Telic States: Theory and Applications

• Telic states are defined as event representations characterized by inherent endpoints, distinguishing bounded (telic) from unbounded (atelic) events.
• They are modeled through formal semantics and empirical approaches, with methods like distributional embedding and regression analysis revealing key linguistic cues.
• Telic states extend to multimodal affordance and computational representations, advancing research in NLP, cognitive science, and reinforcement learning.

A telic state is a formally defined linguistic, cognitive, or computational construct denoting an event, experience, or state representation that is characterized by the specification or attainment of a natural endpoint, goal, or conventional function. The notion of telicity originates in the theory of lexical aspect, distinguishing between predicates or events that comprise an internal, inherent endpoint (telic) and those that do not (atelic). Recent computational and theoretical research has extended this notion to domains including formal semantics, distributional LLMing, multimodal grounded semantics, embodied simulation, sign linguistics, and goal-directed state representation learning.

1. Telic States in Aspectual and Formal Semantics

Telic states are rooted in lexical aspect theory as events (typically encoded by verb phrases) whose very meaning entails a clear endpoint or culmination. Formally, a telic predicate encodes a bounded event, leading to completion, while atelic predicates denote ongoing or unbounded eventualities without such a culmination (Kober et al., 2020, Kovalev et al., 8 Jun 2025). In the domain of compositional semantics, especially as formalized in dependent type theory, the telicity of an event is directly linked to the boundedness of its undergoer (object): events with a bounded undergoer (e.g., “three apples” as the object of “eat”) are modeled as telic, while those with an unbounded undergoer (“apples”) are atelic. The dependent event calculus presented in (Kovalev et al., 8 Jun 2025) formalizes this connection, defining for actors (Act) and undergoers (Und^b, with b ∈ {B,U}) the event type Evt^b(a, und), where b indicates boundedness—and hence telicity in the verbal domain.

Culminativity is an additional property of telic events, indicating not only that an event has a natural endpoint but also that this endpoint is attained, yielding a resulting state. The framework in (Kovalev et al., 8 Jun 2025) provides formal machinery—via the function Result(evt) and the proposition isCul(evt)—to encode the entailment from completed telic events to their resultant states. For instance, a culminating event “John ate three apples” entails the resulting state “three apples are eaten”.

2. Distributional and Contextual Modeling of Telic States

Telicity, while semantically compositional, is also detectable using distributional semantic modeling. Embedding-based approaches, such as those described in (Kober et al., 2020), leverage pre-trained word representations (e.g., from skip-gram or word2vec models) for verbs and compose these with embeddings of nearby context words to produce composite representations:

$x' = x + \sum_{c \in C} c$

where $x$ is the verb embedding and $C$ the set of contextual word embeddings. Logistic regression over $x'$ can reliably distinguish telic from atelic events. The strongest indicators are found in tightly constrained local contexts, especially through closed-class words (prepositions, particles, determiners) that encode grammatical and relational cues signaling goal-directed or resultative constructions.

Empirical evaluation with conversational datasets annotated for aspectual class demonstrates that models using such local context windows and focusing on closed-class markers outperform those relying solely on verb or sentence-level embeddings. Furthermore, telicity correlates not only with lexical content but also with discourse genre and communicative goals: dialog genres characterized by instructions, directions, or goal-driven discourse contain higher frequencies of telic expressions.

3. Telic Affordance in Multimodal and Embodied Frameworks

The concept of telicity extends into multimodal semantic representation and embodied agent-environment interaction via the notion of telic affordance. Distinct from Gibsonian affordances, which capture low-level physical interaction potential based on object structure, telic affordances encode the conventional, intended function of an object in a specific habitat or context (Lee et al., 2023, Chen et al., 2023). For example, a glass affords use as a container only when oriented correctly.

In formal terms, the telic affordance structure $TA(o)$ for object $o$ is defined as the set:

$$TA(o) = \{ f \mid f\!: \text{Object} \to \text{Function} \wedge \text{Habitat}(o) \models C_f \}$$

where $f$ is the intended function and $C_f$ the contextual condition for activation.

Research using annotated multimodal datasets confirms that scenes exhibiting telic affordance elicit less generic and more function-specific linguistic descriptions in image captions. Statistical modeling (e.g., multiple linear regression) quantifies this effect, showing a negative coefficient for telic affordance with respect to the use of generic “holding-verbs” in captions, indicating that telic affordances sharpen the observer’s linguistic focus towards the object’s conventionally assigned purpose. Multimodal cues—including perceptual salience, object number, gaze, and ecological niche association—jointly determine when and how telic affordances are realized in grounded language.

In embodied simulation environments such as VoxWorld, telic states govern agent-object interaction: an agent can only activate and use an object’s telic affordance if the context (habitat) is appropriate. This enables dynamic inference and goal-driven adaptation in human-agent collaboration, as well as enhancing the disambiguation and contextual appropriateness of linguistic understanding in simulated or robotic systems.

4. Kinematic and Cross-Modal Expression of Telicity

In sign languages, telic and atelic distinctions are physically realized in the kinematic properties of sign production (Krebs et al., 8 May 2024). Motion capture in Austrian Sign Language reveals that telic verbs (denoting events with natural endpoints) are produced with higher peak velocity and shorter duration than atelic verbs (ongoing or non-finalized events). Linear mixed-effects models, controlling for participant and item variability, demonstrate statistically significant effects for verb type on both velocity and duration. This physically grounded modulation of movement provides evidence for the cross-modal universality of telic state encoding and has implications for sign language technology—including recognition, avatar animation, and pedagogy—where accurate modeling of aspectual distinctions is contingent on capturing these kinematic markers.

5. Goal-Directed State Representation and Computational Telic States

Recent computational frameworks generalize telic states beyond language and object interaction to the broader domain of goal-directed learning and decision-making (Amir et al., 20 Jun 2024, Amir et al., 20 Aug 2025). Here, telic states are defined as equivalence classes of experience distributions, partitioned according to a goal-dependent preference relation. For a goal $g$, telic state representation is given by the quotient:

$\mathcal{S}_g = \Delta(\mathcal{H}) / \sim_g$

where $\Delta(\mathcal{H})$ is the simplex of probability distributions over possible experience histories, and $\sim_g$ identifies distributions equivalent under the agent’s goal preference.

This formalism unifies descriptive aspects (state modeling of the world) and prescriptive/evaluative aspects (goal, reward), positing that both co-emerge from goal-directed interaction. The agent’s current policy $\pi$ induces a distribution $P_\pi$ over experience, and convergence towards a telic state $S$ can be measured by the KL divergence:

$R = \min_{P \in S} D_\text{KL}(P \Vert P_\pi)$

Algorithms for telic-controllable representations further refine state partitions so each telic state can be reached within bounded policy complexity. This structure supports flexibility in shifting goals while managing policy complexity and provides a mechanism—via deliberate abstraction from goal-irrelevant distinctions—for scalable representation learning. Empirical and computational evidence underscores the relevance of this approach for reinforcement learning, robotics, cognitive neuroscience, and behavioral modeling, providing a formal and unified perspective on purpose-driven cognition and AI.

6. Implications, Applications, and Theoretical Significance

Identifying and modeling telic states have consequences across multiple domains:

• Natural Language Processing: Accurate classification of telic vs. atelic events enhances event ordering, textual entailment, temporal reasoning, and discourse coherence models (Kober et al., 2020).
• Cognitive Science and Neuroscience: Telic state concepts yield testable predictions regarding goal-sensitive neural coding (e.g., hippocampal mapping, orbitofrontal representations) and flexible learning strategies, offering an explanatory mechanism for behavioral adaptation and cognitive map reorganization (Amir et al., 20 Aug 2025).
• Multimodal Reasoning and Embodied Simulation: Telic affordances and conditional state activation improve the contextual grounding and interpretability of AI systems in collaborative, situated, and multimodal environments (Lee et al., 2023, Chen et al., 2023).
• Computational Learning: Telic-controllable state representations facilitate efficient exploration, intrinsic motivation, and transfer learning in artificial agents, by centering learning and abstraction on goal-relevant equivalence structures (Amir et al., 20 Jun 2024).
• Formal Semantics: The type-theoretic analysis of telicity within Agda formalizations demonstrates precision and inferential clarity, supporting cross-linguistic and computational investigations with explicit, computer-verified reasoning (Kovalev et al., 8 Jun 2025).

This body of work demonstrates that telic states constitute a unifying concept spanning linguistic, cognitive, and computational domains, centering productive research on the principled integration of endpoints, goal-induced structure, and context-sensitive abstraction in the representation and processing of events and experiences.