Grounding Large Language Models in Reaction Knowledge Graphs for Synthesis Retrieval
Speaker: Lorenzo Di Fruscia
Abstract: Large Language Models (LLMs) can aid synthesis planning in chemistry, but standard prompting methods often yield hallucinated or outdated suggestions. We study LLM interactions with a reaction knowledge graph by casting reaction path retrieval as a Text2Cypher (natural language to graph query) generation problem, and define single- and multi-step retrieval tasks. We compare zero-shot prompting to one-shot variants using static, random, and embedding-based exemplar selection, and assess a checklist-driven validator/corrector loop. To evaluate our framework, we consider query validity and retrieval accuracy. We find that one-shot prompting with aligned exemplars consistently performs best. Our checklist-style self-correction loop mainly improves executability in zero-shot settings and offers limited additional retrieval gains once a good exemplar is present. We provide a reproducible Text2Cypher evaluation setup to facilitate further work on KG-grounded LLMs for synthesis planning.
GenComp: Generative Models for Spatial Compositionality Problems
Speaker: Amelia Villegas Morcillo
Abstract: Compositionality—the ability to construct novel solutions by recombining known components—is central to spatial reasoning tasks such as puzzles, LEGO assemblies, and molecular structures. While recent diffusion-based generative models have explored aspects of compositional generation, they typically do not explicitly model how spatial parts assemble into coherent intermediate sub-solutions. In this talk, I first introduce spatial compositionality metrics for diffusion models that use connectivity graphs extracted during sampling to evaluate whether meaningful part-part and part-assembly relationships emerge and persist across intermediate timesteps. Building on these insights, I then present ongoing work on GenComp, our proposed joint generative framework with two parallel, interacting diffusion processes over spatial and topological variables, in which graph structure guides how components connect and move together during generation. Preliminary results on cross-cut puzzles in a denoising diffusion setting show how enforcing compositional constraints in the forward process affects both topological and spatial prediction and influences generative performance, laying the groundwork for diffusion models that explicitly reason over reusable spatial substructures.