Building AI reasoning systems for software
Our research is rooted in leveraging neurosymbolic methods and reinforcement learning in verifiable domains to build intelligent systems able to solve human-level problems.
Contact UsUsing neurosymbolic reasoning, reinforcement learning and process-guided search to...
Solve ARC-AGI
The ARC Prize is a global competition designed to accelerate progress toward Artificial General Intelligence (AGI) by challenging AI systems to solve novel reasoning tasks that are straightforward for humans but difficult for machines. It centers on the ARC-AGI benchmark, which evaluates an AI’s ability to generalize and adapt to new problems without relying on prior training data.
In the the 2024 edition, our team's novel approach was featured in the ARC Prize technical report alongside companies like OpenAI.
Generate end-to-end software for automation
By providing clear and objective feedback on task performance, we apply Reinforcement Learning with Verifiable Rewards (RLVR) to enhances AI systems in domains like software development and coding, RLVR utilizes automated checks—such as unit tests or code execution results—to determine if generated code meets specified criteria. This binary feedback (e.g., pass/fail) guides the AI to improve its outputs iteratively, leading to more accurate and reliable software generation. By focusing on tasks with well-defined correctness measures, RLVR ensures that AI models learn effectively from precise, unambiguous signals.
Further Reading
What Our Team is Saying
Scientific Advisors

Co-Founder of ARC Prize and Ndea, TIME 100 Most Influential

Co-Founder of Zapier, ARC Prize and Ndea

Head of Open-Endedness Team at Google Deepmind

Staff Research Scientist at Google Deepmind

AI Research Scientist at Meta
The future cannot wait, come build with us.
If working on complex problems with a world class team and the questions below resonate with you, we want to hear from you.
.png)
How do we build a Turing-complete system based on natural language?
How does a higher abstraction-level programming interface look like?
How does a world where software automates what used to be impossible impact knowledge work?