The Question I
Took With Me

The question after the talk

In late June I went to Deep Tech Week in San Francisco, where Dipan, Voaige's founder, gave a talk on Test-Time Cognition, the company's approach to making a model's search for an answer both more accurate and cheaper at the moment it runs.

It was a late afternoon session, and I had been thinking about going home after the one before it. I sat a few rows from the front. Most people were sitting farther back with their laptops open. They looked like engineers, as far as I could tell. The questions from the floor were about benchmarks and how they had got their numbers. There was real excitement in the room for the results. I was excited by them too, especially because of the neuroscience that had inspired the approach. Nobody asked about that part.

Voaige's bet is that intelligence can usefully be treated as search. Training buys a model a very good first guess, but harder problems require some of the search to happen when the problem arrives. Spending more computation at run time helps, and often multiplies the cost of each problem. Voaige calls its approach to that tradeoff: Test-Time Cognition. What mattered to me in the room was what neuroscience was doing in it.

Companies often invoke the brain when they talk about AI, and the invocation does not always survive a follow-up question. So I asked two.

The first was why the accuracy gains on Dipan's per-task slide were larger on the medium problems than on the hard ones. His answer was that, on the hardest problems, the model was operating furthest from what it had seen during training. It therefore had a poorer sense of which step was likely to come next and when to stop. Allocating more inference-time compute had less leverage there: it could keep an unproductive trajectory going instead of bringing the model closer to an answer.

The second was what level he took the neuroscience to operate at in LLMs: gross function, area, cortical column, circuit. My own bias ran a layer above columns, since that is where most of my work sat. His view was that a language model is not one microcolumn but a group of them, and that the allocation layer works between those groups.

The comparison is functional, not anatomical. Biological plausibility is a guide here, not a requirement. It put the question at a level where we could ask what should be allocated, what should do the allocating, and which results from neuroscience might bear on either. At the time it changed how I read the results he had just shown and what I would have suggested trying next. He has since convinced me that the comparison is useful for language models.

I kept asking after the talk ended. The answer about the hard problems was the one I kept coming back to.

Anyone who runs these systems has seen models collapse into loops and spirals, repeating themselves, with or without any particular inference scaffolding.

That word is starting to appear in the AI literature. Behind it sits forty years of systems neuroscience on how circuits settle into stable patterns and how they get pushed out again. The work runs from the thirty neurons of the crustacean stomatogastric ganglion, where neuromodulators reconfigure one circuit into several different rhythms (Marder & Bucher, 2007), to whole-brain models built for individual epilepsy patients and taken into a surgical trial (Jirsa et al., 2023). In the epilepsy work the question was never only descriptive. Somebody has to get a person out of the state.

None of that solves the AI problem. It gives you measurements, candidate mechanisms and results worth checking.

I thanked Dipan afterward and told him something I had believed for a while: that neuroscience might help the language-model field move faster by narrowing the space of things worth trying. The new part was hearing someone say it from a stage and treat it as an engineering plan.

So I went home and read. The next day I sent Dipan a brief on how neural circuits escape steady states, and on machine-learning work that had already borrowed from that literature without identifying the connection. Voaige had already spent years building a research program around the premise that neuroscience could yield useful computational principles for AI. The brief gave us a concrete point of contact within that larger program: a specific literature connected directly to a live problem. We had arrived at the same underlying conviction from different directions, and it showed me where my own background could add leverage. The conversation continued, and I joined about three weeks later.

The opportunities in composing

Joining gave me the chance to keep finding and composing those connections inside a live research program. In my first two months, I have seen how an adjacent result in neuroscience is turned into a question that can survive the constraints of an AI system.

I spent about a decade in neuroscience taking biological systems apart: most of it in mice, with shorter stretches in flies, rats, monkeys and people. In practice that meant years in dark rooms watching cells light up as a mouse ran a belt and recognized where the reward was. It also meant twenty or thirty minutes with each animal beforehand, letting it run over my hands and arms until it was comfortable enough with me to do the experiments together. Over hundreds of mice, that adds up.

The most direct transfer is the work I had already done as a translator between people who ran experiments and people who modeled them. On my first paper, on the dorsal and ventral hippocampus and how saliency drives the ventral preferentially, my central role was carrying a question back and forth between two groups who described it in incompatible vocabularies until it came out as something both could act on.

That practice is useful here too, with different vocabularies. One habit from it comes up almost daily: suspicion of functional labels.

Prior work using optogenetic manipulations, recordings and activity-dependent labeling had suggested that the ventral hippocampus contains partly separate populations associated with positive and negative experiences. That work helped motivate the hypothesis that vCA1 represents valence, roughly whether something is good or bad. We used two-photon calcium imaging across six experimental paradigms to test that hypothesis. We found robust representations of stimulus identity, sensory modality and intensity, but not a general positive-versus-negative valence code (Biane et al., 2025). Showing that a region helps control valenced behavior is not the same as showing that its neurons represent valence.

The same habit applies to a model that repeatedly returns to the same line of reasoning. Calling it a loop identifies the behavior, but not the mechanism. The repetition could come from candidate generation producing too little diversity, evaluation failing to recognize that the search is no longer improving, or the control process continuing to spend compute after the useful options have been exhausted.

Those explanations imply different experiments. To test the first, vary candidate diversity while holding evaluation and stopping behavior fixed. To test the second, keep generation fixed and change how progress is measured. To test the third, vary the stopping threshold and ask whether the system exits earlier without losing accuracy. A useful result is a pattern across these conditions that rules out some explanations and clarifies the next decision.

Work on attractor dynamics, uncertainty, and transitions between stable states can suggest hypotheses and measurements for this kind of failure. Those ideas become useful here when they yield predictions that can be tested in the model, under the constraints of the system actually running.

What I have seen since joining is that this work does not follow a fixed handoff between disciplines. A proposal has to meet what is already known about the model, the measurements the system can support, the compute available, and the alternatives the result needs to distinguish. The team says no to far more ideas than it executes because an interesting idea alone does not justify the cost.

Having a neuroscience philosophy is cheap. Building around one is not: it determines which questions earn experiments, which infrastructure gets built, and which results count as evidence. It also means setting aside plausible directions that will not resolve an important uncertainty. That is the difference I can now see from inside between invoking neuroscience and building with it.

From separate results to one system

Much of what Dipan presented that afternoon was Dynamic Compute Allocation, one published example of this approach. I have learned much more about it since joining. DCA sits above a pool of models and decides, for each problem, how much computation that problem is worth and which models to trust with it. On Terminal-Bench 2.0, its published results include configurations that outperform stronger single-model baselines at lower cost. The benchmark, model configurations and cost accounting are in the technical results.

Three questions run underneath DCA. Which model should be trusted on the problem in front of it? How should the system combine what several models produce, or decide to take one answer and drop the rest? How much computation should it spend before stopping?

Those decisions occupy part of a broader loop. A system also has to infer what kind of situation it is in, judge how far to trust an answer, act, and update its estimates from what happens next. Each question appears in a different neuroscience literature. Their results give us candidate operations and quantities rather than a finished design. Set beside one another, they suggest a loop we can build and test.

The hard constraint is compute. Carnot began with steam engines that already worked and asked how much work they could extract from a given quantity of heat, even though the theory of heat was wrong (Réflexions sur la puissance motrice du feu). We begin with models that already work and ask how much useful problem-solving ability can be obtained from a given quantity of compute.

Answering that question does not require us to build a brain or adopt a unified model of decision-making. We are looking for algorithms that remain useful when separated from their biological substrate.

Compression, distillation and inference engineering all contribute. A two-order-of-magnitude efficiency target also requires the system to decide where to spend computation and what information to retain.

Neuroscience does not hand you a blueprint

Access to every component of a system does not guarantee that an analysis will recover how it computes. Jonas and Kording ran common neuroscience analyses on a simulated 6502 processor whose workings were completely known. The analyses recovered interesting structure without recovering the hierarchy by which the processor computes. Knocking out individual transistors produced transistors apparently “for” Donkey Kong, a result the authors call grossly misleading (Jonas & Kording, 2017).

The paper shows how easily familiar analytic tools can support the wrong functional story, even in a system simpler and more accessible than a brain. We take specific operations we might build, such as weighting estimates by their reliability or stopping once evidence crosses a threshold, not a wiring diagram or a functional label.

The other problem is organization. Senden clustered 461,316 neuroscience papers published between 1999 and 2023 and found no research cluster devoted to a theoretical framework (Senden, 2026). Frameworks may be embedded within domains, and the same analysis finds substantial citation between clusters. Even so, the field is organized around mechanisms, diseases, functions and methods, not around the question in front of us: which results from separate literatures will work together in one system.

The literature narrows the candidate operations and gives us reasons to test particular combinations. We build each component separately, then test it alone and with the others. The experiments tell us what each one adds, what changes when they interact, and which ideas survive.

The program advances by identifying principles that generalize across problem types and compose into stronger systems. Each candidate has to earn its place empirically: it should improve the system on its own, retain value across tasks, and contribute when combined with other components. That standard keeps the program scientific while directing effort toward an integrated architecture.

The question organizing the work

The scientific question underneath the work is:

Voaige approaches this through a working decomposition of search: how candidates are constructed, how they are evaluated, how memory changes the search space, and how the system decides where to spend compute and when to stop. The program is now refining the boundaries and interactions among these components, specifying each operation precisely enough to implement, using neuroscience to narrow its design, and testing it in the current substrate: language models.

What interests me is identifying the level at which each principle transfers. Some can be implemented close to their computational form; others become useful only after being translated into the constraints of a language model. The work is to make that translation explicit, test it, and assemble the principles that hold into a coherent system.

Published results across several base models already show that the approach is not tied to one model. They include configurations that improve performance without a large increase in cost and configurations that lower average cost, providing a concrete foundation for the broader program.