ABSTRACT. Knowledge graphs (KGs) are increasingly used in autonomous driving (AD) to make traffic scene entities and relations explicit for prediction and reasoning. Existing traffic-scene KGs, however, primarily emphasize trajectory forecasting or general scene understanding. This paper presents the nuScenes Intention Knowledge Graph (NIKG), a focused Web Ontology Language (OWL)-based semantic representation and construction pipeline for ego-centric vehicle intention context. NIKG instantiates nuScenes scenes as a self-describing Turtle graph and promotes three intention-relevant relations to first-class properties of each scene participant: ego-relative motion, spatial configuration, and time to collision (TTC) risk. The pipeline uses a small input contract available in many ego-centric driving datasets: ego pose, tracked object annotations, and temporal keyframe order. NIKG is released as an ontology-backed, self-describing graph artifact together with a reproducible construction pipeline and validation protocol for ego-centric intention context. Applying this pipeline to nuScenes mini yields a graph of 211,335 triples across 404 keyframes; under the validation protocol, ego-relative motion labels achieve 92.5% forward consistency. Coverage, logical consistency, and temporal coherence analyses over this artifact are further reported. NIKG occupies the space between scene-oriented traffic KGs and intention-prediction models: a small, explicit semantic layer intended to support downstream work such as knowledge graph embedding (KGE), SPARQL Protocol and RDF Query Language (SPARQL)-based feature extraction, or hybrid neuro-symbolic reasoning.

ABSTRACT. Large Language Models (LLMs) perform well on general NLP tasks but often underperform in domain-specific settings where proprietary knowledge and concept hierarchies are required. Chain of Concepts (CoC) mitigates this by injecting structured conceptual guidance via manually curated graphs, but expert curation is costly and hard to scale or reproduce. We propose Automatic Chain of Concepts (AutoCoC), a fully automated pipeline that (i) induces a rooted concept tree directly from domain documentation, (ii) enriches nodes with representative examples for in-context learning, and (iii) linearizes the tree into a pedagogically ordered prompt sequence via breadth-first traversal. We evaluate AutoCoC on OWL ontology generation (Turtle) across 10 domains, comparing against handcrafted CoC and reverse-engineered baselines. AutoCoC achieves the best mean composite score (0.832) and produces substantially richer ontologies (21.7 classes/file vs. 13.5 and 14.8). Semantic constraint compliance (OntoClean-inspired) is marginally higher (90.6% vs. 89.4%), while confidence is competitive (0.893 vs. 0.864), though per-domain confidence remains higher for the handcrafted baseline in several domains. A self-consistency analysis over 200 runs shows high concept-level stability (0.889) but moderate structural variability (0.625). These results suggest that automated concept-structure induction is a viable way to scale concept-based prompting for structured, domain-specific generation tasks without relying on expert-authored hierarchies. Overall, AutoCoC contributes a reproducible and scalable alternative to expert-authored concept hierarchies by automatically deriving domain-specific prompt structures from documentation, making concept-based prompting more practical for structured knowledge generation.

ABSTRACT. In computational law, formalization of legal rule-based reasoning relies heavily on the subsumption phase, where natural language case descriptions are aligned with formal logical fact formulas. However, several challenges arise due to semantic gaps. This paper investigates the capability of Large Language Models (LLMs) to identify these semantic gaps during legal subsumption. We define a structured methodology categorizing semantic gaps into four distinct types (contradicted, unmentioned, open-textured, and specified) and preliminarily evaluate off-the-shelf LLMs (GPT-4o mini, GPT-4.1, Gemini 3.0 Pro, and Gemini 3.0 Flash Lite). The results show that all models perform well on the task, with greater agreement in identifying contradicted and unmentioned gaps, but less agreement in identifying open-textured and specified gaps.

ABSTRACT. Large language models asked a legal question will always produce an answer. In fragmented treaty regimes, where overlapping international instruments create genuinely contested obligations, this is an ethical problem rather than a technical achievement. This paper argues that the capacity for formal irresolution-the ability of a system to identify and communicate that the applicable law does not yield a determinate answer is an ethical requirement for AI systems operating in high-stakes legal domains. We ground this argument in empirical findings from LLM-based legal information retrieval over a Hohfeld-Structured Normative Knowledge Base of 377 provisions from three marine biodiversity treaties, where LLMs achieve modest precision and middling recall, systematically missing dispersed obligations. We distinguish formal irresolution from technical approaches to epistemic uncertainty in machine learning, arguing that legal indeterminacy is a structural property of fragmented regimes rather than an epistemic gap that better models could close. We illustrate the proposal through a worked scenario demonstrating three distinct output types: determinate answers, structured presentations of competing obligations, and acknowledgements of incompleteness. Formal reasoning frameworks grounded in legal theory can distinguish between settled and contested legal questions, preserving the interpretive authority of human legal actors and preventing what we characterise as a form of hermeneutical injustice through technical architecture.

ABSTRACT. Large Language Models (LLMs) are increasingly used as coordinators in multi-agent systems (MAS), particularly in settings composed entirely of LLM-based agents; however, their ability to coordinate symbolic or hybrid agents under partial observability remains underexplored. In this paper, we investigate whether LLM-based orchestrators can reliably detect and resolve coordination deadlocks in heterogeneous MAS with horizontally partitioned knowledge. We evaluate this problem using a distributed zebra-style constraint satisfaction task involving three specialized agents with restricted reasoning capabilities: equivalence reasoning, relational reasoning, and negation reasoning. We compare fully neural (LLM-only) and neural-symbolic (LLM + Prolog) execution settings under the same orchestration framework, augmented with explicit plan generation, symbolic verification, execution monitoring, and structured planning traces. Our results show that while LLM-based coordination performs effectively in LLM-only settings, it struggles to resolve stagnation and deadlocks when coordinating symbolic agents, despite the symbolic agents maintaining high role-faithfulness and logically valid local reasoning. We further observe that successful LLM-only coordination frequently coincides with violations of role-restricted reasoning boundaries, suggesting that apparent coordination success may partially arise from cross-partition inference leakage rather than robust long-horizon planning. These findings highlight a failure modeinLLMorchestration for heterogeneous MAS and motivate evaluation metrics that consider role compliance and coordination validity in addition to task completion.

ABSTRACT. Multi-hop question generation is the task of generating complex questions that require reasoning over multiple information from multiple contexts. Though the recent works on multi-hop question generation have shown a great impact with excellent outcomes using sequence-to-sequence models and large language models (LLMs), in this paper, we have shown the impact of knowledge graph-based prompting approach on multi-hop question generation. For this purpose, we have proposed a framework, KGP-QG (Knowledge Graph Based Prompting for Question Generation), where, at first, we have generated knowledge graph from the input contexts. The generated knowledge graph is integrated back into the input context to form a knowledge graph-based prompt. Generated knowledge graph enriched prompt is fed to LLMs, where the prompt is used for fine-tuning in two LLMs- BART and T5 models. For this research, the HotpotQA dataset is used for result evaluation and fine-tuning. The result of our proposed framework, KGP-QG, has outperformed the existing methodology

ABSTRACT. Clinical reasoning requires weighing the quality of evidence, assessing its relevance to a particular patient, and reaching a judgement that accounts for both. Doctors perform this reasoning implicitly; it is difficult to articulate and impossible to audit. Large language models can extract structured information from the biomedical literature, but their outputs lack the transparency and traceability that clinical decision-making demands. This article proposes a neurosymbolic framework in which an LLM performs the extraction and a symbolic layer performs the reasoning. The symbolic layer combines three components: Toulmin's model for organising the evidence, epistemic commitments for identifying the beliefs that bear on the decision, and a reason-based reasoning architecture for weighing those beliefs through the hierarchy of evidence and the characteristics of the patient. The framework is non-monotonic (the same evidence yields different recommendations for different patients) and auditable (every recommendation traces back through explicit reasons to source evidence). A case study on semaglutide prescription illustrates how three patient profiles produce three outcomes (prescribe, undecided, withhold) from the same evidence base, with full provenance at every stage.

ABSTRACT. Agentic AI systems that use large language models (LLMs) to implement advanced applications are becoming more and more popular. However, such systems are known to be prone to hallucinations, opaque in the way they make decisions, and not offering any safety guaranties. This introduces the need for principled ways to analyze and design such systems, such that their behavior is verifiable against requirements specifications, and compliant and explainable according to predefined behavioral and decision making rules and strategies. We propose a requirements-driven neuro-symbolic approach to systematically developing such systems. Our approach uses agent-based goal models to represent the system requirements and allowable strategies for achieving them. Goal models are then translated into programs in the IndiGolog logic-based agent programming language to orchestrate agent behavior in compliance with the requirements. LLMs are integrated as components for pursuing specified tasks and rendering formalizable results, amenable to symbolic verification and explanation generation. In this paper, we introduce our approach and examine how it can be applied to the design of a meeting scheduling application.

ABSTRACT. Aggregating inconsistent judgments from multiple reasoners is a long-standing problem in artificial intelligence, with mature foundations in consensus theory, belief pooling, and epistemic logic. The topic has gained renewed significance with the rise of LLM ensembles, where the judgments of specialist LLMs, retrieval-augmented agents, and judge models often diverge even on factual matters. The core problem is that such divergent judgments usually reflect differences in the information sources accessed, not random noise. How can an aggregation rule recognize this distinction and reconcile beliefs without forcing agents that cannot represent the same fine-grained truth to converge? In this paper, we propose a symbolic-probabilistic framework that unites symbolic epistemic structure with probabilistic belief dynamics. Agents maintain probability distributions constrained by their epistemic partitions, while a critic agent computes a dynamic target belief and specialists shift only toward epistemic projections of that target. The framework supports a convergence guarantee and provides a diagnostic interpretation of the remaining disagreement. The contribution is a concise symbolic-statistical approach to feasible consensus in settings where agents cannot necessarily represent the same knowledge.

ABSTRACT. The growing adoption of symbolic knowledge-based decision systems highlights the need of principled methods to quantify their interpretability. While symbolic models are often considered inherently understandable, the cognitive effort required to interpret them varies significantly depending on both their structural properties and on user-specific factors. Existing approaches typically rely on coarse structural proxies, neglect the user perceptions and are often limited to specific representation formats. This work introduces a representation-agnostic and parametric framework for modelling the cognitive complexity of symbolic knowledge bases. First, a taxonomy is proposed to characterise structural and procedural elements of knowledge representations contributing to cognitive complexity, including size, depth, logical composition, and evaluation dynamics. Building on this taxonomy, a unified formal model is defined. The model decomposes cognitive complexity into micro-level components, capturing the internal structure of knowledge units, and macro-level components, accounting for their global organisation and evaluation process. The model also incorporates user-defined tolerance parameters, enabling a profiled assessment of perceived complexity. The proposed framework is validated through an experimental analysis on a set of knowledge bases with different structural properties. Results show that the model captures meaningful differences in cognitive effort across representations and user profiles, providing a flexible tool for designing, evaluating, and comparing interpretable symbolic models.

ABSTRACT. Frontier reasoning-tuned language models still fail on deductive tasks at depth, and the cost of improved performance through extended internal reasoning scales poorly. Symbolic delegation offers a complementary route: a language model translates the problem, while a solver performs the inference. However, current autoformalization pipelines for logic programming are typically bespoke integrations tied to particular tasks or agents. We introduce \textsc{PrologMCP}, a task-agnostic, open-source server that exposes Prolog as a stateful tool through the Model Context Protocol (MCP). Its compact tool interface, structured error reporting, and per-session isolation make the \emph{translate--run--inspect--repair} loop a reusable primitive for MCP-capable agents. We evaluate a formalizer agent enhanced with \textsc{PrologMCP} against standard and reasoning LLMs (Claude Sonnet 4.6, GPT-4.1, and o4-mini) on two subsets of \texttt{PARARULE-Plus}: a general-purpose sample and a more challenging one targeting a specific failure mode of natural-language reasoning. On the general sample, the formalizer matches or exceeds reasoning LLMs (accuracy 1.00 vs.\ 1.00 / 0.998), with the largest gains over standard models (0.762 for GPT-4.1). On the challenging subset, the formalizer remains near-perfect (1.00 / 0.99) while reasoning LLMs drop to 0.95 / 0.94. These results suggest that delegating inference to Prolog via MCP is a robust and inspectable alternative to extended natural-language reasoning.

ABSTRACT. Automatic translation of Natural Language (NL) sentences into logic representation – such as First-Order Logic (FOL) formulae – is a task of great interest for many communities, including Knowledge Representation (KR), Natural Language Processing (NLP), normative Artificial Intelligence (AI), and AI in general. Recently, an increasing number of works have explored the use of Large Language Models (LLMs) to translate NL into FOL, with promising results. However, existing works mostly focus on the translation of a single NL sentence at a time, resulting in independent formulae that do not share a common predicate and constant name space. This work presents a new framework – namely Natural Language to Uniform First Order Logic (NL2UNIFOL) – that leverages LLMs to translate NL sentences into a uniform FOL theory, where formulae share the same vocabulary. NL2UNIFOL is an end-to-end pipeline that: (i) translates NL sentences into preliminary FOL formulae; (ii) identifies and safely merges predicate and constant names to obtain a uniform vocabulary; (iii) finally generates the final FOL theory. We use NL-FOL pairs from the MALLS dataset to validate our framework. Results show that in the majority of cases NL2UNIFOL is able to correctly create clusters of predicate and constant names to be merged together, which allows to obtain a uniform FOL theory. We also recognise that there are still a numerous amount of cases in which the represented name and the representative name semantically diverge, which motivates future work to further improve the noun clustering process.

ABSTRACT. We present a first formal model of ranking-based human–AI aggregation. We represent both human and AI judgments by ranking functions over possible worlds and study how AI advice affects the epistemic quality of a collective decision. To obtain probabilistic guarantees, we introduce a top-choice abstraction that translates ranking aggregation into a voting problem and connects our setting to a recent model from the voting literature. This allows us to model a specific interaction setting that we refer to as AI-follow assistance with concurrent panel revision. The resulting analysis yields a formal account of when AI influence can be epistemically harmful by inducing dependence among human agents.

ABSTRACT. In this work we study offline reinforcement learning (RL) under temporally extended task constraints expressed in Linear Temporal Logic over finite traces (LTLf). Recently, transformer-based approaches such as Trajectory Transformers and Decision Transformers have been adopted to address RL as a sequence modeling problem. However, these methods optimize purely for reward and do not account for high-level temporal requirements. Here, we introduce a neurosymbolic framework that injects LTLf background knowledge into such transformer-based RL policies. Our approach compiles LTLf formulas into deterministic finite automata (DFAs) and integrates them into the learning process through a differentiable representation and a logic-based loss function. In particular, we derive differentiable satisfaction signals from DFA progression and use them as a regularization term during training. The resulting method is architecture-agnostic across different models. We evaluate the proposed framework on navigation environments with specification suites covering combinations of safety and reachability temporal properties. Experimental results show that incorporating background knowledge not only improves constraint satisfaction, but also maintains competitive return compared to vanilla baselines.

ABSTRACT. Retrieval-Augmented Generation is as a paradigm that integrates the parametric knowledge of a Large Language Model with external non-parametric information through a retrieval component. While RAGs have shown strong performances on several tasks, their answers may still be prone to errors as of the approximate nature of the retrieval process they utilize. In this paper, we argue that the output of a dense retriever can be interpreted as a set of noisy samples that could be retrieved via a query that it is implicitly executed over the data. A symbolic representation of this latent intention can then be reconstructed from the samples and used to compute the final retrieval result. We instantiate this idea in a concrete retrieval architecture, which we call Symbolic-Assisted Dense Retriever, and present a preliminary use case scenario and experimental evaluation.

ABSTRACT. Temporal logics and their automata-based counterparts have risen as languages for reasoning about processes and verifying their properties. Increasingly, though, the events in a process refer to observations perceived and classified by a neural network, calling for a mixed formalism capable of dealing with the uncertainty of these classifications. We propose Deep Weighted Finite Automata, a new neuro-symbolic architecture combining formal temporal specifications and neural classifications. We also show how to make probabilistic inferences and compute the most likely execution given a sequence of perceptions.