Hidden-state inference aligns attention and neural representational geometry during flexible behaviour

Technical Summary: Hidden-state inference aligns attention and neural representational geometry during flexible behaviour

Problem Statement
Flexible decision-making in uncertain environments necessitates the ability to infer latent task states and prioritize information that is behaviorally relevant. While reinforcement learning (RL) models have traditionally been used to explain adaptive behavior, they often struggle to account for rapid context switches where the underlying rules of the environment change. The central problem addressed is understanding the computational and neural mechanisms that allow humans to identify which stimulus dimensions are currently relevant, adapt to context switches, and how these cognitive processes are reflected in both attentional allocation and neural representational structure.

Methodology
The study employed a multi-modal approach combining behavioral modeling, oculomotor tracking, and neuroimaging:

• Task Design: Human participants performed a serial reversal learning task. In this paradigm, participants had to identify which stimulus dimension(s) were currently predictive of rewards, requiring them to infer the latent state of the environment as it switched.
• Behavioral Modeling: The authors compared a hidden-state inference model against multiple reinforcement-learning variants. The hidden-state model was evaluated on its ability to predict both participant choices and their inferred contexts.
• Oculomotor Tracking: Eye movements were recorded to provide converging evidence for selective attention. Metrics included gaze concentration on relevant features, the tracking of trial-by-trial belief updates, and the dynamics of gaze following high prediction errors.
• Neuroimaging: Magnetoencephalography (MEG) was utilized to examine the neural correlates of latent-state inference. The analysis focused on how neural activity encoded the inferred context and how it was modulated by recent prediction errors.
• Representational Analysis: The study employed representational geometry analysis to investigate the structure of neural representations, specifically looking for changes in the distance between neural patterns representing different stimulus values.

Key Results

• Behavioral Adaptation: Participants rapidly adapted to context switches. The hidden-state inference model provided the best explanation for these adaptations, significantly outperforming various RL variants in predicting both choices and inferred contexts.
• Attentional Dynamics: Oculomotor data confirmed that gaze became increasingly concentrated on behaviorally relevant features. This attentional focus tracked trial-by-trial belief updates. Notably, following high prediction errors, gaze patterns transiently broadened, suggesting a temporary shift in attentional scope.
• Neural Encoding: Neural representations successfully encoded the currently inferred context. These representations were modulated by recent prediction errors; specifically, larger prediction errors were associated with weaker context representations on the subsequent trial.
• Reorganization of Neural Geometry: Successful performance was linked to a systematic reorganization of neural representational geometry. This reorganization was characterized by the selective amplification of representational distances for task-relevant features and an increased separation of behaviorally relevant stimulus values.

Key Contributions
The paper provides a systems-level account of flexible decision-making by demonstrating that latent-state inference, attention, and neural representational geometry are tightly coordinated. It establishes that:

1. Hidden-state inference is a superior computational framework for explaining rapid behavioral adaptation in uncertain environments compared to standard RL models.
2. Oculomotor behavior serves as a direct, trial-by-trial marker of belief updates and attentional allocation during context switching.
3. Neural representational geometry is not static but dynamically reorganizes to support adaptive behavior, specifically by amplifying the neural distinction between relevant stimulus values when performance is successful.

Significance
The authors claim that their findings elucidate how behaviorally relevant information is selectively prioritized and represented to support adaptive behavior. By linking computational models of latent-state inference with specific oculomotor signatures and dynamic changes in neural representational geometry, the study offers a comprehensive mechanism for how the brain manages uncertainty and maintains flexibility in decision-making. The work suggests that the brain does not merely update value estimates but actively restructures its internal representational space to align with the inferred task context.