Science
Identify and explore the untested question
Research means reading what’s known, checking it and trying something new. We build environments for agents doing that work and contribute to the benchmarks that measure it.
RL environments for science and research agents
RL environments for science put an agent in the middle of a research problem. It reads what is known, proposes a test, runs it with a tool or a simulator and decides what to try next from a result that may be messy.
Many science benchmarks are well specified questions with one right answer. They measure knowledge. Humanity’s Last Exam, which our researchers contributed to, has 2,500 expert-written questions at the edge of human knowledge, and the best model scored 8% at launch. Research work is different. The question is underspecified, and the most difficult part is choosing what to test.
A research agent also needs to predict how its world will respond. Our world modeling work trains language models to stand in for an environment while an agent trains. GlobeBench tests whether models can predict the effects of an action. The best success rate was 49.6%, and it fell to 5.1% when a change had to spread to dependent objects.
Generative Simulators extend this idea. Tasks, world dynamics and rewards are generated together, so difficulty can rise as the agent improves.
Published research
Humanity’s Last Exam
A multi-modal benchmark at the frontier of human knowledge.
World Modeling for Agentic RL
A language model that stands in for an environment while an agent trains.
Generative Simulators
Environments that generate their own tasks, world dynamics and rewards.
Questions about AI for scientific discovery
What are RL environments for science?
They are environments where an agent takes research steps, such as reading papers, running code or simulations and analyzing results, and earns a reward for the outcome. Unlike question benchmarks, the agent has to act and live with the consequences.
Can AI agents do scientific research today?
Parts of it. Models answer many textbook questions well and still score low on expert-level ones such as Humanity’s Last Exam. Choosing what to test and reading a failed result remain hard, and that is where we focus.
Should science tasks be well specified or realistic?
Both have a place. Well specified tasks are easy to grade. Realistic prompts are closer to how a scientist works. We lean toward realistic prompts and do the extra work to grade them.
Which fields do you cover?
We are exploring engineering and scientific computing, alongside research skills that cross fields, such as reading prior work and designing a test. Tell us the field and the work, and we will say plainly whether we can build it.
How do world models help science agents?
A world model predicts how an environment responds to an action. It lets an agent practice in more situations than anyone could build by hand. Our papers measure where these models still fail, such as changes that ripple through many objects.