We pursue two connected lines of research to build state-of-the-art,
commercially deployable systems.
Our hypothesis
A research program connecting learning, episodic memory, and explicit search to build more powerful adaptive intelligence.
Read Intelligence Is SearchTrack 1
The Bitter Lesson has two halves: learning and search. The pre-training and post-training eras scaled learning; we believe the next frontier will arise from cognition, which treats search as a first-class citizen of intelligent systems.
We hold model weights and prompts fixed, and explore everything else. Our research discovers levers within LLMs that improve reasoning at test time, guided by computational principles drawn from neuroscience.
AI today encodes intelligence into weights and retrieves it at inference. Hard problems demand more: active reasoning that balances accuracy with efficiency, and knows when to think harder.
Track 2
We identify a key principle for identifying where search and planning should occur in robotics, instantiate it with per-skill policies and world models composed at test time, and show novel compositions, inferred subgoals, and replanning without composite demonstrations.
Perturbing a small fraction of routing decisions in frozen MoE reasoning models—a little noise helped, a lot hurt, and the tuning curve suggests a new search surface inside the model.
Dynamic Compute Allocation improves agent performance by deciding, at each turn, where additional computation is worth spending.