Tutorial at
the 23rd International Conference on Principles of Knowledge Representation and Reasoning (KR 2026)
co-located with the 9th Federated Logic Conference (FLOC) 2026
July 19, 2026
Lisbon, Portugal
Knowledge Representation and reasoning (KR) has long provided a foundation for building reliable AI systems. Logical formalisms and reasoning engines allow for precise, verifiable conclusions. Yet KR faces a persistent challenge: the knowledge acquisition bottleneck. Encoding knowledge into formal representations is time-consuming, and many end-users are reluctant to communicate in formal languages.
Large Language Models (LLMs) offer a strikingly different approach. They are capable of processing and generating natural language fluently and performing diverse tasks with little explicit training. However, this flexibility comes at a cost: LLMs are fundamentally unreliable, prone to inconsistencies, hallucinations, and errors.
Recent work has explored combining LLMs with KR methods to leverage the strengths of both paradigms. By doing so, researchers can build systems that are intuitive to use while remaining trustworthy. This tutorial provides an accurate, non-hyped introduction to LLMs. We will cover how they work, why they can fail, and how their outputs can be constrained using formal methods from logic and automata theory. Participants will learn strategies for using LLMs as interfaces between natural and formal languages and for integrating them with reasoning tools to build more reliable intelligent systems.
The tutorial aims at researchers in Knowledge Representation and Reasoning (KR) and other logic related disciplines who are interested in exploring the potential of Large Language Models for their work. No expertise in machine learning or neural networks is assumed to make the tutorial accessible to KR researchers who want to understand LLMs from a practical, research-oriented perspective.
We will mostly focus on high-level ideas, but a basic understanding of algorithms, formal languages, linear algebra and probability theory as usually taught in an undergraduate computer science degree will be helpful to understand some technical parts.
During the tutorial, participants will:
Understand the core principles behind LLMs, including their mechanics, capabilities, and limitations.
Gain insight into common failure modes of LLMs, such as hallucination, inconsistency, and bias.
Learn methods to bridge natural language and formal languages with LLMs.
Explore strategies for combining LLMs with reasoning engines to improve reliability and correctness.
Acquire a realistic, non-hyped perspective on the opportunities and challenges of using LLMs in KR research.
Become able to identify research directions where LLMs can complement traditional KR methods and apply them in their own work.
The tutorial will be interactive and leave time for questions and discussions. We will try to cover the following topics, but may skip parts depending on the amount of interaction and the audience’s interests.
Introduction and Motivation (30 minutes)
Why LLMs matter for KR and logic
Strengths and limitations of LLMs
Essentials of LLMs (90 minutes)
History, core architecture, and key training ideas
Prompting, in-context learning, and fine-tuning basics
Typical failures (hallucination, inconsistency, bias)
Translating Between Natural and Formal Languages (60 minutes)
Ideas based on prompting, fine-tuning, and constrained generation
Examples from logic programming, temporal logic, and knowledge graphs
Combining LLMs with Reasoning Engines (75 minutes)
Neuro-symbolic approaches
Agentic approaches
Challenges and Outlook (15 minutes)
Sample References
Yongchao Chen, Rujul Gandhi, Yang Zhang, Chuchu Fan: NL2TL: Transforming Natural Lan guages to Temporal Logics using Large Language Models. EMNLP 2023: 15880-15903.
Saibo Geng, Martin Josifoski, Maxime Peyrard, Robert West: Grammar-Constrained Decoding for Structured NLP Tasks without Finetuning. EMNLP 2023: 10932-10952.
Adam Ishay, Joohyung Lee: LLM+AL: Bridging Large Language Models and Action Languages for Complex Reasoning About Actions. AAAI 2025: 24212-24220.
Adam Ishay, Zhun Yang, Joohyung Lee: Leveraging Large Language Models to Generate Answer Set Programs. KR 2023: 374-383.
Antonello Meloni, Diego Reforgiato Recupero, Francesco Osborne, Angelo A. Salatino, Enrico Motta, Sahar Vahdati, Jens Lehmann: Assessing Large Language Models for SPARQL Query Generation in Scientific Question Answering. ISWC 2024.
Shubham Ugare, Tarun Suresh, Hangoo Kang, Sasa Misailovic, Gagandeep Singh: SynCode: LLM Generation with Grammar Augmentation. Transactions on Machine Learning Research 2025 (2025).
Honghua Zhang, Liunian Harold Li, Tao Meng, Kai-Wei Chang, Guy Van den Broeck: On the Paradox of Learning to Reason from Data. IJCAI 2023: 3365-3373.
I received a PhD in Computer Science in 2016 from the University of Hagen in Germany, and had postdoc positions at the University of Osnabrück (2016– 2020), the University of Stuttgart (2020 2022), and Imperial College London (2022– 2023). Since 2023, I’m a lecturer at Cardiff University, where I’m leading the Knowledge Representation and Reasoning group. I’ve been doing research in AI for 15 years on topics such as probabilistic logics, computational argumentation, description logics, knowledge graphs, and inconsistency tolerance. In recent years, I have focused in particular on combining reasoning and machine learning methods.
I started working with statistical language models as an undergraduate when they were still largely based on probabilistic graphical models. During my PhD, I followed their transition to deep learning approaches and, as a postdoctoral researcher, I taught LSTM-based language models for several years in an advanced AI course at the University of Osnabrück. I followed the introduction of the attention mechanism in RNNs and how it led to the transformer architecture that underlies today’s LLMs. While considerable progress has been made in the predictive components of generative language models over the past decades, their basic generation mechanism has remained largely unchanged and explains many strengths and weaknesses of LLMs from a purely technical perspective. My LLM-related research focuses on neuro-symbolic architectures that aim at combining the benefits of formal and generative methods.
More information