SakanaAI/AI-Scientist-v2

AI-Scientist-v2 is the first system to get a fully AI-generated scientific paper accepted at a peer-reviewed ML workshop with zero human intervention. Not a writing assistant. Not a research helper. An end-to-end agent that generates hypotheses, designs experiments, runs them on GPUs, analyzes results, writes manuscripts with citations, and iterates based on reviewer feedback — all autonomously. The core innovation in v2 is progressive agentic tree search. Instead of the linear pipeline in v1 (which required human-authored templates for each research domain), v2 uses best-first tree search (BFTS) with an experiment manager agent that explores multiple research directions in parallel. Think of it like a chess engine, but for science — it evaluates which experimental paths are most promising, allocates compute to the best branches, and prunes dead ends automatically. A Vision-Language Model feedback loop handles something researchers usually do manually: looking at figures and deciding if they actually communicate the findings. The VLM iterates on chart aesthetics, label placement, and visual clarity until the figures meet publication standards. That's the kind of detail that separates this from a GPT wrapper that generates papers. The system runs experiments using real GPU compute with PyTorch and CUDA. It integrates Semantic Scholar for literature search, supports GPT, Gemini, and Claude (via AWS Bedrock) as the reasoning backbone, and costs approximately $15-$20 per experimental run. That cost-per-experiment figure is remarkable — a human researcher's time on a single experiment often exceeds that by an order of magnitude. Sakana AI, the Tokyo-based company behind the project, explicitly acknowledges v2's tradeoffs. It doesn't necessarily produce better papers than v1 when a strong template exists. v1 excels at well-defined tasks within narrow domains. v2 trades that reliability for generalization — it can tackle open-ended research questions across ML domains without templates. Lower success rate, but dramatically broader capability. The security warning is worth taking seriously: the system executes LLM-generated code on your hardware. Sakana strongly recommends Docker containers for sandboxing. Any paper generated must include a disclosure that it was created by AI — the repo provides mandatory wording for this.