Large language models (LLMs) and multimodal large language models (MLLMs) have unlocked impressive reasoning ability.
However, when tasks demand precise logical guarantees, explicit structure, or reliable multi-step deduction,
purely neural approaches often fall short. Their reasoning can be fluent yet brittle: explanations may sound convincing
while quietly violating basic logical constraints.
Neuro-symbolic approaches aim to close this gap by coupling LLMs with symbolic representations and reasoning procedures.
Symbolic reasoning enable compositionality, support program-like manipulation of knowledge, and good traceability and interpretability.
Neural models, in turn, offer strong capabilities in learning, perception, and pattern discovery. When combined, they allow us to:
- Improve faithfulness: Reasoning steps can be tractable and verifiable, thus enabling faithful reasoning.
- Enhance interpretability: Symbolic traces and proofs provide human-inspectable explanations.
- Enable compositional generalization: Logic and constraints naturally capture reusable structures and rules.
- Support robustness & safety: Formal checks help enforce invariants in high-stakes scenarios.
- Extend beyond text: In multimodal settings, symbols provide a common reasoning layer over images, text, and other signals.
This page collects a set of Neuro-Symbolic Large Model projects—covering both symbolic-guided LLM reasoning
and multimodal symbolic reasoning—that explore these ideas in different settings and benchmarks.
