A Taxonomy of Cognitive Capability Gaps in Generative and Agentic AI
A taxonomy-driven survey published on arXiv identifies five dimensions of cognitive capability gaps in generative and agentic AI: persistent state modeling, goal-directed autonomy, self-monitoring and control, environment interaction, and learning and adaptation. The paper proposes an Adaptive Cognitive Intelligence Architecture (ACIA) and discusses cognition-centric evaluation.
The paper 'A Taxonomy of Cognitive Capability Gaps in Generative and Agentic AI' surveys limitations in current AI systems across five cognitive dimensions. It reviews recent advances, identifies recurring limitations, and outlines open research challenges. The authors propose a conceptual Adaptive Cognitive Intelligence Architecture (ACIA) and examine emerging directions in cognition-centric evaluation, providing a unified framework for organizing research and guiding future cognitive AI development.
The taxonomy highlights that current generative and agentic AI systems lack sustained reasoning, adaptive behavior, persistent memory, and self-regulation. The five dimensions—persistent state modeling, goal-directed autonomy, self-monitoring and control, environment interaction, and learning and adaptation—represent fundamental gaps. The proposed ACIA architecture suggests a path toward integrating these capabilities, while cognition-centric evaluation may shift benchmarking from task performance to cognitive function assessment.
This research signals a growing recognition that commercial AI systems need more than language generation and task execution to operate reliably over extended time horizons. The identified gaps may influence future product roadmaps for AI platforms and agentic frameworks, as enterprises demand more robust, self-regulating AI. The taxonomy could become a reference for prioritizing R&D investments in cognitive capabilities.
Addressing these cognitive gaps could enable AI systems that maintain context over long interactions, autonomously pursue complex goals, and self-correct errors, reducing human oversight costs and expanding AI applicability in enterprise workflows, autonomous systems, and personal assistants. The taxonomy provides a structured framework for assessing and communicating AI system capabilities to stakeholders.
Next signals to watch include follow-up papers implementing or evaluating the ACIA architecture, adoption of the taxonomy in AI benchmarking suites, and industry announcements of cognitive AI features (e.g., persistent memory, self-monitoring) in commercial products. Research funding calls or collaborations focused on cognitive AI may also emerge.