PaA – Partnership Area

Project

PaA-7

Integration of Logical Tensor Networks into LLMs for Explainable and Efficient Reasoning

Project Heads

Martin Eigel, Alex Goeßmann, Sven Wang

Project Members

Janina Schütte

Project Duration

01.09.2025 − 31.08.2027

Located at

WIAS

Description

Despite human-like chains of thought in reasoning language models, generated reasoning steps merely rely on trained next token probabilities. To strengthen explainability and efficiency, we capture specific problem structures by logic tensor networks for rigorous deduction in a hybrid LLM reasoning method.

Related Publications

Alex Goessmann, Janina Schütte, Maximilian Fröhlich, and Martin Eigel. A tensor network formalism for neuro-symbolic AI. arXiv preprint arXiv:2601.15442 (2026).

Related Picture

Visualization of the Computation-Activation-Network

In the tensor-network framework tnreason, both hard and soft constraints can be represented. The computation layer, shown in black, encodes hard logical constraints, such as A1⇒F. The activation layer, shown in multiple colors, assigns weights to these constraints; for example, an implication may hold only with a given probability.

Project Software

Current software for the tensor-network framework tnreason is available at https://github.com/tnreason.