AI Agents Tackle Full Research Cycle in New Framework
Scientists often spend years on literature reviews, experiments, and writing reports. A new tool aims to change that by handing much of the work to AI. In a paper presented at EMNLP 2025, researchers led by Samuel Schmidgall introduce Agent Laboratory, a system that uses large language model (LLM) agents to handle the whole research process from a starting idea.
How It Works
Start with a basic research concept from a human, and Agent Laboratory kicks off. It goes through three main steps: scanning existing papers for a literature review, running experiments, and drafting a final report. Along the way, it creates a code repository for the work and lets people step in with feedback at any point.
The team tested it with top LLMs like o1-preview. They also ran a survey where other researchers guided the process with input and then judged the results.
- Versions powered by o1-preview produced the strongest papers overall, according to the evaluators.
- The AI-generated machine learning code matched or beat current top methods in performance.
- Adding human guidance at key stages made the final output better.
- On costs, it cut expenses by 84% next to earlier AI research setups.
Schmidgall and colleagues see this as a way to let researchers focus on big ideas instead of routine tasks, speeding up discoveries in fields like science and tech. You can read the full paper here from the ACL Anthology.
