Animals must decide whether to persist with a resource patch or move on. Investigators reported mice faced stay-or-leave patch foraging choices for water in environments labeled poor or rich, where local reward outcomes changed with time. The cohort showed that choices were not driven solely by the most recent reward. Instead, decision-making incorporated immediate outcomes, recent local reward history, and a longer-timescale expectation of reward magnitude — described by the authors as a learned value that favored greater time investment in rich-environment patches. This pattern emerged across dynamic patches, indicating animals adjusted strategies when background reward statistics shifted. The abstract did not list sample sizes or full task parameters; those methodological specifics were not provided in the excerpt used here.
Computational analyses revealed a structured bias in decision-making that varied over the course of a patch visit. The study team fitted models to behavior and isolated a time-varying bias term whose temporal profile mirrored the shape of a constructed value function. In plain terms, mice appeared to carry an evolving expectation about how much a patch was worth, and that expectation altered the threshold or propensity to leave as time passed. The investigators framed this as consistent with classical reinforcement learning algorithms, where a value signal accumulated from past experience modulates action selection. The abstract did not present full model equations or parameter values; those details were not reported in the available text.
Electrophysiological and/or neurochemical readouts were used to link neural signals to behavior. Data indicated that dopamine activity in the ventral striatum followed a temporal-difference pattern consistent with a reward prediction error — the difference between expected and received reward predicted by reinforcement learning theory. The authors noted that such an RPE could operate as a teaching signal, updating value expectations over time so that animals learn which patches are more valuable. The abstract did not specify recording modalities, signal timing relative to choices, or statistical metrics; those methodological specifics were not included in the excerpt available for this rewrite.
To test causality, investigators employed calibrated optogenetic manipulation of dopamine to create fictive reward signals. When dopamine was artificially modulated to mimic prediction-error-like events, the cohort showed a recapitulation of the value-related decision bias observed naturally. In other words, creating a synthetic teaching signal altered subsequent patch persistence in the direction predicted by the learned value function. The study team interpreted these manipulations as direct evidence that dopaminergic RPEs can sculpt temporally resolved value expectations and thereby bias foraging decisions. The abstract did not supply stimulation parameters, control conditions, or the behavioral effect sizes; the excerpt lacked those experimental details.
The authors posted this work as a preprint on October 22, 2026 and declared no competing interests. The abstract explicitly frames findings within a reinforcement learning account and links dopamine to a temporal-difference teaching signal, but it does not provide exhaustive methodological reporting in the excerpt used here. Specifically, the abstract did not report sample sizes, exact task timing, statistical values, or the precise optogenetic protocols used. Readers should note the preprint status: these results had not been certified by peer review at the time of posting. Open questions remain and were not resolved within the abstract: how stable are these learned value functions across contexts, what is the precise temporal resolution of the dopaminergic RPE relative to behavioral events, and how generalizable are the findings across species or naturalistic foraging settings. Future peer-reviewed publication and access to the full methods and datasets will be necessary to evaluate reproducibility and mechanistic detail.
Acknowledgments and provenance: the work was authored by Lyle E. Kingsbury, Hugo May, and Naoshige Uchida, affiliated in the abstract with Harvard University; the preprint appeared on bioRxiv with DOI https://doi.org/10.64898/2026.10.21.760765. The authors stated no conflicts of interest in the provided text.