State on Demand: Could We Build Cognitive Modes?

Almost everyone who has done hard intellectual work has noticed that the work goes differently at different temperatures of the mind. A proof that stalls at 2 a.m. can fall open after a walk; a difficult conversation that would have gone badly before lunch goes well after it; a composer who cannot find a melody while anxious finds three while half-distracted. The folk vocabulary — “in the zone,” “sharp,” “foggy,” “wired,” “flat” — implies a question the series keeps circling: could the differences between those conditions be moved from luck into deliberate control?

The version of that question here is narrower and more testable than “can we control mood.” It asks whether a person might deliberately enter and leave a small set of recurring modes — analytical, creative, learning, social, and recovery. A mode, in this sense, is not a personality and not a skill. It is a configuration of several variables at once: arousal, the breadth of attention, the readiness to inhibit a habitual response, sensitivity to error, the balance between exploiting what you know and exploring what you do not, and the willingness to feel reward. Because careful sequential reasoning and loose association reward different configurations, a single permanent “optimal state” is an incoherent target. The question is whether the right configuration can be reached on purpose, and what would have to be true for that to be more than a metaphor.

What is already being switched, and by what

Nobody regulates cognition from nothing. Ordinary life is full of crude mode control, and the crude versions are informative because they show which levers actually move the underlying variables.

A deadline narrows attention and raises arousal. A specific location — the same desk, the same carrel — becomes a cue that a particular kind of work happens here. Music sets tempo and mood, sometimes better than intention does. Coffee shifts arousal upward; nicotine shifts attention and alertness through separate pathways. A ritualized opening of a notebook, or a fixed block of deep-work time, borrows the machinery of habit to reduce the startup cost of a demanding task. Social context changes what a person will attempt: an audience can raise performance on well-learned tasks and degrade it on novel ones.

These levers work unevenly, and for reasons that map onto the variables above rather than onto any single global “focus.” The coffee that helps a person plow through email can make them jittery and worse at reading a technical paper. What modern AI adds is three capabilities older techniques lacked: personalization, continuous measurement, and adaptation as the person’s response changes. Whether those can be assembled into something that reliably and safely switches modes is the engineering question; whether they should be is the ethical one, because the same loop that helps a person enter a mode can be aimed at an objective that is not the person’s own.

The variables that actually shift between modes

Before any switching is plausible, it helps to be precise about what is being switched. The literature offers at least four axes, and they do not move as a single dial.

Arousal. The oldest finding in this area is that performance on many tasks is best at intermediate arousal and worse at both extremes — the inverted-U. The 1908 experiments that gave the curve its name were tiny, a few animals per condition, and a 2024 reassessment in Trends in Cognitive Sciences argues that the curve’s famous precision was never justified by the original data (Nieuwenhuis, 2024). That does not make the shape false. A 2024 study in PNAS recorded roughly 3,570 trials per participant across 28 people and found a genuine inverted-U between pupil-indexed arousal and perceptual sensitivity, with a circuit model — disinhibition via SST and VIP interneurons — proposed as the mechanism (Beerendonk et al., 2024). Arousal is real, measurable through indirect proxies like pupil size, and non-monotonic: too little is as bad as too much.

Aston-Jones and Cohen argued that the locus coeruleus–norepinephrine system toggles between a phasic mode that sharpens the current task and favors exploiting a known good option and a tonic mode that broadens responsiveness and favors exploring alternatives (Aston-Jones & Cohen, 2005). Exploit-versus-explore is one of the axes that separates analytical work from creative work. Arousal is not simply “energy”; it is a bias about which kind of move to make next.

Inhibitory and attentional breadth. The second axis concerns how narrowly attention is focused and how strongly habitual responses are suppressed. Monsell’s review of task switching documented that changing tasks carries a measurable cost, on the order of 200 milliseconds, that preparation reduces but never eliminates, leaving a “residual” cost that looks like the previous task set bleeding into the new one (Monsell, 2003). That residual is the price of having been in a different mode; anyone who has tried to move from editing prose to debugging code knows the bill.

Tobias Egner’s 2023 synthesis adds a crucial refinement: the ability to hold a task focus and the ability to switch flexibly are two largely independent dimensions, not two ends of one spectrum (Egner, 2023). People can be good at stability and poor at flexibility, or the reverse, and they lean on cheap heuristics — recency of the last task, recognition of familiar cues — to decide when to switch. The regulation of these demands adapts to changing conditions while being dysregulated in conditions including ADHD, autism, schizophrenia, and Parkinson’s disease. Switching is neither free nor uniform, and a person’s switchability is a separate trait from their focus.

Creative association. A third axis is the one creative work depends on: cooperation between brain networks usually described as opponents. Beaty and colleagues reviewed evidence that creative production involves joint engagement of the default network, which supports spontaneous associative thought, and the executive control networks that manage and evaluate it (Beaty et al., 2016). The pattern is not “turn off control to be creative”; it is a specific temporal coordination, generate broadly and then evaluate sharply. That is a different signature from sustained analytical attention, and it helps explain why a state excellent for proofreading is often poor for brainstorming.

A reward and error-monitoring axis. The fourth axis is motivational. Whether a person is oriented toward seeking reward, avoiding error, or monitoring for surprise changes what they notice and how quickly they update. The locus coeruleus appears again, in its role in signaling unexpected events and reorienting behavior, and the system is less unified than early models assumed: a 2020 review describes it as heterogeneous, with distinct subpopulations of its roughly 3,000 rodent neurons apparently encoding different cognitive processes rather than all broadcasting one signal (Poe et al., 2020). That heterogeneity matters for any claim that a single internal dial could be turned to a “creative setting.”

Why “mode” is a useful abstraction and a dangerous one

The mode metaphor compresses a messy multidimensional state into a label a person or a system can act on. The danger is the same compression: a label invites treating the state as a thing that is either present or absent.

Greene and colleagues proposed three criteria for calling something a brain state: it should recur, it should be stable for a behaviorally significant period, and it should be distinguishable from other states (Greene et al., 2023). Sleep stages qualify cleanly. Waking cognitive modes qualify less cleanly, because they overlap, change on the timescale of seconds to minutes, and do not map one-to-one onto any measurement; a person can be in a broad-attention mode and an error-sensitive mode at once. These modes are more like recurring regions of a landscape than discrete rooms.

There is a second reason for caution. States are not independent of the tasks performed in them, and the causal arrow runs both ways. A 2024 study in Nature used precision functional mapping — roughly eighteen MRI visits per participant — to track what a 25 mg dose of psilocybin does to functional connectivity, with 40 mg of methylphenidate as an active comparison (Siegel et al., 2024). The psilocybin condition produced more than three times the connectivity change, most of all in the default-mode network, with a persistent decrease in hippocampus–default-network coupling lasting weeks. A pharmacological intervention can move a person into a connectivity configuration they would not otherwise occupy — but the tool is blunt: the drug offers no menu of modes, and the finding concerns the acute and subacute effects of a single high dose rather than a controllable switch.

The nearest real example of a mode that can be entered on demand

The closest laboratory demonstration that behavioral modes can be entered and left deliberately comes from animals. Tervo and colleagues trained rats to alternate between a strategic choice policy and a stochastic, exploratory one (Tervo et al., 2014). The switch depended on noradrenergic input from the locus coeruleus to the anterior cingulate cortex; suppressing that input left the animals stuck in one policy. The mode was real, reproducible, and tied to a specific circuit, but the animal did not decide to switch the way a person decides to buckle down; it was gated by an internal signal the experimenter could manipulate. The study proves that strategic versus exploratory choice is a genuine, switchable state in a mammalian brain. It does not prove that a human can select such a state at will.

Learning and recovery as modes with their own clocks

Two of the proposed modes are special because they are dominated by processes that run on time scales longer than a work session.

Learning has an encoding phase that favors alert, attentive wakefulness and a consolidation phase that favors sleep. The canonical review of this division describes the waking brain as optimized for encoding new information and the sleeping brain as optimized for consolidating and integrating it, with slow-wave sleep doing much of the systems-level work of redistributing memory traces into longer-term storage (Rasch & Born, 2013). Part of “learning mode” is therefore not entered during the study session at all. Consolidation happens later, and a person who trades sleep for extra study time may be trading consolidation for encoding, often a bad exchange. A system that helped someone enter an attentive encoding mode while protecting the sleep that consolidation requires would be doing something genuinely useful; one that optimized only for time-on-task would quietly degrade the thing it claims to improve.

Recovery is the other long-clock mode, and the one most easily corrupted. It is not the absence of activity. It involves processes whose markers a device can see but whose function it cannot infer — heart-rate variability, slow-wave sleep, the return of baseline reactivity — and it depends on deprivation, which is not the same as preference. Here the vocabulary of modes becomes dangerous, because a system optimized for subjective comfort or short-term productivity can label a state “recovered” while the underlying deficit persists.

The state-space view, and where optimization enters

Put the axes together and a different picture becomes available, one drawn from 2025 work on arousal as an embedding. Raut and colleagues showed, in mice, that a single scalar measure of arousal derived from pupil diameter could reconstruct a large fraction of multidimensional spatiotemporal brain dynamics through a delay-embedding construction (Raut et al., 2025). The claim is not that arousal is the only variable; it is that a surprisingly low-dimensional latent variable organizes a high-dimensional state space, so the manifold of states a brain occupies is smaller and more orderly than raw measurements suggest.

If that picture generalizes even partially to humans, the mode question becomes a search problem. Modes are regions of a state space, and the space is not infinite, because physiology constrains it. Some regions are easy to reach through known levers — sleep, exercise, social presence, drugs, attentional focus. Others may be reachable in principle and rarely visited in practice. The engineering temptation is to treat this as optimization: define a target function for “creative” or “focused,” search the space of interventions, and steer the person into the highest-scoring region.

That is where this article hands off to its neighbor. Whether a search could find genuinely novel and useful configurations, and what it would mean to accept a state a person has never occupied, is harder than whether familiar modes can be entered on purpose: it is a search over a space whose objective function has not been agreed on and whose side effects may take months to appear.

What “deliberate switching” would actually require

Take the optimistic reading and ask what a person or a system would need in order to switch modes reliably. Four requirements are visible in the evidence above.

The first is a measurement of the current state faster than the state’s own drift. Pupil diameter, reaction time, and task performance all qualify; a monthly self-report does not. The second is a model of how this particular person moves between regions — which levers work, how strongly, and with what delay. The delay matters: Monsell’s residual switch cost shows that the state left behind persists, so any intervention has to account for inertia rather than assume an instant transition.

The third is a target. The person has to specify what state is wanted, for what task, and for how long. This step looks like a user-interface problem and is actually the whole ethical game. A tool that accepts a target from the user is an instrument; a tool that infers the target from an engagement metric or an employer’s productivity dashboard is a management system wearing an instrument’s clothes.

The fourth is a reversibility guarantee. A deliberate switch should be inspectable while it is happening and reversible after it ends. This is where the thermostat analogy fails instructively: the furnace and the room are separate objects, and the thermostat’s state does not alter the setpoint. A person is both furnace and room, and a change deep enough to alter which mode is available may also alter what the person wants the mode for. The sharpest clinical statement of this problem comes from work on state-dependent stimulation: the same electrical stimulation at the same site produced a calming effect when a patient was anxious and a worsening effect when the same patient was low on energy (Scangos et al., 2021). The sign of the effect depended on the state it was applied to, so a switching system that ignores its own starting state will sometimes push in the wrong direction.

The case for the useful version

The strongest version of this idea is modest and durable: the deliberate, learned management of the variables above — arousal, attention breadth, the stability-flexibility balance, error sensitivity, reward orientation — using well-understood means, with measurement to close the loop and personalization to match the response to the individual.

Even at that modest level the gains are real. Starting-cost is a measurable tax, and a reliable transition routine pays it down. Context carryover is another, and knowing how long a residual switch cost lasts lets a person schedule a buffer instead of blaming themselves for the fog. Recognizing that creative work calls for a generation-then-evaluation rhythm rather than one long “focus” session keeps a person from spending their best hours in the wrong configuration.

The gains also compound, because a technique that works only because a device steers it leaves nothing behind when the device is gone, while one the person understands and can run by hand is a capability rather than a dependency.

How mode control goes wrong, in order of how quietly it arrives

The loud failure of mode control is coercion, and it is not the most likely one. The quiet ones are more interesting.

The first is narrowing the definition of good work. If the mode system is tuned on a proxy — words written, lines committed, tickets closed — it will optimize the proxy. Creative and learning modes are especially vulnerable, because their value is partly latent: a slow afternoon of apparently unproductive associative wandering is sometimes the state from which next month’s insight emerges. A system that cannot represent delayed value optimizes the observable part and erodes the rest.

The second is flattening individual differences. The population distributions of optimal arousal differ between people, and the same person’s optimum differs between tasks and times. A system trained on group averages will push individuals toward a mediocre central tendency that feels smooth and works poorly. Personalization is not a feature; it is the minimum requirement for the thing to be worth using.

The third is capture by the surrounding institution. The same measurement and inference capabilities that let a person shape their own cognitive modes let an employer, platform, or school shape the modes of everyone it has leverage over. Consent is necessary here but not sufficient, because a worker who clicks “agree” may be choosing between a mode system and a job. The relevant question is whether a person can decline without penalty and leave without loss.

The fourth is the substitution of state for skill. A mode that reliably produces fluency can mask the absence of the underlying ability, and the fluency may not survive removal of the scaffolding. This is the boundary between assistance and control described in the series hub: a system assists when it widens a person’s range of voluntary action and controls when it narrows that range, even if the narrowing feels pleasant. A mode system that leaves the person more able to enter the mode unaided is assisting; one that leaves them unable to work without it has substituted itself for a capability.

How this would be judged

There is a temptation, when a topic touches subjective experience, to accept a person’s report that a state was good. That is not enough here, for the same reason it is not enough in medicine: the states being altered are partly the states that produce the report. The better questions are diagnostic. Can the person enter the mode again next week without the intervention, or with a reduced one? Does the benefit transfer to a new task and a new environment, or is it bound to the exact setup where it was trained? Do the objective measures move in the same direction as the felt experience, or do they diverge? Does the system leave the person’s own reflective priorities intact, so that the goals it serves are still recognizably theirs?

Applied over months rather than minutes, those questions separate a capability from a dependency.

Where the question lands

The five modes in the opening question are not five buttons. Arousal, attention breadth, the stability-flexibility balance, error sensitivity, and reward orientation can each be nudged by ordinary means, and the evidence that these nudges change performance is real and often precise: the inverted-U is measurable, the switch cost is measurable, the exploit-explore tradeoff has a circuit, and the sleep-consolidation division has decades of work behind it. What does not yet exist, outside a few clinical settings, is a loop that measures a person’s state continuously, personalizes a model of how it moves, and steers reliably and safely among the modes. That loop is plausible engineering, not a demonstrated consumer capability, and no device on the market today should be described as switching a person’s cognitive modes.

The more durable point is about the shape of the problem. Cognitive modes are a genuine feature of how minds work, and the case for treating them as switchable-to-a-degree is better than the case for treating them as fixed. The case for treating them as fully plastic, and for delegating the choice of mode to an optimizing system, is weaker than it looks, because the system’s objective is not given by the biology. The brain supplies the axes. Something else supplies the target, and that something should remain the person whose brain it is.