Definition
Plain language
A simulated science-lab environment for testing whether AI agents can carry out experimental procedures.
As stated in the literature
A text-based interactive environment for evaluating LLM agents on multi-step science tasks like growing plants or measuring temperatures, built on a PDDL-style simulator.
Also called: SciWorld
Why it matters: Multi-step procedural tasks expose failure modes that single-question benchmarks miss, like agents that get partway through a procedure and then forget what they were doing.
For example, the agent might be told to grow a plant, and has to find seeds, fill a pot with soil, add water, and place it near a light source — each as a discrete action in the simulator.
Heard on the show
“On the science-experiment environment, SciWorld, when they scaled their method up to twenty-five parallel trajectories, it outperformed fully fine-tuning the same model with GRPO.”Episode 119 — Beating Reinforcement Learning Without Ever Touching the Model's Weights