Our research
Powered by the researchers who helped shape how AI learns.
Our team created some of the most widely used tools and methods in machine learning. Now we’re building AI that improves itself at every layer, from data and tools to the models themselves, with every experiment run in production.
Create
Automate the craft of building AI
Can the craft of building AI be learned? We build systems that can optimize themselves, generate their own data and curricula, and create their own tools.
Own
Specialize efficiently, from any data
How little data and compute does it take to master a domain? We work on distillation, lightweight updates and learning from limited data.
Evolve
Learn without end
How can AI keep learning after deployment without forgetting? We work on learning at test time and from rich feedback.
Discover
Reach beyond the known
How does AI find answers no one has found? We work on search, reasoning, and teams of models that propose, test and refine.
Recent Publications
Learning continues after deployment
Rather than prompting a frozen model to search, we let it keep training at test time on the specific problem in front of it. New state of the art across mathematics, GPU-kernel engineering, and algorithm design — evidence that a model specialized on the fly beats a fixed general one.
Learning more from less via self-distillation
Standard RL collapses feedback into a single scalar and discards nearly everything a failing unit test tells you. SDPO learns from the structured signal directly — tests, runtime errors, profiling — which is where the sample efficiency comes from.
Smarter data beats bigger models
Synthetic-data curricula that keep scaling where ordinary distillation plateaus: a 7B model trained this way matched teachers 20× larger. It’s how a specialist gets built when the customer has no data to train it on.
Learning to orchestrate tools, including new ones
ToolOrchestra (arXiv) and AgentFlow (arXiv).
Hard problems need many tools and models, and the right mix shifts as new ones appear. Both systems learn which ones to call and when by practicing on live, multi-step tasks. Small models trained this way outperform much larger ones: ToolOrchestra’s 8B orchestrator beats GPT-5 at 30% of the cost and works with tools it never saw in training.
The harness in a real lab
The same modular, RL-trained agentic system pointed at the gut microbiome: screening nearly 2,000 bacterial genes to identify a DNA-repair axis governing fitness under inflammatory stress, designing a four-strain consortium that attenuated colitis in mice, and generating a commensal-sparing antibiotic cocktail — each experimentally validated. Co-discovery with scientists, not a demo.
Learning continuously from imperfect data
Introspective X Training (arXiv).
Real-world data is mixed in quality, and filtering out the weak parts wastes data that is often scarce. Introspective training learns from all of the data at any stage of training while achieving up to 2.8× compute efficiency. The key idea is to condition on a critique of each piece of data, so the model trains with awareness of its quality.
Want the thinking behind the research? Read our vision, The Plural of Intelligence.