Overview
Neural networks have achieved remarkable success across a broad range of domains, including human-like question answering, game playing (Schrittwieser et al. 2020), solving International Mathematical Olympiad problems (Trinh et al. 2024), and code generation (Zhu et al. 2025). Despite these advances, large language models (LLMs) continue to display unpredictable behaviour (Park et al. 2024), struggle with seemingly simple abstract-reasoning tasks (Lampinen et al. 2024), and occasionally produce correct answers supported by incorrect explanations (Creswell, Shanahan, and Higgins 2023).Decomposing complex problems into intermediate steps can improve reasoning performance (Wei et al. 2022a), while code-based prompting may significantly strengthen the causal-reasoning capabilities of LLMs (Liu et al. 2025b). Nevertheless, it remains unclear whether these models can achieve the rigour of symbolic reasoning—or, more fundamentally, whether LLMs genuinely reason at all (Mitchell 2023). This uncertainty raises important concerns about the potentially serious and unpredictable risks that these systems may pose to society (Bengio et al. 2024). Recent developments invite researchers to reconsider the foundations of neural reasoning. The Sphere Neural Network demonstrates rigorous syllogistic reasoning without relying on training data (Dong, Jamnik, and Liò 2025), while related work argues that conventional vector embeddings may oversimplify conceptual representations by omitting non-zero radii (Dong, Jamnik, and Liò 2026). However, human decision-making is also influenced by unconscious desires, emotions, transference, embodiment, power dynamics, and interpersonal relationships (Piotrowska, A. 2026) — factors that conventional neural networks may already capture to some extent. Sphere embeddings may provide a simple yet promising neural foundation for integrating formal reasoning with these less formal dimensions of human inference. Together, these advances motivate closer collaboration among researchers and practitioners working on LLMs, knowledge graphs, and neuro-symbolic reasoning.
This workshop will encourage participants to explore new approaches to reliable and creative reasoning. Relevant research directions include neural-symbolic collaborative distillation (NesyCD), which seeks to transfer the complex reasoning capabilities of LLMs (Liao et al. 2025), and neural-symbolic unification for applications such as trustworthy digital identity systems. Promising directions include scientific discovery and humour research. In drug discovery, for example, SynFlowNet embeds symbolic chemical-reaction rules and available building blocks within a neural generative framework, enabling the model to generate diverse molecules together with feasible synthesis pathways. This integration illustrates how symbolic domain knowledge can guide neural models towards scientifically valid and practically useful solutions. Humour requires contextual understanding, logical inference, and the recognition of incongruity. Combining the pattern-recognition capabilities of neural models with the interpretability and control of symbolic systems could support more reliable humour detection and generation. These diverse forms of reasoning may share a common basis in spatial cognition. In Mind in Motion: How Action Shapes Thought, Barbara Tversky argues that spatial thinking provides a foundation for abstract thought, enabling people to organise and reason about concepts, actions, relationships, and transformations.
Building on this perspective, the workshop will investigate whether logical, causal, creative, linguistic, and social reasoning can be understood and represented within a unified spatial framework and implemented in the vector space. Our objective is to develop a neural model capable of simulating these diverse reasoning processes through learned spatial representations and transformations, while integrating symbolic structures to provide rigour, interpretability, and control.
Topics
This workshop aims to enable an exchange of ideas to enhance the determinacy of neural networks in NLP and knowledge graph reasoning, and to promote AI’s explainability, reliability, and safety. Topics include, but are not limited to (1) Neural-symbolic knowledge representation; (2) Knowledge introspection and localization in LLMs; (3) Symbolic knowledge induction and distillation from LLMs; (4) Neural-symbolic collaborative reasoning; (5) Neural-symbolic distillation and model compression; (6) Tool graphs, procedural knowledge, skills and capability evolution; (7) Benchmarks, datasets, and evaluation methodologies for measuring neural-symbolic reasoning, knowledge consistency, faithfulness, interpretability, generalization, and reliability; (8) Hybrid neral-symbolic systems for autonomous agents, digital economy, education, healthcare, humour, law, manufacturing, robotics, scientific discovery, and other knowledge-intensive applications.Keynote Speakers
<
