Proceedings

Humans lie and detect lies. Deceivers deceive successfully by inferring whether a victim will detect them, and victims detect deception by inferring how a sender might lie. Deception and counter-deception require a set of cognitive traits: the ability to infer what others are thinking, the capacity to reason how likely others will accept an unfair outcome, and other qualities. These behaviors have been formalized as recursive social reasoning. While deceptive dynamics have been tested and recorded in humans, should we expect modern AI agents to follow the same rules and conventions that guide human behavior, or can we postulate otherwise? In this work, we examine whether modern AI agents (LLMs) replicate complex deceptive and counter-deceptive dynamics, and why they sometimes fail to do so. We will further propose a neuro-symbolic (model synthesis) architecture to endow LLMs with missing social inference capacities.

Current LLM agents are commonly organized around a unitary objective, supplemented by memory, critics, planning, or temporary subagents. This position paper proposes a different organization: one persistent agentic identity containing multiple persistent, specialized cognitive Parts that can influence, but cannot independently actuate, the agent. The design draws inspiration from Jungian complexes, Internal Family Systems (IFS), Minsky's Society of Mind, Global Workspace Theory, and predictive-processing accounts. Emotion is used only as an inferred salience signal: unusually significant events may create or reactivate narrow, intentionally biased Parts. Activated Parts bid for influence, make predictions, and revise local beliefs from later outcomes, while external action remains centralized in a single executive identity. We hypothesize that this architecture can preserve important minority concerns, support long-horizon relational learning, increase solution diversity, and make internal motivational conflict more inspectable than either unitary agents or disposable subagent swarms.

Trustworthy knowledge-based AI needs extensive amounts of explicit knowledge. Manual acquisition of such knowledge is often considered too expensive to be practical at scale. We present a learning approach that uses large language models (LLMs) to drastically lower the cost of knowledge acquisition for knowledge-based AI. The pipeline outsources cognitively costly steps to LLMs while reserving the final validation of new knowledge for human experts. Our empirical results show that this learning approach produces quality knowledge at practical scales, and suggest a trend of quality improvement over time.

Scripts, also called complex events, are events with component subevents. Ontologically grounded agents use scripts to encode procedural knowledge. Scripts which represent mental procedures (agenda management, learning, etc.) are called mental scripts. Object scripts encode knowledge about procedures which involve interaction with an agent's physical environment. We present recent advances in the methods that ontologically grounded agents can use to learn object scripts during operation. Specifically, we outline a new mental script which allows agents to learn task-relevant object scripts through natural language dialogue. We also present a demonstration of this mental script in which an agent assisting in a ship's engine room learns to replace a fuse in real time.

In our system, a Soar cognitive architecture instance acts as a cognitive control unit: it maintains an explicit task-decomposition model of a robotic sanding application and invokes neural models - currently a fine-tuned object-detection and pose-estimation pipeline - only as tools for the perception subtasks they perform best. The Soar-ROS 2 bridge, soar_ros, is open source and separately evaluated. The current work-in-progress version is a semi autonomous agent gated by human-in-the-loop validation steps before irreversible sanding actions. A planned lab study will add an explanation-chat interface to the same system and measure its effect on shop-floor workers' attitude towards robots under controlled failure conditions. Post experiment interviews will explore the perceived quality of the explanations and use of the system.

We present a responsibility contract for language-model proposals against stale inventory observations: the adapter owns modeled facts, the generator proposes, and a symbolic controller admits and commits atomically. Under a sound closed adapter, accepted transfers preserve encoded invariants; intent correctness and physical safety remain outside this guarantee. A reproducible CPU fixture checks the implementation and exposes conservative refusals. This is a design and comparison agenda, not a new validation algorithm or a small-model performance result.

Autonomous agents increasingly act under delegated authority, yet conventional authorization mechanisms represent credential validity more readily than the continuing grounds that make an action appropriate after the agent's world model changes. This paper frames authority maintenance as a cognitive task of meta-level control. It proposes a temporal ontology in which each grant depends on explicit justificatory propositions whose epistemic status, freshness, and revalidation conditions are represented symbolically. World-model updates can therefore change actionability while the stored permission remains unchanged. Neural components may interpret documents, dialogue, or perceptual inputs and propose belief updates, while the symbolic layer validates those updates and retains execution control. Five epistemic-control states distinguish supported, stale, contradicted, suspended, and revoked authority. A comparative evaluation design tests whether this representation improves safe execution, avoids unnecessary interruption, and produces explanations grounded in the conditions that actually governed action. The resulting framework connects world modelling, metacognitive monitoring, and high-level execution within a concrete challenge problem for symbolically orchestrated agents.

In the 2026 BEHAVIOR Challenge, a policy receives RGB, depth, and proprioception, while the task evaluator can inspect symbolic simulator state that the policy cannot access. The evaluator can therefore mark a household task complete before the robot has evidence that supports the same claim. We ask when an embodied agent has enough evidence to say DONE.

We distinguish oracle completion from warranted completion. A claim for a task is warranted only when its symbolic goal is true in every world history compatible with the robot's evidence. If compatible worlds disagree, the verdict remains UNKNOWN. A missing observation cannot establish a negative fact, and a final scene cannot establish facts that depend on hidden state or earlier events.

Robotic systems increasingly combine separately trained encoders with symbolic planners. The encoders may use different latent coordinates, so the planner must know whether shared observations suffice for its own readouts. Let Q_* be the unknown rotation or reflection from source to target coordinates. If X and Y collect the source and target encodings of shared observations, then Y = Q_*X + E. Let U span the source-space directions read by the task. In zero noise, every orthogonal map fitting the anchors gives the same task readouts if and only if U ⊆ span(X); otherwise a reflection fixes all anchors but changes a task direction. We also give a directional noisy bound and an idealized resource law: if k = dim U < r, then k chosen exact anchors are necessary and sufficient, whereas r generic passive anchors are needed almost surely. This yields an auditable accept/abstain/request-more interface. A preliminary non-robot test shows that orthogonality is only approximate.