the experimental results for OncoVLM's multi-teacher distillation. The 92.4% PubMedQA accuracy came from experiments designed and executed autonomously.
The Meta-Story
AI designing AI experiments. The researcher becomes the researched. Each successful run teaches us more about what autonomous systems can accomplish.
Connected work