Neural landscape diffusion resolves conflicts between needs across time

The study addresses how the brain resolves conflicts between co-occurring physiological needs (for example, hunger and thirst) when both relevant rewards are equally available. Classical views treat needs as distinct forces acting directly on behaviour, raising the Buridan’s ass problem of indecision under equal needs. The authors posit that a more complete framework must (1) relate the intensity/salience of individual needs to momentary choices, (2) identify neural bases for choices, and (3) explain the dynamics of switching between need-appropriate behaviours. They introduce Buridan’s assay to probe how mice organize choices across time under conflicting needs and hypothesize that an internal, persistent goal-like neural state, modulated by needs and noise, guides behaviour and transitions.

Prior work established neural circuits for detecting individual physiological imbalances and generating goal-directed behavioural and neural states for thirst and hunger. Historical treatments (Aristotle; Buridan’s ass) highlight the conceptual challenge of decision under equal needs. Studies compared thirst and hunger circuits and interactions but not their moment-by-moment conflict resolution under equal access to rewards. Broader literature shows brainwide population dynamics representing motivational states and mixed selectivity across neurons, suggesting that distributed network states could support goal selection. This context motivates searching for a distributed, persistent neural goal state and a dynamical framework (inspired by Langevin dynamics and energy landscapes) to explain stochastic switching between behaviours under conflicting needs.

Behavioural paradigm (Buridan’s assay): Head-restrained mice were food- and water-restricted and placed before two equidistant spouts delivering water or salted liquid food. An olfactory Go/No-Go design was used: a Go odour indicated reward availability (67% of trials), and the mouse’s first lick direction (left/right) determined whether water or food was delivered; a No-Go odour (33%) signalled no reward, requiring lick withholding during the inter-trial interval. Trained mice performed hundreds of trials per session, incrementally collecting rewards until satiety, under conditions of food-only restriction, water-only restriction, both, or none. Behavioural analyses: Choice sequences were analysed for bout structure and persistence. Bout length distributions were compared to a sticky Markov (geometric) process. Momentary relative thirst versus hunger was operationally defined as the cumulative future amount of water versus food the animal would collect until satiation; a normalized thirst−hunger metric (−1 to +1) was computed per trial. Choice predictability from needs versus previous choice was evaluated via linear regression and support vector machine decoding (AUC). Transition/self-transition probabilities were quantified as functions of relative need and time between choices (assayed via numbers of No-Go trials between Go trials). Optogenetics: To causally modulate thirst, RXFP1+ osmotic thirst neurons in the subfornical organ were transduced with AAV5-Ef1α-DIO-hChR2(H134R)-YFP in Rxfp1-2A-Cre mice. Brief stimulation epochs (for example, 10 s, 20 Hz) were delivered during Buridan’s assay in hungry-only mice and in sated mice, and also during a separate quiet waiting period (no odour/spouts) in Neuropixels sessions to probe neural effects without behavioural confounds. Probabilistic transitions and persistence after stimulus offset were measured. Electrophysiology: Large-scale extracellular recordings were performed using four acute Neuropixels 1.0 probes (1,536 channels) with trajectories spanning frontal/motor cortices, basal ganglia, thalamus, hypothalamus, and midbrain. Hundreds of single units per session were isolated; anatomical locations were recovered by atlas alignment. Baseline (1 s pre-odour) activity was emphasized to avoid movement and cue confounds. Regression analyses quantified information about upcoming choice and other task variables; mixed selectivity was characterized; decoding analyses assessed population predictiveness of future choice. Functional coupling and population variance explained were also evaluated. Neural population analyses: A ‘goal dimension’ was defined as the vector separating average baseline population activity before upcoming water versus food choices; trial-wise neural states were projected along this dimension to track goal-related dynamics across time and around behavioural switches. Modelling: A forward generative model was formulated using Langevin-like stochastic differential equations in a low-dimensional need subspace. Population neural activity diffuses across an energy landscape with wells corresponding to goals (food, water, neither); well depths are scaled by dynamic need magnitudes (hunger, thirst), which decrease as rewards are consumed. Dynamics are governed by the landscape gradient (needs) and additive white noise (fast-timescale influences). Three fixed parameters were fit from trial-by-trial data using theoretical expressions for equilibrium and transition probabilities derived from non-equilibrium statistical mechanics: scaling of gradient, scaling of noise, and a weight for thirst/hunger relative to other needs. Simulations of Buridan’s assay and optogenetic perturbations were run forward in time, repeated 128 times per dataset to capture stochastic variability. Model predictions were compared with experimental behavioural and neural data, including transition probabilities over time intervals and neural goal-dimension dynamics before switches and under thirst stimulation.

• Behavioural structure: Under simultaneous hunger and thirst, mice organized choices into persistent bouts with sudden, stochastic transitions, inconsistent with hierarchical, deterministic relative-need, or random-choice models. Bout length distributions matched a geometric distribution from a sticky Markov process (maximum likelihood shape parameter P=0.061, 95% CI [0.05, 0.074]).
• Needs vs history: The probability of choosing water on a trial correlated with normalized relative need (thirst−hunger) with R²=0.92 and slope=0.426, but the most recent previous choice predicted current choice significantly better than need magnitudes (SVM AUC higher for previous choice vs current needs; paired t-test, n=22 sessions, t=5.89, P=6.28×10⁻⁷). Repeat-choice probabilities remained high (>80%) and exceeded 90% when needs were approximately balanced (−0.25 ≤ thirst−hunger ≤ 0.25). Self-transition probability varied linearly with relative need (water choice: R²=0.612, slope=0.07; food choice: R²=0.844, slope=−0.077). Transition probability increased with longer intervals between choices under constant needs, matching model theory (model vs experiment R²=0.411).
• Optogenetic causal tests: Brief activation of subfornical organ RXFP1+ osmotic thirst neurons in sated mice elicited probabilistic transitions to water choices that persisted beyond stimulation; in hungry-only mice, stimulation induced probabilistic switches from food to water. Transitions within epochs were stochastic and not dictated by immediately preceding rewards; water choices persisted for ≥10 s after stimulus offset, indicative of a persistent internal state.
• Distributed neural correlates: About 20% of recorded neurons carried significant information about upcoming choice in baseline (1 s pre-odour) firing rates, with informative cells distributed across hypothalamus, midbrain, striatum, and cortex (hypothalamic/midbrain regions exhibited higher aggregate baseline information than cortex). Most neurons showed mixed selectivity to multiple task variables (satiety, odour, choice). Population baseline activity predicted upcoming choice with high accuracy and outperformed subtle movement predictors; prediction improved with more simultaneously recorded neurons and accounted for ~10% of trial-by-trial variance in the pre-odour period. Neurons with goal information were more functionally coupled within and across regions.
• Goal-dimension dynamics: Projecting population activity onto a goal dimension (water vs food) revealed persistent state-like activity within choice bouts and rapid trajectories during behavioural transitions, both in data and model simulations. Predictiveness of baseline activity dropped immediately before switches, consistent with noise-driven transitions near a decision boundary; the distance from midpoint along the goal dimension predicted the probability of an upcoming switch. Baseline goal-dimension magnitude alone could predict imminent switches.
• Generative model performance: The neural landscape diffusion model reproduced key behavioural statistics: choice persistence-length distributions, choice probabilities vs relative need, self-transition modulation by need, and the increase in switching probability with time between choices. Simulated optogenetic thirst (transient deepening of the thirst well) produced probabilistic switches and post-stimulation persistence, with decay dynamics matching experiments.
• Causal neural state modulation: Optogenetic thirst stimulation during a quiet period shifted neural activity along the goal dimension toward water seeking and decayed slowly after offset, while a control (Go vs No-Go odour dimension) showed no significant change. Individual stimulation epochs produced variable neural goal-state trajectories, consistent with stochastic dynamics predicted by the model.

Findings show that mice resolve conflicts between hunger and thirst not by instantaneous comparison of needs but via a persistent, distributed neural goal state that governs future choices and transitions stochastically. Behavioural persistence with probabilistic transitions aligns with a Markov-like process modulated by relative need but dominated by internal state history. Baseline neural population activity widely distributed across brain regions encodes this goal state and predicts upcoming choices; loss of discriminability along the goal dimension precedes switches, supporting noise-driven transitions across an underlying energy landscape. The neural landscape diffusion model parsimoniously links needs (via reshaping well depths), state persistence (landscape gradient), and stochasticity (noise) to produce observed behaviour and neural dynamics without invoking an external transition controller. It resolves the Buridan’s ass paradox by positing that the internal goal state evolves continuously; if delayed, the state can diffuse and cross decision boundaries even under equal needs. The framework suggests that global goal context can serve as an initial condition coordinating downstream, task-evoked dynamics to implement specific actions after sensory cues, consistent with hierarchical control theories. Broader implications include understanding how the balance between gradient and noise regulates transition rates and stability of internal states; deviations could map onto psychiatric phenotypes with excessive switching or pathological persistence. The model provides mechanistic predictions for how neuromodulators or circuit mechanisms that adjust landscape shape or noise levels could alter behavioural organization across time.

This work introduces Buridan’s assay to study moment-by-moment resolution of conflicting physiological needs and identifies a persistent, distributed neural goal state that organizes behaviour into stable bouts with stochastic transitions. Through large-scale Neuropixels recordings, optogenetic perturbations, and a minimal Langevin-inspired landscape diffusion model, the authors show that needs bias but do not directly determine choices; instead, they reshape an energy landscape across which a goal state diffuses. The model quantitatively recapitulates behavioural statistics, neural population dynamics, and optogenetically induced effects with few parameters. Main contributions: (1) empirical demonstration of persistent, stochastic choice organisation under conflicting needs; (2) discovery of widespread baseline neural correlates predictive of future choice and of impending switches; (3) a generative neural landscape diffusion model unifying needs, state, and noise; (4) causal evidence that thirst modulates the internal goal state. Future directions include identifying neurobiological mechanisms that remodel the landscape (for example, neuromodulators, synaptic reweighting), dissecting circuit substrates that set gradient and noise scales, extending the framework to additional homeostatic and affective drives, and testing generalization to freely moving conditions and to human cognition and psychiatric disorders.

• Brain coverage: Although recordings spanned many regions, only a fraction of the brain was sampled; a primary locus controlling transitions could have been missed.
• Preparation and task constraints: Head-restrained Go/No-Go design and olfactory cues may not capture all aspects of naturalistic decision-making; although freely moving results were referenced, detailed data were in extended materials.
• Model simplifications: The low-dimensional needs subspace, white-noise approximation, and fixed parameterization abstract away biological complexities; the exact neural implementation of the energy landscape and noise sources remains unidentified.
• Specific drives: The study focuses on hunger and thirst; generalization to other needs, motivations, and affective states requires further testing.
• Mixed selectivity and movement: While neural predictors outperformed movement variables, subtle movements and mixed selectivity could still confound some neural-behaviour relationships.