Papers & preprints

R-APS: Compositional Reasoning and In-Context Meta-Learning for Constrained Design via Reflective Adversarial Pareto Search

João Pedro Gandarela, Thiago Rios, Stefan Menzel, André Freitas

arXiv preprint

A reflective adversarial Pareto-search protocol decomposes reasoning modes, stress-tests robustness, and maintains explicit rule memory for constrained mechanism design.

Language Models Refine Mechanical Linkage Designs Through Symbolic Reflection and Modular Optimisation

João Pedro Gandarela, Thiago Rios, Stefan Menzel, André Freitas

arXiv preprint

A symbolic “lifting” interface lets language models read, compare, and critique a kinematic simulator’s output, refining linkage designs through modular optimisation and carrying across model families.

Reflective Reasoning for SQL Generation

Isabelle Mohr, João Pedro Gandarela, John Dujany, André Freitas

arXiv preprint

Carries the reflective, verify-in-the-loop reasoning approach beyond mechanisms to SQL generation, evidence that the symbolic-interface idea transfers across domains.

Controlled Agentic Planning & Reasoning for Mechanism Synthesis

João Pedro Gandarela, Thiago Rios, Stefan Menzel, André Freitas

arXiv preprint arXiv:2505.17607

An agentic protocol gives each kind of reasoning (proposing, stress-testing, correcting) its own context and treats worst-case robustness as a first-class objective, so a small, frozen language model becomes reliable on mechanism synthesis through organisation rather than scale.

Inductive Learning of Logical Theories with LLMs: An Expressivity-graded Analysis

João Pedro Gandarela de Souza, Danilo Carvalho, André Freitas

Proceedings of the AAAI Conference on Artificial Intelligence, 39(22), pp. 23752–23759

An expressivity-graded analysis of how well large language models induce logical theories, mapping precisely where neural models succeed and fail at formal reasoning.

Automated, LLM-enabled Extraction of Synthesis Details for Reticular Materials from Scientific Literature

Viviane Torres da Silva, Alexandre Rademaker, Krystelle Lionti, Ronaldo Giro, Geisa Lima, Sandro Rama Fiorini, Marcelo Archanjo, Breno W S R Carvalho, Rodrigo Neumann Barros Ferreira, Anaximandro Souza, João Pedro Gandarela de Souza, Gabriela Valnisio, Carmen Paz, Renato Cerqueira, Mathias B Steiner

AI for Accelerated Materials Design - NeurIPS 2024

An LLM pipeline that extracts synthesis procedures for reticular materials from the literature, accelerating materials-discovery workflows.

Dataset of Reticular Materials' Syntheses Automatically Created from PDFs by using LLMs

Viviane T Silva, Rodrigo Neumann Barros Ferreira, João Pedro Gandarela de Souza, Ronaldo Giro, Geisa Lima, Carmen Nilda Mena Paz, Alexandre Rademaker, Gabriela de Valnisio

American Chemical Society (ACS) Fall Meeting

Hypergraph neural networks with logic clauses

João Pedro Gandarela de Souza, Gerson Zaverucha, Artur S d’Avila Garcez

2024 International Joint Conference on Neural Networks (IJCNN), pp. 1–8

Encodes logic clauses inside a hypergraph neural network: the “sound-by-construction” pole of neuro-symbolic AI, where structure lives inside the model.

Lessons learned in Knowledge Extraction from Unstructured Data with LLMs for Material Discovery

Viviane T Silva, João Pedro Gandarela de Souza, Renato Fontoura de Gusmao Cerqueira

American Chemical Society (ACS) Spring Meeting

Knowledge extraction pipeline with foundation models for material discovery

Viviane T Silva, Breno William Santos Rezende de Carvalho, Marcelo Archanjo Jose, Sandro Rama Fiorini, João Pedro Gandarela de Souza

American Chemical Society (ACS) Fall Meeting