The Ankyra project implements a neuro-symbolic reasoning approach that combines large language models with deterministic symbolic solvers. This method aims to mitigate hallucinations in language models by using formal logic to verify facts and derive conclusions, treating the symbolic solver as a structural framework for the model's output.
The reasoning process operates in cycles, or waves. In each cycle, the language model extracts facts and rules from text to build a theory, while the symbolic solver attempts to derive a solution. If the solver cannot reach a conclusion, the model is prompted to provide additional facts or rules. Testing showed that most tasks were resolved within seven waves.
The system automatically selects the appropriate logic type based on the theory extracted from the input text. Supported logic types include first-order logic with Horn clauses, logic with quantifiers and negation, as well as arithmetic and systems of linear equations for specialized tasks.
The author tested the algorithm using the DeepSeek V4 Flash model across several benchmarks, including AI2, ProntoQA, FOLIO, and LSAT-style logic puzzles. Performance was compared against existing implementations like Logic-LM and LINC. On the L2 benchmark slice, the system achieved 29–31 correct out of 45 in live mode.
