Advances & challenges in foundation agents: Introduction to Chapter 2 – Cognition
This article is the Introduction to Chapter 2 of a series of articles featuring Liu and colleagues’ book Advances and Challenges in Foundation Agents: From Brain-Inspired Intelligence to Evolutionary, Collaborative, and Safe Systems.
Human cognition is a dynamic system of information processing that supports learning, memory, reasoning, and goal-directed behavior. At its core lies the concept of a mental state—a structured, internal representation encompassing beliefs, knowledge, context, and intent. This mental state provides the substrate over which cognitive operations act, enabling humans to flexibly adapt to new situations, abstract over experiences, and make context-sensitive decisions.
Decades of cognitive neuroscience have revealed a modular1 yet integrated architecture2 underlying these capabilities: perception systems that translate sensory input into internal symbols; memory systems that encode and retrieve experience; reasoning systems that formulate decisions; action systems that translate decisions into environmental interactions; reward signals that guide behavior through reinforcement; and emotion systems that modulate attention and resource allocation.
LLM-based agents offer a new computational paradigm that begins to approximate aspects of this architecture. These agents construct and manipulate internal mental states—though often implicitly—via large-scale hidden representations, memory buffers, or intermediate reasoning steps. Learning arises through gradient-based updates or contextual inferences, while reasoning often involves generating or selecting structured hypotheses, subgoals, or actions based on current context. Despite differences from biological systems, the underlying principles—modular operations on internal representations, guided by experience and adaptive goals—recur.
In this chapter, we first examine learning as the process by which agents improve or restructure their internal state. We then turn to reasoning, understood as the agent’s internal deliberation process that selects or constructs actions through search, generation, or inference. Figure 2.1 shows an overview of different learning and reasoning paradigms, which will be discussed in more detail in the following sections.
Next part: Section 2.1 – Learning.
Article source: Liu, B., Li, X., Zhang, J., Wang, J., He, T., Hong, S., … & Wu, C. (2025). Advances and challenges in foundation agents: From brain-inspired intelligence to evolutionary, collaborative, and safe systems. arXiv preprint arXiv:2504.01990. CC BY-NC-SA 4.0.
Header image: AI is Everywhere by Ariyana Ahmad & The Bigger Picture / Better Images of AI, CC BY 4.0.
References:
