Decision-Making & Cognitive Control

Contributed Talk Session: Tuesday, August 4, 10:15 – 11:15 am, Skirball Theater

A structured population code for a symbolic action grammar

Talk 1, 10:15 am

Lucas Y. Tian¹, Daniel J. Hanuska¹, Kedar Garzón Gupta¹, Yue Liu², Xiao-Jing Wang³, Joshua B. Tenenbaum⁴, Winrich Freiwald¹; ¹Rockefeller University, ²Florida Atlantic University, ³New York University, ⁴Massachusetts Institute of Technology

Presenter: Lucas Y. Tian

A hallmark of intelligence is the ability to flexibly solve problems by reusing, or generalizing, prior knowledge. Compositional generalization—the ability to recombine discrete units of knowledge in a rule-based (or grammatical) manner—is considered essential for solving problems never previously encountered, by allowing access to a combinatorial space of new representations (e.g., drawing a new animal by recombining legs, heads, and arms). Compositionality is broadly implicated, from language and reasoning to communication and action, and is a basis of imagination and symbolic abstraction. Yet, despite this importance, its neural basis remains unclear. Two question are critical: How does the brain (i) represent discrete units of knowledge (symbols), and (ii) implement grammars to recombine those units? To address the first question, a recent study by Tian et al. (2025) established a model system for compositional generalization, combining a compositional drawing task with large-scale neural recordings in macaque monkeys, and identified a neural population encoding discrete, recombinable action categories action symbols in ventral premotor cortex (PMv). Here, we build on this system to address the second question—how does the brain implement action grammars? We trained macaques to use grammatical rules to draw geometric figures, revealing a striking capacity for macaques to generalize action grammars. Using multi-area neuronal recordings and microstimulation, we identified population activity patterns that explicitly code for variables in grammars, specifically in the presupplementary motor area (preSMA). The geometric structure of activity encodes “slot subspaces” that can explain behavioral capacities for generalization. Thus, the grammatical ability to recombine discrete actions into novel sequences reflects a structured population code in preSMA. This finding establishes a foundation for understanding how PMv, preSMA, and interconnected areas recombine symbols in a grammatical fashion to enable compositionality.

Orthogonal Task Representation Enables Flexible Task Switching in Human OPM-MEG and Neural Networks

Talk 2, 10:25 am

Xiaoyi Liu¹, Leigh Nystrom¹, Mark Pinsk¹, Nicholas DePinto¹, William A. Wolf¹, Nathaniel D. Daw¹, Sabine Kastner¹, Jonathan D. Cohen¹, Harrison Ritz²; ¹Princeton University, ²Queen's University

Presenter: Xiaoyi Liu

Task switching requires flexible reconfiguration of task representations over time. To characterize the underlying neural computations, we recorded brain activity with optically-pumped magnetometer-based MEG (OPM-MEG) during a task-switching paradigm, in which participants maintained instructions for an upcoming task while performing a current task. In both human and recurrent neural networks (RNNs) trained on similar paradigms, we found that the current and future task identities were encoded in orthogonal dimensions. In human but not RNNs, the current-task representations generalize across epochs, consistent with a compositional representation organized by functional relevance.

Dynamic belief state representations in human orbitofrontal cortex under uncertainty

Talk 3, 10:35 am

Theo AJ Schäfer¹, Ondrej Zika², Tobias H. Donner³, Nicolas W. Schuck¹; ¹Universität Hamburg, ²University College Dublin, ³University Medical Center Hamburg-Eppendorf

Presenter: Theo AJ Schäfer

Adaptive behavior requires predicting the occurrence and timing of rewards. Some rewards, like annual celebrations, are precisely timed, while others, such as a decisive sports goal, are more uncertain. Because sensory input often provides ambiguous information about time, animals rely on probabilistic belief states to guide their expectations and learning from outcomes. Past fMRI work has implicated the orbitofrontal cortex (OFC) in the representation of hidden states, and animal work has suggested medial prefrontal cortex (MPFC) to be involved in temporal belief states. We investigated 36 human subjects with fMRI, eye-tracking, and computational modeling in a temporally dynamic belief state task. We provide evidence that belief states are encoded in OFC and expressed in subjective expectations, pupil dilation, and gaze behavior.

Shared and Distinct Neural Responses at Event Boundaries and Task Transitions

Talk 4, 10:45 am

Zhuoyang Li¹, Christopher Honey¹; ¹Johns Hopkins University

Presenter: Zhuoyang Li

Our mental lives are punctuated by transitions: we complete one chore and turn to the next, or we finish reading one chapter and begin the next. In what ways do our brains re-organize at such transitions? The two most studied forms of mental transitions — event segmentation and task-set switching — have been investigated independently. Here, we ask whether there are common neural processes operating at boundaries between narrative episodes and boundaries between goal-directed subtasks. We collected fMRI data from seven participants across three tasks: movie watching, word generation, and explanation generation. We observed boundary-locked BOLD increases in all three tasks both in the default mode network and in prefrontal regions. However, while the boundary-associated voxel patterns in default mode network regions generalized across all task pairs, the prefrontal boundary patterns were more task-selective. We conclude that the activity in the DMN likely reflects a context-updating process which is common to movie boundaries and subtask boundaries. Conversely, the more selective patterns in the PFC may track differences in control demands between movie viewing and active generation tasks.

Model-based error signals in the frontal cortex precede successful inference and hippocampal context updating during human latent-state inference

Talk 5, 10:55 am

Daniel Deng¹, Hristos Courellis¹, Ueli Rutishauser²; ¹California Institute of Technology, ²Cedars-Sinai Medical Center

Presenter: Daniel Deng

Adaptive behavior requires evaluating reasons for failure. A negative outcome may reflect suboptimal behavior or signal structured changes in the environment. Recording single-units from subjects performing latent-context inference, we asked how the brain distinguishes between these possibilities and how it updates internal models when failures suggest a change in latent state. We found that error-responsive neurons in the frontal cortex (FC), especially dorsal anterior cingulate cortex (dACC), distinguished errors signaling switch from those signaling genuine mistake based on unexpectedness, quantified with a novel reinforcement-learning (RL) model. This was true only in sessions where subjects successfully performed inference. On the other hand, hippocampal (HPC) neurons adapted their encoding of latent context to the new context only after successful switch behavior. These findings suggest that FC/dACC initiates contextual switching triggered by errors during inference, whereas HPC stably expresses context only after switching was successful.

From Impulsivity to Engagement: Latent Behavioral States Reveal Distinct Decision Strategies in Rats

Talk 6, 11:05 am

Aya Akhmetzhanova¹, Adam Goldring², Timothy Hanks², Rishidev Chaudhuri¹; ¹University of California, Davis, ²UC Davis

Presenter: Aya Akhmetzhanova

Behavioral variability in perceptual decision-making is often treated as noise, yet it may reflect meaningful fluctuations in internal cognitive state. We asked whether latent behavioral states could explain trial-to-trial variability in both behavior and neural population dynamics. In our work, we propose that latent behavioral states drive variability in perceptual decisions and simultaneously modulate neural dynamics across decision-related brain regions. To test this, rats were trained on an auditory decision-making task in which they reported increases in the click rate of a Poisson-generated stimulus. Fitting the Generalized Linear Model-Hidden Markov Model (GLM-HMM) model revealed two latent states that best captured behavioral variability across animals (Ashwood et al., 2021; Linderman et al., 2020). In an impulsive state, animals responded rapidly, producing frequent false alarms with little dependence on stimulus strength. In an engaged state, false alarm rates were markedly reduced, enabling stimulus-dependent correct detections. Large-scale Neuropixels recordings revealed no meaningful signature of behavioral state in mean firing rates or within-region correlations. Yet state identity was decoded from population activity trial-by-trial with high accuracy. Our findings provide a mechanistic link between internal engagement, neural population dynamics, and decision variability—challenging the assumption that engagement simply scales overall neural activity. Instead, latent behavioral states reflect discrete internal strategies encoded in population-level firing patterns.