The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic...
summary
An end-to-end system runs agentic tree search with an experiment progress manager that creates parallel nodes for prototype builds, hyperparameter sweeps, main experiments and ablations, uses VLM feedback on visuals and an LLM evaluator to prioritize nodes, and produces manuscripts via a stronger reasoning model. Three AI-generated papers were submitted and one paper accepted (scores 6, 6, 7; average 6.33, placed ~top 45%); outputs still hallucinate (wrong citations like Goodfellow for LSTM), struggle to generate deep novel hypotheses, and require time‑intensive human verification; code and experiment data are open sourced.
AI-written summary · about
chapters (10)
- 0:00AI Scientist V2 intro
- 2:12Full research pipelines
- 5:29V1 limitations and V2 improvements
- 8:38V2 multi-stage tree search
- 11:30Peer-reviewed AI papers
- 14:03Automated science challenges
- 15:45Responsible AI science
- 17:51AI capability growth evidence
- 18:36Future of AI-driven discovery
- 19:42Open-source V2 and Q&A
related talks
15mWhat’s Next in Medical AI?
28mBuild an AI Research Agent
19mAutonomous Agents for Scientific Tasks
36mBuilding Agentic AI Assistants: Lessons from productionising agents in Preclinical R&D
42mCreating the Engine for Scientific Discovery
16mBuilding Applications with AI Agents
01h 29mBuild Your Own Data Science AI Agents
46mFullstack AI: The Building Blocks for Production Ready AI-Powered Software