An Epistemic Perspective on Agent Awareness

AI summary

Overview

Research area: Artificial Intelligence, specifically epistemic logic and formal models of agent awareness (knowledge representation and reasoning).

Technical level: Advanced. The paper is a contribution to mathematical modal logic: it defines a new formal language and semantics, states an axiomatisation, and proves soundness and completeness using a set-theoretic "matrix"/frame construction over ordinals. Readers will need comfort with modal logic, Kripke-style semantics, and consistency/extension proofs.

Scope (one sentence): The paper proposes that "agent awareness" be treated as a form of knowledge, introduces separate modalities for de re and de dicto awareness, and proves a sound and complete logical system describing how those modalities interact with a standard knowledge modality.

What This Paper Is About

Existing literature on awareness treats awareness as a concept distinct from knowledge, often splitting it into awareness of objects and conceptual awareness. The authors instead adopt the dictionary reading of awareness as "knowledge that something exists" and give it a formal epistemic treatment.

Working out that treatment, they observe that one agent can be aware of another in two different ways: aware of a physical object (de re) or aware of someone matching a description (de dicto). The paper's goal is to build a logical language with distinct modalities for these two forms and to axiomatise how they interact with ordinary knowledge.

Key Contributions

1. An epistemic reinterpretation of agent awareness. The paper breaks with the tradition in the awareness literature by treating awareness of another agent as a form of knowledge rather than as a separate primitive concept, drawing on the Cambridge Dictionary definition of awareness as "knowledge that something exists."

2. Two new modalities for de re and de dicto awareness. The language Φ extends a knowledge modality K (read as "knows φ about herself") with R ("de re aware about someone with property φ") and D ("de dicto aware about someone with property φ"). Their meanings are specified using a version of 2D-semantics via a ternary satisfaction relation w, a ⊩ φ between a world, an agent, and a formula.

3. A sound and complete axiomatisation. The system includes the standard epistemic axioms (Truth, Negative Introspection, Distributivity) plus Self-Awareness, Introspection of Awareness, Unawareness of Falsehood, Disjunctivity, and General Awareness, with Modus Ponens, Necessitation, and two Monotonicity rules. Soundness is Theorem 1; completeness is established for the full system.

4. A novel adaptation of the "matrix" completeness technique. The proof adapts the recently proposed matrix technique of Naumov and Tao (2023) — matrices whose rows are worlds and columns are agents — by adding an explicit awareness relation, row labels λ_w, and the requirement that each cell set be λ_w-assured. The authors call these augmented matrices "frames."

Main Findings

• Soundness. Theorem 1 states that if a formula is a theorem of the system (⊢ φ), then it is satisfied at every world and by every agent present there in every epistemic model. The paper notes soundness is straightforward for most axioms and proves the General Awareness axiom separately as Lemma 2.

• Completeness. The system is shown to be sound and complete for the class of epistemic models defined in the paper — the main technical result claimed in the abstract.

• De re awareness defined as physical-object awareness. By item 5 of the semantics, agent a is de re aware of φ if there is some agent b present in the current world with property φ, and b is present in every world indistinguishable by a from the current world. The running example has Ann de re aware of an unmarked police vehicle outside the museum, but not de re aware of the WeRide vehicle (the actual-world WeRide agent b is absent from world w₂).

• De dicto awareness defined as description-based awareness. By item 6, agent a is de dicto aware of φ if every world indistinguishable by a contains at least one agent with property φ. In the example, Ann is de dicto aware of a WeRide near the museum because a WeRide exists in both w₁ and w₂ — even though she cannot see the vehicle.

• The λ-assured set notion. A set X of formulae is λ-assured if X ⊬ Aⁿ ¬λ for each n ≥ 0, where A φ abbreviates R φ ∨ D φ ("generally aware of φ"). Intuitively, this states that only the existence of absolutely undetectable "ghost" agents with property ¬λ is consistent with X, so all detectable agents must have property λ. This notion is what makes the frame construction work.

• Why the frame machinery is needed. The paper explains that defining awareness in a partially constructed model in the usual way could cause an agent to become "unaware" of another agent once a new world is added. To prevent this "loss of awareness effect," frames carry an explicit reflexive awareness relation ↝_w, connected to the syntactic modalities by item 5(b) of the frame definition.

• Requirement-based completion. Five types of "requirements" corresponding to the five clauses of the completeness definition are enumerated; any finite frame with an active unfulfilled requirement can be extended to a finite frame where it is fulfilled (Lemma 8), and any finite frame can be extended to a complete frame (Lemma 11), so the limit of the resulting chain is complete.

• No empirical results. The paper contains no experiments, benchmarks, datasets, model evaluations, or implementations. All claims are mathematical.

Methodology in Plain English

The authors start from a modelling choice: treat an agent's awareness of another agent as something the agent knows, not as a separate mental state. They set up "epistemic models" containing a set of worlds, a set of agents, a presence relation (which agents exist in which worlds), an indistinguishability relation for each agent, and a valuation for propositional variables.

They then define what it means for a formula to be satisfied by a specific agent at a specific world — a ternary relation rather than the usual world-only relation. This is the "egocentric" style of Prior (1968), combined with the two-dimensional semantics that lets a formula be evaluated relative to a world and an agent at once. In that setting, knowing φ about yourself means φ holds in all worlds you cannot distinguish; de re awareness means there is a specific agent who has the property now and exists in all such worlds; de dicto awareness means every such world contains at least one agent with the property.

They write down axioms intended to capture intuitions about these notions, discuss each informally, and then prove soundness directly against the semantics. For completeness, the usual "canonical model" approach does not fit 2D-semantics because a single maximal consistent set must be split into a world and an agent. They solve this with the matrix technique: build an infinite table whose rows are worlds and columns are agents, fill each cell with a maximal consistent set of formulae, and add an explicit awareness relation on columns together with per-row labels to keep awareness stable as the table grows. Starting from a finite frame, they enumerate all outstanding "requirements" and extend the frame step by step; the limit of this chain is a complete frame from which a satisfying model is obtained.

Why This Matters

Impact on research. The paper adds to a body of work on specialised forms of knowledge — know-how, know-who, know-whether, know-why, and know value — by proposing know-awareness (in de re and de dicto forms) as another such form, and provides a quantifier-free modal system where most earlier treatments of the de re/de dicto distinction rely on quantifiers. It also complements Epistemic Logic with Assignments and the de re/de dicto "know who" modalities of Epstein, Naumov, and Tao (2023), and the dataset-based de re/de dicto knowledge system of Jiang and Naumov (2025), none of which the authors say can express either form of awareness.

Real-world applications (as motivated in the paper):

• A war robot that must minimise casualties when it is aware of civilians present at the theatre of operations.
• A self-driving car that must stop at a yield (give way) sign when aware of an approaching vehicle, or stop for an approaching emergency vehicle.
• A medical doctor who must offer help when aware of someone being sick.
• A machine whose values align with humans' that must apologise when aware of someone being offended, or a well-mannered robot that should not take the last piece of cake if aware of somebody else wanting it.

Industry relevance. The de re/de dicto split maps onto engineering distinctions: a robot series i17 must introduce itself if a guard is aware of its presence as a physical object (a de re reading — otherwise the introduction would be pointless if the guard already knew the series), while a guard must report any sighting of an i17, a de dicto reading. These distinctions matter for specifying autonomous system behaviour and multi-agent protocols.

Future Directions

• Extending the language beyond the three modalities K, R, and D — for instance to combinations involving iteration of de re and de dicto awareness — is not addressed in the truncated content.
• The paper notes that a positive-introspection-style axiom for the de re modality R is "generally speaking, not valid," leaving open what weaker introspection principles for R might be sound and complete.
• Generalising the completeness construction beyond the specific class of epistemic models defined here (for example, dropping or weakening the assumption of given transworld identity, which the authors flag as a contested subject in the philosophy of language) is a natural next question.
• Connecting the abstract system to concrete computational settings — the authors note that existing systems based on Epistemic Logic with Assignments and dataset-inferred de re/de dicto knowledge cannot express awareness, suggesting integration work remains open.

Target Audience

Researchers in epistemic and modal logic, multi-agent systems, and knowledge representation who work on awareness, de re/de dicto distinctions, or completeness proofs using canonical-model techniques. It will also interest AI safety and autonomous-systems researchers who need formal languages for specifying what a machine must know about whom is present before acting. The paper assumes substantial background: general readers without training in modal logic will find the axiomatisation and the frame construction difficult, though the running "museum" example is designed to make the two awareness modalities intuitive.

Note: The paper content provided is truncated, so the detailed completeness proof and appendices are only partially visible in the source above; the summary reflects what is present.

Authors’ abstract

The paper proposes to treat agent awareness as a form of knowledge, breaking the tradition in the existing literature on awareness. It distinguishes the de re and de dicto forms of such knowledge. The work introduces two modalities capturing these forms and formally specifies their meaning using a version of 2D-semantics. The main technical result is a sound and complete logical system describing the interplay between the two proposed modalities and the standard "knowledge of the fact" modality.