Anxious Minds Mistake Random Noise for Real Change, and Deep Sleep Corrects It

When the world around us is unpredictable, the brain faces a subtle but crucial computational problem: it must decide whether an unexpected outcome reflects genuine change in the environment or merely random noise. A new study published in iScience by Rakshita Deshmukh and Arjun Ramakrishnan of the Indian Institute of Technology Kanpur suggests that people with high trait anxiety struggle with exactly this distinction, treating random fluctuations in rewards as evidence that the rules of the game have shifted. Strikingly, the researchers also found that a single night of deep, slow-wave sleep, known as N3 sleep, appears to recalibrate this miscalibration, pulling anxious learners’ strategies back toward what a mathematically optimal learner would do. The findings, based on two preregistered experiments, offer some of the clearest evidence yet that anxiety is not a generalized learning deficit but a highly specific distortion in how the brain infers the structure of uncertainty.

To disentangle the two kinds of uncertainty, the team designed a three-option probabilistic reversal-learning task framed as a fishing game, in which participants chose among three islands to maximize their catch. The researchers independently manipulated two dimensions: stochasticity, the random variability of outcomes under stable rules, and volatility, the rate at which the underlying reward contingencies actually changed. In the low-stochasticity environment, the best patch paid off on 80 percent of trials; in the high-stochasticity environment, it paid off only 60 percent of the time, making individual outcomes far less informative. Volatility was manipulated by reversing which patch was best every 25 trials in slow conditions and every 10 trials in fast ones. Because normative learning theory predicts that learning rates should rise under volatility but fall under high stochasticity, this design allowed the researchers to test whether anxious individuals violate these prescriptions in a selective way.

In the first experiment, fifty participants completed six 50-trial blocks spanning all combinations of stochasticity and volatility, and trait anxiety was measured with the State-Trait Anxiety Inventory. The results were unambiguous: participants with high trait anxiety, defined as a trait score of 45 or above, earned significantly fewer rewards than their low-anxiety peers, but only in environments that were stable yet noisy, that is, under slow volatility combined with medium or high stochasticity. When the environment genuinely changed rapidly, the two groups performed identically. This selectivity is the study’s central surprise. It rules out the idea that anxiety produces a blanket overestimation of environmental change; instead, anxious individuals remain sensitive to true volatility but fail to discount random outcome noise appropriately when the environment is actually stable.

Computational modeling revealed the mechanism behind this performance gap. The researchers fit a hierarchical Bayesian reinforcement-learning model with separate learning rates for rewards and punishments, allowing them to quantify how strongly each participant updated their beliefs after positive and negative feedback. Contrary to the widely held view that anxiety amplifies learning from negative outcomes, the elevated learning rates in high trait anxiety were reward-driven. Under slow volatility, higher trait anxiety predicted increasing reward learning rates as stochasticity rose, with 95 percent highest density intervals well above zero in both medium and high stochasticity conditions. Under fast volatility, reward learning rates did not vary with anxiety at all. Punishment learning rates, meanwhile, remained below task-optimal levels in both groups and did not differ between them, indicating that punishment sensitivity could not explain the performance differences.

To determine whether these elevated learning rates were actually harmful, the team simulated reinforcement-learning agents across nearly sixty thousand parameter combinations to identify the reward-maximizing learning rate for each environment. As normative theory predicts, the optimal learning rate increased with volatility and decreased with stochasticity. Against this benchmark, high trait anxious individuals significantly exceeded the optimal reward learning rate in the stable, noisy environments, with Bayes factors exceeding ten billion in the medium-stochasticity condition. Higher learning rates were in turn associated with lower reward accumulation, and anxious participants clustered in the region of high learning rates and low earnings. In other words, anxious learners were overfitting noise, updating their beliefs on the basis of random reward fluctuations as if each surprise signaled a genuine shift in the world.

The decision process itself was also altered. Using hierarchical drift-diffusion modeling, which decomposes reaction times into parameters such as drift rate, the speed of evidence accumulation, and boundary separation, the threshold of information required before committing to a choice, the researchers found that in the slow-volatility, high-stochasticity environment, high trait anxious participants showed drift rates centered near zero, dramatically slower than those of low-anxiety peers, along with elevated decision boundaries. Under fast volatility, none of these differences appeared. This means anxious individuals were not only updating their beliefs too aggressively but also accumulating evidence more slowly and cautiously when valuing the most rewarding option, a double computational burden confined precisely to environments that demand a clean separation of signal from noise.

The second experiment asked whether sleep, a physiological state tightly intertwined with anxiety, could modulate these computational phenotypes. Forty participants performed the task before and after sleeping in the laboratory, with overnight sleep recorded using a validated wearable EEG headband. Greater N3 sleep, the deepest stage of non-REM sleep characterized by slow delta oscillations and strongly linked to synaptic plasticity, was associated with a larger evening-to-morning reduction in state anxiety. Contrary to the researchers’ initial prediction, it was the high trait anxious participants who derived the greatest anxiolytic benefit, showing the steepest dissipation of morning state anxiety as their N3 duration increased. No comparable associations emerged for lighter N1 or N2 sleep or for REM sleep.

The restorative benefit of deep sleep extended directly into the realm of learning. Post-sleep reward accumulation was higher than pre-sleep performance overall, and greater N3 sleep specifically predicted higher post-sleep earnings and more accurate identification of the most rewarding patch in high trait anxious individuals, particularly under slow volatility. At the computational level, reward learning rates decreased from pre- to post-sleep under high stochasticity across all participants, consistent with normative adjustment, and greater N3 sleep was associated with significantly reduced post-sleep reward learning rates in the anxious group. When compared against the simulated optimal benchmarks, participants with more N3 sleep showed learning rates that were statistically indistinguishable from optimal values after sleep, whereas those with less N3 sleep continued to exceed the optimal benchmark. Post-sleep drift rates also rose, indicating faster, more confident evidence accumulation favoring the best option.

Taken together, the two experiments sketch a coherent picture in which trait anxiety selectively corrupts the inference that separates stochastic noise from genuine environmental change, inflating reward learning rates in stable but noisy settings and slowing the accumulation of decision evidence, while N3 sleep acts as a state-dependent corrective that nudges these parameters back toward their task-optimal values. The authors caution that N3 sleep was measured rather than experimentally manipulated, so its effects remain partly correlational, and that related constructs such as intolerance of uncertainty were not assessed, leaving open whether the observed effects are unique to trait anxiety. The sample was also drawn from a single campus community, and circadian factors were not controlled. Even so, because the trade-off between stochasticity and volatility is fundamental to countless real-world learning problems, from financial markets to social relationships, the study raises the provocative possibility that protecting deep sleep could be a tangible, physiological lever for improving how anxious minds learn under uncertainty.

Subject of Research: How trait anxiety and N3 sleep shape reward learning and decision-making under stochasticity and volatility

Article Title: Trait anxiety selectively amplifies reward learning under uncertainty, while N3 sleep attenuates it

Article References: Deshmukh, R., & Ramakrishnan, A. (2026). Trait anxiety selectively amplifies reward learning under uncertainty, while N3 sleep attenuates it. iScience, 29(10), Article 117745. https://doi.org/10.1016/j.isci.2026.117745

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.117745

Keywords: trait anxiety, reward learning, uncertainty, stochasticity, volatility, N3 sleep, reinforcement learning, drift-diffusion model, reversal learning, decision-making, sleep EEG, computational psychiatry