Multi-agent system verification comprises formal techniques for modeling distributed and intelligent systems, expressing system properties, and verifying those properties. Agents within such systems may possess distinct profiles (e.g. personality traits, levels of experience, or physical characteristics) or may specialize during execution (e.g. through software or hardware updates, or restrictions from access control policies). These aspects are captured through the novel concept of capacity, which determines the subset of actions that an agent can perform, thereby constraining the agent’s behavior throughout the interaction. This article considers multi-agent systems with imperfect information regarding both states and actions, where capacity assignments are secret at system initialization. This general framework enables the modelling of a broad range of multi-agent systems; however, verification is undecidable in the general case. To address this challenge, we investigate memory constraints and structural restrictions. Our contributions are as follows: (i) we introduce Capacity Alternating-time Temporal Epistemic Logic (CapATEL), a logic for expressing strategic, temporal, and epistemic properties of imperfect information multi-agent systems with capacities, (ii) we analyze and compare distinct semantics that capture agents with fixed profiles or agents specializing, (iii) we study the model-checking problems depending on agent information and memory capabilities and we prove complexity bounds, and (iv) we illustrate the relevance of CapATEL in the active cybersecurity context.
Trajectory prediction and planning are critical for the safe operation of autonomous driving (AD) systems. Existing data-driven and imitation learning pipelines typically rely on geometric pattern fitting and short-horizon statistical correlations, struggling to generalize in complex, interactive scenarios. Although large language models (LLMs) can offer profound semantic understanding and causal reasoning, their massive computational overhead and mismatch with continuous trajectory generation severely hinder their development in real-time AD. In this paper, we propose an LLM-enhanced framework for joint trajectory prediction and planning that combines fast kinematic inference with adaptive semantic reasoning. Inspired by the free energy principle, we introduce a novel Hierarchical Contextual Encoder with an uncertainty-driven trigger. This module dynamically allocates language-model reasoning capacity based on scene complexity and uncertainty, utilizing a lightweight language model (LM) for routine scenarios, and selectively invoking a high-capacity LLM only when deep causal reasoning is required. To enhance the reasoning ability of the LM, we develop a CoT-based latent distillation paradigm that transfers structural reasoning and contextual priors from a teacher LLM to a student model through hidden-space alignment. In addition, we design a Cross-modal decoder that injects semantic intent into geometric trajectory generation, enabling more socially compliant and context-aware decision making. Extensive experiments on multiple real-world benchmarks demonstrate that LLM-TP2 achieves state-of-the-art performance in both prediction and closed-loop planning, while ensuring the computational efficiency required for practical deployment.
Meta-learning, which extracts meta-knowledge from historical tasks to facilitate learning new tasks, has achieved great success in various applications. Representative meta-learning algorithms like MAML assume tasks are similar and propose to learn a globally-shared meta-initialization for all tasks. However, real-world environments are usually complex, making task model parameters diverse and a single meta-initialization is insufficient to capture all the meta-knowledge. To deal with this challenge, we propose multi-subspace meta-learning in this paper. Task model parameters are structured into multiple subspaces, and each subspace represents one type of meta-knowledge. We propose two novel algorithms MUSML and PE-MUSML to learn the meta-parameters (i.e., subspace bases) and theoretically establish the generalization bound of the learned subspaces. Extensive experiments on regression and classification benchmark datasets demonstrate the effectiveness of subspace meta-learning.
Automated planning is a prominent area of Artificial Intelligence, and an important component for intelligent autonomous agents. A critical aspect of domain-independent planning is the domain model, that encodes a formal representation of domain knowledge needed to reason about a given problem. Despite the crucial role of domain models in automated planning, there is a lack of tools supporting the knowledge engineering process by comparing different versions of the models, in particular, determining and highlighting differences the models have.In this paper, we formalise the notions of structural equivalence and strong equivalence of domain models, and, on top of these, we introduce a novel concept of similarity of domain models. To measure the similarity of two models, we introduce a directed graph representation of lifted domain models that allows the formulation of the domain model similarity problem as a variant of the graph edit distance problem. We propose an Answer Set Programming approach to optimally solve the domain model similarity problem, that identifies the minimum number of modifications the models need to become strongly equivalent, and we demonstrate the capabilities of the approach on a range of benchmark models.
Successful human-agent teaming relies on an agent being able to understand instructions given by a (human) principal. In many cases, an instruction may be incomplete or ambiguous. In such cases, the agent must infer the unspoken intentions from their shared context, that is, it must exercise the principal’s Theory of Mind (ToM) and infer the mental states of its principal. We consider the prospects of effective human-agent collaboration using large language models (LLMs). To assess ToM in a dynamic, goal-oriented, and collaborative environment, we introduce a novel task, Instruction Inference, in which an agent assists a principal in reaching a goal by interpreting incomplete or ambiguous instructions.We present Tomcat, an LLM-based agent, designed to exhibit ToM reasoning in interpreting and responding to the principal’s instructions. We implemented two variants of Tomcat. One, dubbed Fs-CoT (Fs for few-shot, CoT for chain-of-thought), is based on a small number of examples demonstrating the requisite structured reasoning. One, dubbed CP (commonsense prompt), relies on commonsense knowledge and information about the problem. We realized both variants of Tomcat on three leading LLMs, namely, GPT-4o, DeepSeek-R1, and Gemma-3-27B. To evaluate the effectiveness of Tomcat, we conducted a study with 52 human participants in which we provided participants with the same information as the CP variant. We computed intent accuracy, action optimality, and planning optimality to measure the ToM capabilities of Tomcat and our study participants. We found that Tomcat with Fs-CoT, particularly with GPT-4o and DeepSeek-R1, achieves performance comparable to the human participants, underscoring its ToM potential for human-agent collaboration.
