Program

Large language models have dramatically expanded the apparent reach of AI systems, enabling fluent interaction, flexible task execution, and increasingly sophisticated agentic behavior. Yet their impressive performance also exposes a fundamental limitation: generated outputs are not, by themselves, justified conclusions. LLMs may hallucinate facts, overlook constraints, fail to respect preferences, or produce reasoning that is persuasive but invalid. This is especially problematic when LLM-based agents are used in settings where correctness, compliance, explainability, and safety matter.

In this talk, I will argue that traditional research in logic programming, non-monotonic reasoning, and knowledge representation is not merely relevant to the current wave of AI systems; it is likely to be essential for building reliable ones. Logic programming provides declarative ways to specify rules, constraints, policies, defaults, exceptions, and goals. Non-monotonic reasoning provides principled mechanisms for reasoning with incomplete information, revising conclusions in light of new evidence, handling defaults and exceptions, and representing competing preferences. These are precisely the kinds of reasoning challenges that arise in LLM-based agents.

I will discuss two complementary roles for logic programming and non-monotonic reasoning in future AI architectures. First, they can help train, evaluate, and structure models so that they acquire more disciplined forms of reasoning. Second, and perhaps more importantly for safety-critical applications, they can serve as external verification and control layers: checking LLM-generated plans, explanations, tool calls, policy decisions, and intermediate conclusions before these are shown to users or passed to subsequent stages of an agent pipeline.

The talk will highlight scenarios in which such techniques are particularly valuable, including agents that must satisfy policy constraints, respect policy preferences, reason under incomplete information, justify their conclusions, and avoid unsafe or non-compliant actions. I will suggest that the future of high-quality AI agents is not a choice between neural and symbolic methods, but a careful integration of generative capabilities with declarative specifications, non-monotonic reasoning, and formal verification. For the logic programming and non-monotonic reasoning community, this moment presents both a challenge and an opportunity: to bring decades of work on formal reasoning into the design of trustworthy AI systems.

Business processes underpin many real-world domains, from healthcare and manufacturing to e-commerce, by defining how activities are coordinated to achieve organizational goals. Process Mining (PM) aims at discovering, analyzing and monitoring business processes from event logs. While classical approaches assume complete observations in logs and deterministic process models, real-world applications are inherently affected by uncertainty. Event logs may contain missing or ambiguous events, while process models may need to distinguish between mandatory requirements and soft recommendations rather than strict Boolean constraints.

This talk introduces the foundations of “Declarative” Process Mining based on the DECLARE language and reviews the current state of the art on uncertainty management in this field. It then shows how probabilistic logics, and in particular Probabilistic Logic Programming (PLP), provide a natural semantic framework for modeling uncertainty in both process models and log activities.

By transferring the Distribution Semantics from PLP to Declarative PM, probabilities can be associated with process models’ constraints and uncertain events, yielding a unified framework for the main tasks in PM: (probabilistic) conformance checking and (probabilistic) process discovery. Recent developments on efficient conformance checking algorithms and PLP-based discovery are also presented.

The talk concludes by discussing current research directions, including scalable inference, causal process discovery, and richer forms of uncertainty over event data.

Reinforcement Learning (RL) is a highly successful branch of machine learning for training an agent to maximize rewards in a given environment. RL comes in different incarnations and flavors, and is actively advanced. In the recent years, there has been increasing interest in using symbolic AI approaches, including Answer Set Programming (ASP), for improving the expected behaviour of RL trained agents, especially with regard to trustworthiness and guarantees.

In this talk, we review different usages of ASP for RL, where we put a focus on enforcing ethical behavior or norms based on deontic constraints that the agent should adhere to. Notably, such constraints can emerge after the training and require instant action, such as fixing the agent’s policy for staying operational, and retraining the agent, possibly on the fly. Real-time requirements and large state spaces make this a challenging application domain for ASP, which needs to be addressed with further advancing ASP technology.