An AI Scientist That Actually Got Published

What it does

AI Scientist-v2 is an autonomous research agent that generates hypotheses, runs experiments, analyzes results, and writes scientific manuscripts. It generalizes across machine-learning domains and sheds the human-authored templates that constrained its predecessor. A progressive agentic tree search, managed by an experiment agent, drives the exploration.

The interesting bit

The system produced the first workshop paper written entirely by an AI and accepted through human peer review. It trades the high success rate of template-driven generation for open-ended exploration, so it can wander into novel territory but also fails more often.

Key highlights

• Generated the first fully AI-authored workshop paper accepted after peer review.
• Uses progressive best-first tree search to explore experimental paths in parallel.
• Generalizes across ML domains without requiring human-written code templates.
• Runs on Linux with NVIDIA GPUs and chains multiple LLMs (OpenAI, Gemini, Claude via Bedrock) for ideation, experimentation, writing, and citation.
• A full run costs roughly $20–$25 and takes several hours, with success rates heavily dependent on the foundation model chosen.

Caveats

• Success rates are lower than v1, especially when a strong human template is available; v1 remains the safer bet for well-defined tasks.
• The system executes LLM-generated code with full autonomy, creating risks of dangerous packages, uncontrolled web access, and unintended processes, so sandboxing is effectively mandatory.
• Requires Linux, NVIDIA GPUs, CUDA, and several paid API keys.

Verdict

Researchers curious about automating open-ended ML exploration should look here; if you just need to iterate on a known template, stick with v1. Anyone running it should treat the sandbox as non-optional.