The Ankyra project introduces a neuro-symbolic engine designed to address the issue of hallucinations in large language models. The core problem with standard LLMs is their probabilistic nature, which prevents them from distinguishing between factual information and generated content that appears plausible but lacks a basis in reality.
The project implements a separation of roles where the language model acts as a proposer, while a symbolic engine handles the decision-making process. This approach utilizes formal logic methods to verify reasoning, moving away from relying solely on the model's internal patterns to determine correctness.
Ankyra is released as an open-source project. The source code, documentation, and verification tools are available for unrestricted use. This development aims to provide a more reliable framework for reasoning tasks compared to existing techniques like chain-of-thought or tree-of-thought, which do not guarantee accuracy.
