NeuroProlog: Multi-Task Fine-Tuning for Neurosymbolic Mathematical Reasoning via the Cocktail Effect

Researchers introduce NeuroProlog, a neurosymbolic framework that improves mathematical reasoning in Large Language Models by converting math problems into executable Prolog programs. The multi-task 'Cocktail' training approach shows significant accuracy improvements of 3-5% across different model sizes, with larger models demonstrating better error correction capabilities.

• →NeuroProlog framework converts math word problems into verifiable Prolog programs to ensure logical consistency in LLM reasoning.
• →Multi-task Cocktail training strategy achieves 3-5% accuracy improvements across model sizes from 3B to 32B parameters.
• →Larger models (32B) can transform unfixable errors into correctable ones with 92.7% overall correction rate.
• →The research reveals critical capacity thresholds where smaller models (8B) eliminate syntax errors but introduce semantic failures.
• →Execution-guided decoding pipeline enables iterative program repair and quantifies model self-debugging capabilities.