Solving Hard Problems Requires Efficient Search. Neuroscience Teaches Us How.
Reinforcement learning is, at its core, amortized and memorized search. A way of compressing the results of exhaustive trial-and-error into policy parameters, so that at inference time, the right action can be retrieved without re-running the search. It works. But it has a fundamental ceiling: the quality of the compressed policy is bounded by the quality of the search that produced it.
Hard problems, genuine reasoning, novel planning, open-ended generalization, cannot be fully amortized. They require search at test time, not just at training time. This is why scaling training alone is not enough. But naively scaling test-time compute is not enough either. What is needed is carefully designed inference: systems that know when to search, where to search, and how to do so efficiently.
The brain solved this problem long ago. Biological cognition doesn't brute-force search over exponential spaces; it navigates them through hierarchical abstraction, dynamic gating, and selective pruning, performing structured, resource-constrained search on 20 watts. What makes this possible is an underlying capacity for adaptation: assessing difficulty on the fly, allocating compute where uncertainty is high, scaling back where it is not. This is what makes biological search both efficient and generalizable. Better generalization, the ability to transfer competence to new domains and remain useful in the open-ended complexity of the real world, is not, in our view, a separate property to be optimized for. It is a natural consequence of a system that can adapt its own inference.
At Voaige, we study these computational principles. Our goal is to understand how the brain conducts search, what algorithmic strategies it employs, and where the real gains come from, then implement principled approximations of those mechanisms using the representational machinery of large language models.
This is what we call Test Time Cognition: search-capable inference that is architecturally grounded in neuroscience, not just scaled compute.