The team designed a behavioral paradigm that allowed them to measure and even influence how participants made perceptual decisions across two distinct tasks. In the setup, animals and human participants had to judge stimuli along two different feature dimensions, and critically, they had to figure out on their own which of the two tasks was currently in effect. Because the task rule changed unpredictably, participants had to maintain and update an internal belief about which task was relevant, informed by whether previous trials had been rewarded. This design meant that the experimenters could track, trial by trial, how confident each participant was about the current task and how that confidence shaped the quality of their perceptual judgments.

The behavioral results were striking and consistent across species. Both humans and monkeys made less accurate perceptual decisions when the task was uncertain, and both were slower to respond under those conditions. An ideal observer analysis showed that this drop in perceptual accuracy was not a necessary consequence of switching between tasks; a perfectly rational agent could, in principle, change its task belief without sacrificing perceptual precision. Yet the biological participants consistently did. This mismatch between what is achievable and what brains actually do set the stage for the central question of the study: what, mechanistically, causes perception itself to degrade when the task is ambiguous?

To generate mechanistic hypotheses, the researchers turned to recurrent neural networks. They trained two types of networks on the same tasks the monkeys performed. One network, called the correct choice network, was trained simply to produce the best possible answers. The other, called the monkey choice network, was trained to replicate the actual, suboptimal choices of the monkeys, including their delays in switching tasks and their errors. Remarkably, only the network trained to mimic the monkeys reproduced the uncertainty-related drop in perceptual accuracy. The correct choice network switched tasks efficiently and maintained high perceptual performance throughout. This dissociation suggested that the behavioral cost was not an unavoidable property of the task architecture but instead reflected something specific about how biological brains manage uncertainty.

Peering inside the trained networks revealed the mechanism. Using distance covariance analysis, the team quantified how strongly the two stimulus features, the relevant and the irrelevant one, were represented in the network’s hidden layer activity during the interval between stimuli. In the monkey choice network, after unrewarded trials that lowered task certainty, the representation of the irrelevant feature became substantially stronger and the two feature representations lost their orthogonality, becoming entangled with one another. In the correct choice network, by contrast, the irrelevant feature was suppressed and the two feature axes remained nearly orthogonal even under high uncertainty. In other words, the network that replicated animal behavior literally mixed together the neural codes for features that should have been kept separate.

Armed with this hypothesis, the researchers turned to the brain itself. Recording from populations of neurons in the primary visual cortex, V1, of monkeys performing the task, they found the same signature of feature interference. Under high task certainty, V1 population activity primarily encoded the believed relevant feature, and information about the irrelevant feature decayed rapidly after stimulus offset. Under low task certainty, however, the irrelevant feature lingered in the population activity, and the neural representations of the two features became less separable. Decoders trained to read out feature values from V1 performed far better on the irrelevant feature during uncertain trials, mirroring the pattern seen in the monkey choice network and providing a direct neuronal correlate of the behavioral cost.

The team then pushed the analysis further, asking whether the animal’s own perceptual confidence could be decoded from V1 and used to predict behavior. The strength of the V1 representation of feature changes predicted how likely the monkey was to switch tasks after an unrewarded trial: when population activity indicated a strong, obvious change in the believed relevant feature, task switches were more frequent. Even among trials that were behaviorally identical, the decoded confidence from visual cortex carried information about the animal’s subsequent decision. This finding links a precise, measurable property of early sensory cortex to the higher-order process of deciding which task to perform, bridging perception and metacognition within a single neural circuit.

Crucially, the researchers went beyond correlation. Through behavioral experiments in humans, physiological recordings, and causal experiments using microstimulation in monkeys, they demonstrated that feature interference actively causes errors under uncertain conditions rather than merely accompanying them. Strengthening the representation of irrelevant features, whether through the natural dynamics of task uncertainty or through direct manipulation of the circuit, biased perceptual choices in predictable directions. Human participants showed idiosyncratic but systematic couplings between irrelevant and relevant feature judgments, and the monkeys exhibited similar directional interference across sessions, reinforcing the idea that the effect reflects a genuine representational mechanism rather than a generic lapse in attention or motivation.

The broader implications of the study reach into one of the oldest questions in psychology: why are cognitive capacities limited? The authors propose that capacity limitations, whether in multitasking, working memory, or attention, may stem fundamentally from interference between neural representations of different stimuli, tasks, or memories. When the brain cannot fully suppress one representation, it bleeds into another, and the resulting entanglement degrades both. This framework reframes the classic switch costs documented in decades of task-switching research as a consequence of representational geometry: under certainty, the brain maintains orthogonal codes for different features and tasks, but uncertainty erodes that orthogonality, allowing signals to contaminate one another.

The study also showcases a methodological blueprint that is likely to influence the field. By training one network to be optimal and another to match animal behavior, and then comparing their internal representations, the researchers could generate a falsifiable mechanistic hypothesis and test it with both observational and causal tools in biological brains. The monkey behavioral, electrophysiology, and microstimulation data, together with the human psychophysics data, have been made openly available, and the code for network training and analyses has been released to the research community. As artificial neural networks increasingly serve as models of cognition, this work demonstrates both the power and the subtlety of that approach: networks can reveal not only what brains do, but, when trained on behavior rather than on correctness, why brains fall short of their own impressive potential.

Subject of Research: Neuronal mechanisms of task uncertainty and feature interference in perceptual decision-making

Article Title: Feature interference underlies a neuronal basis for the behavioral cost of task uncertainty

Article References: Xue, C., Markman, S. K., Chen, R., Kramer, L. E., & Cohen, M. R. (2026). Feature interference underlies a neuronal basis for the behavioral cost of task uncertainty. Nature Neuroscience. https://doi.org/10.1038/s41593-026-02430-w

Image Credits: AI Generated

DOI: 10.1038/s41593-026-02430-w

Keywords: task uncertainty, feature interference, decision-making, visual cortex, neural networks, perception, task switching, neuroscience, Nature Neuroscience, cognitive capacity, microstimulation, psychophysics