Parallel Minds: How the Brain Weighs Choice and Confidence at the Same Time

This post dives into new emerging research with novel insights from Vivar-Lazo and Fetsch, who asked a deceptively simple question: when we make a decision, does the brain first decide what to do and then figure out how confident it is, or does it compute both in parallel? Using carefully trained monkeys and single-neuron recordings in the lateral intraparietal cortex (LIP), they show that choice and confidence are supported by concurrent evidence accumulation, rather than a serial, two-stage process.

Why does this matter for people working with biofeedback and neurofeedback? Both approaches are about closing the loop between physiology and awareness: sensors track bodily or brain activity in real time and feed it back so that a person can learn to self-regulate. In very general terms, biofeedback uses measures like heart rate, breathing, or muscle tension, while neurofeedback typically uses brain-based signals such as EEG rhythms or, in research settings, fMRI patterns. In both cases, learning hinges on the brain’s ability to monitor its own performance and to assign confidence to internal signals.

The study by Vivar-Lazo and Fetsch shows that confidence is not some after-the-fact commentary layered on top of a decision. Instead, it seems to be woven directly into the decision process itself. That is a big deal for how we think about training metacognition, error awareness, and performance monitoring in clinical neurofeedback: we may be able to work not only on “being less anxious” or “focusing better”, but also on how the brain evaluates the quality of its own decisions in real time.

This research also ties into broader questions in neuroscience: how the brain integrates noisy sensory evidence, how it decides when to commit to an action, and how internal belief states like confidence arise from the dynamics of neural populations. As we will see, the data suggest that parietal cortex is a key hub where choice and confidence are computed together and made available to downstream systems.

Methods

Two rhesus monkeys were trained on a motion discrimination task in which they viewed random-dot patterns drifting left or right and reported both their perceived motion direction and their confidence with a single saccadic eye movement. Each of four targets on the screen represented a combination of direction (left vs right) and wager (high vs low). A saccade to, say, the “left–high” target meant: “the dots are moving left, and I’m willing to bet high that I’m correct.”

The reward structure was carefully tuned to turn confidence into a meaningful variable. Correct high bets led to a larger juice reward, whereas incorrect high bets incurred a time penalty. Low bets yielded smaller rewards but protected the monkeys from penalties when they were wrong. Importantly, low bets did not correspond to an “opt out”: accuracy on low-bet trials was still high, and both choice and reaction time behaved like genuine deliberative decisions across motion strengths.

Behavior was modeled with several variants of bounded evidence accumulation. In these models, noisy sensory evidence about motion direction is integrated over time until it hits a decision bound, determining both choice and reaction time. The key comparison here was between:

• Serial models: evidence is first accumulated to make a choice; only after that decision is reached does a second accumulation process determine confidence.

• Parallel models: choice and confidence are computed concurrently, with an explicit mapping between the evolving decision variable and the probability of being correct at each moment.

• Hybrid models: an initial phase of parallel accumulation followed by extra post-decision accumulation that refines confidence.

While the animals performed the task, the authors recorded from 407 neurons in the lateral intraparietal area (LIP) in the right hemisphere. LIP is a well-known hub for decision variables in perceptual choice tasks, with neurons that ramp up their firing rate as evidence accumulates in favor of a saccade toward their receptive field.

The authors used a mix of classic single-neuron analyses and more modern population methods. Psychophysical kernels measured how moment-to-moment fluctuations in motion energy influenced both choice and confidence over time. To test the accumulation framework, they examined variance and autocorrelation in firing rates (VarCE and CorCE) for signatures of noisy evidence integration. Finally, they trained logistic classifiers on population spike counts to decode, on a trial-by-trial basis, the evolving decision variable for choice and for confidence, and examined the geometry of these neural codes over time.

Results

Behaviourally, the monkeys did what a good Bayesian decision-maker should do. Accuracy increased with motion strength, and reaction times decreased. High bets were more frequent when the sensory evidence was strong, and when the monkeys bet high, their choices were more accurate and faster than when they bet low. Even within a given motion coherence, trials with shorter reaction times tended to be associated with higher confidence, reproducing classic relationships between speed and confidence.

When behavior was fit with different bounded accumulation models, all variants captured the broad patterns of accuracy and reaction time across motion strengths. However, a critical diagnostic pattern emerged when confidence (the wager) was conditioned on whether the trial was correct or incorrect. Instead of the classic “folded X” pattern (confidence rises with evidence strength on correct trials but falls on error trials), confidence in these monkeys increased with motion strength for both correct and incorrect trials. Serial and hybrid models, which assume that confidence is evaluated after the choice is made, struggled to reproduce this pattern without invoking unrealistically brief post-decision epochs. In contrast, the parallel model naturally captured the observed relationship between evidence strength and confidence on error trials, making it the best-supported account.

The neural data converged on the same story. Single LIP neurons showed ramping firing rates that depended both on the upcoming choice (left versus right) and on the eventual wager (high versus low). Crucially, the divergence in firing rate related to choice and the divergence related to confidence occurred over the same time window, beginning around 200 ms after motion onset. There was no consistent sign that the population first committed to a direction and only later sorted out confidence.

Analyses of variability and autocorrelation in firing rates revealed patterns consistent with noisy evidence accumulation: variance of the latent firing rate increased over time and then collapsed near decision termination, and autocorrelation between time bins showed the characteristic structure of a bounded integrator. These signatures were present for neurons tuned to high and low wager targets alike, arguing against a model in which one pool accumulates evidence for choice first and another pool accumulates later for confidence.

Population decoding made these dynamics even clearer. Logistic classifiers trained to predict choice or wager from LIP population activity produced decision variables that ramped together during evidence accumulation. Prediction accuracy for both choice and confidence rose from chance shortly after motion onset. Around the time of the saccade, choice decoding peaked just before movement, while confidence decoding remained informative slightly afterward, suggesting that confidence signals may continue to update even after commitment.

Interestingly, the decoding vectors for choice and confidence were roughly orthogonal in population space during deliberation, indicating that downstream regions could, in principle, read out either variable without much interference. At the same time, the strength of decoding for the upcoming choice covaried with the likelihood of a high bet, especially for choices toward the contralateral field represented by the recorded neurons. This suggests a tight but geometrically separable relationship between decision and confidence signals in LIP.

Discussion

Taken together, the behavioral modeling and neural data paint a coherent picture: the brain is capable of using a single stream of sensory evidence to serve two computational goals at once—choosing an action and estimating confidence in that action. In the dorsal visual stream, LIP emerges as a key site where this multiplexing happens, with neural populations that track a shared decision variable and express both the direction of the upcoming choice and an evolving sense of how likely it is to be correct.

For clinical and applied work, especially in biofeedback and neurofeedback, this is more than an abstract curiosity. Many problems that bring people to treatment—anxiety, chronic indecision, compulsive checking, impulsive choices, or performance blocks—are not just about raw sensory processing. They often involve a mismatch between evidence and confidence: either feeling too sure when the evidence is weak, or never sure enough even when the evidence is strong. The current study shows that such calibrations are not bolted on at the end of the decision; they are intertwined with the decision variable itself.

In practical terms, this means that when we design neurofeedback protocols to support better decision-making, emotion regulation, or performance, we might think less in terms of a single “attention network” and more in terms of how the brain monitors its evolving choices. LIP and its human homologs sit inside broader parietal–frontal networks that are accessible to surface EEG, especially through parietal midline and parietal–occipital electrodes. Modulating rhythms in these regions—such as posterior alpha for sensory gating and focused attention, or beta activity related to sensorimotor readiness—may indirectly shape how evidence is accumulated and how much confidence is attached to a given action plan.

The finding that choice and confidence signals are nearly orthogonal in population space is also conceptually rich. It suggests that the brain keeps these variables partially separable: the same neural population can carry both, but in a geometry that allows downstream areas to read out one or the other as needed. Translating this to neurofeedback, it reinforces the intuition that a single training signal (for example, an amplitude in one frequency band at one location) may in fact reflect multiple latent functions—attention, motor preparation, and confidence—depending on task context. Good protocol design therefore requires careful task design and symptom formulation, not just a blind focus on increasing or decreasing a given band.

Another important implication concerns timing. Confidence here is shown to be online: it grows as evidence accumulates, is available before the saccade, and continues to evolve immediately afterward. This aligns with clinical observations that people often “feel wrong” or “feel uncertain” even before they can articulate why, and that these feelings can drive either adaptive caution or maladaptive doubt. In neurofeedback, pairing training with tasks that demand graded confidence judgments—for example, continuous performance tasks where clients rate certainty about their responses—could help us tap into these online metacognitive signals instead of only training in resting states.

Finally, the LIP findings sit in a larger ecosystem of metacognitive circuitry that includes prefrontal and cingulate cortices. Confidence signals decoded in parietal cortex could be sent forward to these regions to shape strategy, learning rate, and future behavior. In neurofeedback terms, this encourages a systems-level view: protocols that simultaneously normalize parietal processing and frontal control—through, for instance, combined parietal alpha training and frontal midline theta regulation—may be especially well suited for clients whose main complaint is poor judgment, chronic self-doubt, or difficulty trusting their own perceptions.

The interpretive message here is that confidence is not just a “nice extra” feeling about our decisions. It is built into the dynamics of how the brain accumulates evidence and prepares actions. Any intervention that meaningfully alters those dynamics, including EEG-based neurofeedback, is also likely touching how people experience certainty, doubt, and agency.

Brendan’s perspective

When I read this paper, I immediately thought of the people who sit in front of us and say some version of: “I never know if I’m making the right decision,” or “I decide too fast and regret it later,” or “On game day, my brain suddenly forgets how to trust itself.” What Vivar-Lazo and Fetsch are showing, at a very elegant single-neuron level, is that these experiences are not just stories we tell after the fact. They are baked into how the brain accumulates evidence and tags it with confidence in real time.

In the lab, that story lives in LIP. In the clinic, it shows up as hesitation, impulsivity, overchecking, or performance anxiety. Neurofeedback gives us a way to gently nudge the underlying dynamics rather than just coaching the story at the surface.

One way I like to frame it is this: some brains are biased toward "never sure enough," others toward "too sure, too soon." The first group shows up in anxiety and OCD-like presentations—lots of evidence, never quite enough confidence. The second shows up in ADHD and impulse-control problems—thin evidence, big confident push. Both profiles can exist in the same person in different contexts. The parallel-accumulation story in this paper suggests that when we train attention, sensory gating, or response inhibition, we are also, implicitly, training the confidence system that rides on top of those signals.

For the chronically doubtful, especially those with perfectionistic anxiety or checking behaviors, I often think in terms of stabilising parietal processing and lowering frontal “noise.” On the EEG side, that might look like encouraging a robust but flexible alpha rhythm in posterior sites such as Pz, P3/P4 or Pz/Oz, while gently inhibiting excessive high beta (for example 22–30 Hz) at midline frontal sites like Fz or FCz. The goal is not to sedate frontal control, but to reduce the constant micro-corrections that make every decision feel precarious. In practical terms, this often means pairing neurofeedback with tasks that demand many small decisions under mild uncertainty—simple discrimination tasks where clients must respond and then rate how sure they were. Over time, they can feel the difference between “I actually don’t know” and “my brain is just demanding 110% certainty before it allows me to move.”

On the other side of the spectrum are the "too sure, too fast" brains. Here, the impulsivity is not just a behavioural problem; it is a timing problem in the evidence accumulator. The classic EEG move in ADHD—training SMR (roughly 12–15 Hz) at central sites such as C3, Cz or C4 while inhibiting excess theta (4–7 Hz) and high beta—can be reframed through this paper as a way of improving the quality of the ramping process toward a decision. SMR has long been associated with better motor inhibition and reduced hyperactivity; from a decision perspective, it may also reflect a more stable sensorimotor platform that allows evidence to accumulate just a little longer before the system commits. When we combine this with tasks that require waiting for enough information before responding (for example, continuous performance tasks with variable difficulty), clients are effectively rehearsing the neural version of “pause without freezing.”

Performance and flow bring a slightly different twist. High performers are often very good at accumulating evidence, but their confidence system becomes fragile under pressure: one early error, a spike of self-consciousness, and the internal decision variable is suddenly dominated by noise. Here, parietal–occipital alpha training at Pz/Oz or POz, sometimes in a slightly higher band (for example 10–12 Hz), can help stabilise sensory gating and reduce over-reaction to momentary fluctuations. Paired with tasks that mimic the timing and uncertainty of their sport or art—rapid perceptual decisions, go/no-go with changing rules—we are essentially helping their brain keep the decision and confidence vectors “orthogonal,” in the spirit of the LIP findings: performance signals can fluctuate without dragging confidence up and down with every micro-variation.

A theme that emerges across all of this is that the task matters as much as the protocol. The monkeys in this study never gave a confidence rating in the abstract; confidence was embedded in the action itself, via the wagers. Our clients, likewise, learn most deeply when confidence is embedded in what they are actually doing. That is why I like integrating graded confidence reports directly into neurofeedback tasks: after a block of trials, ask for a simple scale (for example, 1–4: guessing to very sure), and link their brain-based feedback to both objective accuracy and calibration. The brain is not just rewarded for being “calm” or “focused,” but for being appropriately sure given the evidence.

This is also where biofeedback can complement neurofeedback beautifully. Confidence is not only a cognitive or neural calculation; it shows up in heart rate, breathing, and muscle tone. A person can say they are “pretty sure” while their physiology is screaming threat. Simple heart rate variability biofeedback before or during EEG training can help bring the autonomic system into a range where the parietal–frontal decision networks can do their slower, more rational accumulation work. In other words, we are creating the bodily conditions under which LIP-style evidence accumulation can proceed without being constantly hijacked by survival signals.

I also want to highlight the difference between elegant research paradigms and messy clinical reality. In the paper, confidence can be neatly modelled with bounded accumulators and read out by logistic decoders. In clinic, our clients come in with sleep debt, interpersonal stress, trauma histories, and wildly different developmental trajectories. That is why individualisation remains non-negotiable. A qEEG assessment, careful history, and symptom mapping help us decide when a parietal protocol makes sense, when frontal midline theta training for cognitive control should be foregrounded, and when to step back and work on foundational sleep or arousal regulation before we touch metacognition at all.

If there is a practical takeaway for neurofeedback clinicians, it is this: any time we alter how the brain handles incoming information over time—through SMR training, posterior alpha work, or fronto-parietal connectivity protocols—we are implicitly nudging our clients’ confidence systems. We can lean into that by designing tasks and conversations that explicitly explore how it feels to “know enough,” to act with appropriate certainty, and to tolerate the discomfort of not being 100% sure.

My own bias is to see confidence not as a luxury add-on for high performers, but as a core mental health variable. A well-calibrated confidence system lets a person say, “I’ve seen enough; I can choose,” or “I don’t know yet; I’ll wait,” without collapsing into panic or apathy. The LIP data give us a beautifully precise neural story about how that calibration might work. Neurofeedback and biofeedback give us tools to gently shape it in the people who need it most.

Conclusion

This study offers compelling evidence that the primate brain does not wait to “finish” a decision before deciding how confident it feels. Instead, choice and confidence are co-evolving products of the same accumulation process, implemented in sensorimotor regions like LIP that sit at the interface between sensation and action. Parallel computations allow the brain to prepare movements, allocate effort, and adjust learning in real time, guided by a continuously updated estimate of the probability of being correct.

For clinicians and practitioners using biofeedback and neurofeedback, these findings invite a subtle shift in perspective. Rather than seeing confidence as a downstream, purely cognitive construct, we can treat it as a dynamic, trainable property of brain networks that track evidence and prepare actions. By designing protocols and tasks that engage these networks—especially parietal–frontal systems involved in evidence accumulation—we may help clients not only make better decisions, but also feel appropriately confident in them.

The take-home message is simple but powerful: when we help the brain regulate how it builds and evaluates its own decisions, we are not just changing reactions—we are reshaping the very mechanisms of confidence.

References

Vivar-Lazo, M., & Fetsch, C. R. (2025). Neural basis of concurrent deliberation toward a choice and confidence judgment. Nature Neuroscience. Advance online publication. https://doi.org/10.1038/s41593-025-02116-9