Architectural Extensions for Embodiment and Emotion

Emotion is acknowledged as a major factor in human thought and behavior, but is often omitted from large scale models of cognition. Instantiating emotion poses several problems, such as pinpointing the origin of emotion information, constructing a representation, and defining how emotion interacts with the rest of the model components. In this paper, we examine this third challenge using the Common Model of Cognition (CMC), a high level architecture that provides a good account of human brain activity in a variety of tasks and at rest. The original CMC does not include emotion, but a recently proposed expansion (Rosenbloom et al, 2024) has laid out multiple accounts of how an emotion module might be incorporated into the framework. We compare three accounts of emotion interaction using the fMRI data of subjects from the Human Connectome Project (HCP) dataset, assuming the presence or absence of an emotion signal based on task events. We find that the base account, where emotional content affects the brain network only through the perceptual module, does not differ significantly from an account where emotion signals are allowed to affect all network modules (the direct account), an account where emotion signals have a modulatory effect on connections within the network (the modulatory account), or both (the combined account). This suggests that emotional signals cannot be assumed directly from emotional task events, and more precise methods for extracting and representing emotional signals will be required in order to further explore how emotion interacts with whole-brain models of cognition.

This is an in-person presentation on July 19, 2026 (10:40 ~ 11:00 EDT).

The interplay between cognitive and motor processes is central to skilled performance, yet most cognitive architectures treat motor execution as a downstream consequence of cognitive decisions rather than an active contributor. This paper explores how this limitation could be overcome in ACT-R. The task we investigate is figured bass improvisation, a baroque-era musical skill requiring real-time harmonic realization from symbolic bass cues. Although rule-based, the task is also shaped by motor expertise: experienced pianists learn it faster than novices, suggesting that ideomotor principles play a role. We present two ACT-R models simulating figured bass improvisation. The schema-based model processes no motor information. To address this, we developed a memory-based model, which uses anticipated fingerings as retrieval cues for chords, operationalizing the ideomotor principle within ACT-R. This approach faced two challenges: First, the transition graph of fingerings, which we constructed from observing a musician playing six exercises, is cyclic, making it problematic to model anticipation using standard probabilistic models. The predictions obtained from the graph produced some misleading anticipations. Second, the motor cues are often too weak to compete with large differences in baselevel activation, which are due to different frequencies of chords. Our results highlight a gap in the ability of ACT-R to model tasks in which motor knowledge shapes cognition and suggest that associative motor-cognitive interactions require architectural extensions. We discuss potential solutions and argue that figured bass improvisation serves as a compelling testbed for theories of embodied cognition.

This is an in-person presentation on July 19, 2026 (11:00 ~ 11:20 EDT).