In a groundbreaking new study that shines light on the neural underpinnings of decision-making, researchers have unveiled how mice incorporate prior experiences into their choices with remarkable optimality. This ambitious brain-wide inquiry reveals that the animal’s subjective priors—the internal expectations influencing decisions—are principally anchored in their own past actions rather than the sensory stimuli they encountered. This nuanced understanding challenges pre-existing notions and opens up fresh pathways for how complex cognitive processes emerge from neural circuits.
The investigation carried out by Findling, Hubert, and the International Brain Laboratory team employed a sophisticated combination of behavioral experiments and cutting-edge brain-wide neural recordings. The mice were engaged in perceptual decision tasks where sensory uncertainty was manipulated through varying contrasts. Analysis of neural data across multiple brain regions demonstrated that the subjective priors manifest robustly at all levels of processing—from early sensory zones such as the lateral geniculate nucleus and primary visual cortex to associative zones including orbitofrontal and anterior cingulate cortices, extending even to motor areas responsible for executing decisions.
What stands out in this study is the discovery of bidirectional communication loops, as revealed by Granger causality analyses, that shuttle these prior-related signals across cortical and subcortical areas. This multidirectional flow underpins a distributed inference mechanism strikingly reminiscent of Bayesian networks, supporting the brain’s capacity to integrate past actions in shaping future choices. Such brain-wide coordination challenges the simplistic feedforward view of sensory processing, suggesting instead that priors and beliefs permeate even early stages of sensory representation.
These findings call for a shift in perspective. The so-called subjective prior is more than simple motor preparation or mere attentional modulation. Its neural signature correlates strongly with the animal’s eventual choices, particularly in trials devoid of strong sensory evidence. Furthermore, it integrates information over several previous trials rather than reflecting only the immediately preceding choice, indicating a temporal depth to the neural coding of expectations. This layered encoding implies a memory-dependent predictive framework intricately embedded in brain circuits.
Intriguingly, the patterns of neural activity encoding the priors satisfactorily fulfill criteria posited by the theory of linear probabilistic population codes. That is, the Bayes-optimal prior’s log odds can be linearly extracted from population neural activity, and neuronal dynamics capture trial-to-trial modifications of this prior. Such encoding schemes offer a computationally elegant framework, allowing priors and sensory likelihoods to be combined through straightforward linear operations. This alignment bridges neurophysiological observations with long-standing theoretical constructs in computational neuroscience.
The likelihood itself, representing the sensory evidence, has been demonstrated in primate models to obey a similar linear probabilistic population code. Hence, encoding both prior and likelihood in compatible neural formats could significantly streamline the computation of posterior beliefs during decision-making. This aligns strongly with ideas that the brain approximates Bayesian inference by exploiting structured population codes, delivering efficient and flexible adaptations to uncertain environments.
However, the study also acknowledges the challenge of definitively disentangling the neural coding schemes at play. Alternative hypotheses such as sampling-based probabilistic codes cannot yet be ruled out, given ongoing debates about what precise features of neural variability correspond to probabilistic sampling versus other coding strategies. The simplicity of the Bernoulli prior used in the experiments—essentially a binary probabilistic framework—complicates attempts to differentiate these theoretical models experimentally.
Notably, the subjective prior signals are not localized in isolation but emerge as a coordinated phenomenon across disparate brain regions. Early sensory areas embed prior information alongside incoming stimuli, associative cortices integrate and propagate expectations, and motor circuits implement the downstream behavioral choices informed by these complex computations. This holistic, distributed architecture is highly suggestive of a large-scale Bayesian inference network that operates in parallel, with continuous feedback and updating.
The research harnesses an unprecedented brain-wide dataset provided by the International Brain Laboratory, which compiles multi-regional neural recordings synchronized with detailed behavioral measurements. This resource enables the authors to map with fine granularity how past choices modulate ongoing neural activity, and how this modulation predicts future decisions. The richness of the dataset forms a fertile ground for future explorations that could further elucidate the algorithmic properties of neural inference.
Moving forward, the authors emphasize the critical necessity of developing sophisticated neural models capable of simulating Bayesian inference in modular and recurrent networks reflective of the complex brain architecture. Such models would ideally capture the multidirectional loops and temporal integration of priors identified empirically. This remains a pressing challenge, demanding integration of empirical neuroscience, theoretical modeling, and advanced computational frameworks.
This study embodies a pivotal step towards elucidating how brains incorporate history-dependent expectations into moment-to-moment choice behaviors. By revealing that prior information pervades across sensory, associative, and motor regions via dynamic recurrent circuits, it lays down a detailed mechanistic foundation for understanding adaptive behavior. In a broader context, these insights may help clarify the neural bases of learning, memory, and probabilistic reasoning, with implications extending from fundamental neuroscience to artificial intelligence.
Above all, these findings assert that decision-making is a deeply integrative brain-wide operation—a tapestry woven from the threads of past choices, current sensory inputs, and predictive computations. The confluence of experimental rigor, theoretical insight, and comprehensive brain mapping heralds a new era in the neuroscience of inference, promising to unravel how subjective beliefs shape perceptions and actions with elegant precision.
Subject of Research: Neural encoding of prior information and Bayesian inference mechanisms in mouse decision-making.
Article Title: Brain-wide representations of prior information in mouse decision-making.
Article References:
Findling, C., Hubert, F., International Brain Laboratory. et al. Brain-wide representations of prior information in mouse decision-making. Nature 645, 192–200 (2025). https://doi.org/10.1038/s41586-025-09226-1
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s41586-025-09226-1
Keywords: Bayesian inference, subjective prior, decision-making, probabilistic population codes, mouse brain, neural coding, sensory integration, motor preparation, brain-wide neural dynamics
