TLDR: A new research paper proposes that our understanding of the world (state representation) and what we desire (reward function) don’t exist separately but co-emerge from our goals. This “telic states” framework, inspired by Buddhist philosophy, defines states as groups of experiences that are equivalent in achieving a goal. It suggests that learning involves aligning behavior with these goal-defined states, offering new insights for AI and neuroscience, and providing a way to understand how goals themselves are formed.
In the fascinating world of artificial intelligence and cognitive science, understanding how intelligent beings pursue goals is a central challenge. Traditionally, models of purposeful behavior, like those in reinforcement learning (RL), assume that an agent’s understanding of the world (its “state representation”) and what it desires (its “reward function”) are distinct and pre-defined. However, a groundbreaking new paper titled “Goals and the Structure of Experience” by Nadav Amir, Stas Tiomkin, and Angela Langdon proposes a radical alternative: these two fundamental aspects of a world model don’t exist independently but rather emerge together, driven by an agent’s goals.
The researchers introduce a computational framework where the descriptive (what is) and prescriptive (what is desirable) elements of a world model co-emerge from the agent’s interactions with its environment, which they call “experiences.” Drawing inspiration from Buddhist epistemology, particularly the philosopher Dharmakīrti’s ideas on concept formation, the paper defines “telic states.” The word “telic” comes from the Greek “telos,” meaning ultimate end, emphasizing the goal-oriented nature of these states. A telic state is essentially a class of “goal-equivalent experience distributions.” This means that if different sequences of actions and observations lead to the same desired outcome, they are grouped into the same telic state, regardless of their superficial differences. For instance, if a hungry rat has multiple paths to reach food, all those paths, if equally likely to lead to food, would be considered part of the same telic state.
This novel perspective offers a more parsimonious way to understand goal-directed learning. Instead of an agent trying to maximize a pre-set reward, learning becomes a process of minimizing the statistical difference between its current behavioral patterns (policies) and the features of desirable experiences encapsulated within telic states. This framework suggests that what an agent perceives and how it categorizes its experiences are not fixed, objective truths but are dynamically shaped by its current goals. For example, when crossing a busy road, your “state representation” focuses on traffic and pedestrian signals. But when hailing a taxi, it shifts to vehicle types and availability. This dynamic, goal-dependent representation is a core tenet of the telic states framework.
Implications for Artificial Intelligence and Neuroscience
The telic states framework has significant implications for both artificial intelligence and our understanding of the brain. In AI, it offers a fresh approach to long-standing problems such as transfer learning, where knowledge from one task is applied to another; multi-objective reinforcement learning, where agents pursue multiple goals simultaneously; and state abstraction, simplifying complex environments. By aligning an agent’s preferences and experiences with its state representation in a goal-flexible way, telic states can provide a unified framework for addressing these challenges. It suggests that instead of inferring hidden states, agents construct them based on their goals, distinguishing between goal-relevant and goal-irrelevant information.
For neuroscience, the framework aligns with recent findings that brain regions traditionally associated with “cognitive maps” (like the hippocampus and prefrontal cortex) exhibit dynamic, goal-dependent representations. For example, hippocampal place cells, once thought to only encode spatial location, are now known to show preferential representation for goal locations and encode goal-oriented aspects of experience. The paper predicts that neural signatures of a cognitive map would reflect an intrinsic preference among sensorimotor sequences, suggesting that brain regions might jointly encode both state and value information, rather than separately.
Also Read:
- Unlocking Spike-Driven Learning: How Synaptic Bundles and Motor Neuron Counts Impact AI Systems
- Unveiling the Statistical Link: How Preference Models Connect to Proportional Hazards
The Origin of Goals: A New Perspective
One of the most profound questions in goal-directed learning is: where do goals come from? Traditional value-based approaches often face a circularity problem: if goals determine value, how can value be used to select goals? The telic states framework offers a potential way out by introducing the concept of “telic-controllability.” This property relates to how easily an agent can reach all possible telic states given its computational limitations. It allows for goal selection to be regulated by the properties of the induced telic state representation, suggesting a trade-off between the granularity of the representation and the computational resources needed to utilize it. This also sheds light on “transformative experiences”—events that fundamentally change an individual’s preferences and goals—by viewing them through the lens of the new telic state representation they induce.
In essence, “Goals and the Structure of Experience” invites a paradigm shift from an “agent-centric” view, where a fixed agent is the central controller, to a “goal-centric” one, where subjective experience is dynamically shaped by goal-dependent representational structures. This perspective, detailed further in the full paper available at arXiv:2508.15013, provides a unified foundation for studying diverse forms of intelligence, both natural and artificial, by grounding learning in the statistical structure of subjectively structured experience streams, without necessarily invoking the notion of fixed and enduring agents.
