The Artificial Intelligence Research Institute (IIIA) offers two training plans to introduce participants to a research career, in the context of CSIC's JAE Intro 2026 call.

Training plans and mentors:

JAEINT26_EX_1284: Neurosymbolic Computation for an Embodied Approach to Diagramatic Reasoning
Mentor: Dr. Marco Schorlemmer (marco@iiia.csic.es)

To address the current barriers and limitations in the reasoning, abstraction, and analogy-making capabilities of neural-network-based AI systems, particularly in the context of diagrammatic understanding and reasoning, we aim to conduct fundamental research to explore a new paradigm for neurosymbolic AI grounded in the principles of embodied cognition. This paradigm should combine the neural-network-based computation of a given input –the image of an abstract diagram like those used in mathematics and computer science– with the symbolic representation of image-schematic structure and dynamics, thus modelling an ‘embodied understanding' of the information processed by the neural network. (Image-schemas are a framework proposed in cognitive linguistics to capture the recurring dynamic patterns of our perceptual interaction and motor programs that give coherence and structure to our embodied experience.) Some neurosymbolic approaches typically aim to integrate the connectionist level directly with the level of predicates and logic. In contrast, our approach aims to integrate the connectionist level with an intermediate level, where qualitative representations of perception can be combined with image schemas to support an ‘embodied understanding’ of an initial perception. This would enable the neurosymbolic system to perform reasoning, abstraction, and analogy in an integrated manner that is more easily understood by humans, owing to the embodied cognition framework.

JAEINT26_EX_1462: Neuro-Symbolic Reasoning for Logical Games with LLMs and SAT/MaxSAT Technology
Mentor: Dr. Felip Manyà (felip@iiia.csic.es)

The objective of this project is to enhance the reasoning capabilities of large language models (LLMs) and vision-language models (VLMs) in logical and combinatorial games by leveraging SAT/MaxSAT solvers. Many such games (e.g., grid-based puzzles, deduction tasks, and constraint-placement problems), including well-known examples such as Sudoku, can be naturally formalised as constraint satisfaction problems. In this project, the student will use LLMs and VLMs to reason about these games by encoding their rules and instances into SAT/MaxSAT through Python, making use of existing solver APIs to support structured and formally grounded reasoning. The work will involve developing Python-based encodings, designing pipelines in which models generate, refine, or interpret logical constraints, and experimentally evaluating model-only, solver-only, and hybrid neuro-symbolic approaches. The aim is to analyse how the integration of symbolic constraint solving influences correctness, consistency, and robustness in structured reasoning tasks. The expected outcome is a reproducible research prototype and an empirical study demonstrating the strengths and limitations of combining foundation models with SAT/MaxSAT-based reasoning for logical games.