NARE – A framework that "crystallizes" LLM reasoning into fast Python scripts (github.com)

NARE, a newly introduced Skill-Based Cognitive Architecture, aims to optimize large language model (LLM) reasoning by transforming complex inference tasks into efficient Python scripts for zero-shot execution. This innovation allows for deterministic processing of logic tasks with O(1) latency, bypassing the traditionally slower LLM generation process. By dynamically learning from its own reasoning experiences, NARE compiles abstract algorithms during a consolidation phase, leveraging executable reflexes to solve recurring logic problems without the need for API calls. This framework significantly impacts the AI/ML community by enhancing the efficiency of reasoning execution, effectively shifting the computational burden from auto-regressive generation to local, procedural execution. Key features include four routing protocols: REFLEX for immediate procedural execution, FAST for deterministic solution retrieval, HYBRID for context-augmented reasoning, and SLOW for in-depth explorative thinking. Importantly, NARE’s design includes a fault-tolerant skill registry that evaluates generated algorithms in isolated environments, mitigating risks during execution. Empirical evaluations show that NARE not only conserves computation resources—demonstrating 100% token conservation in specific tasks—but also exhibits exponential speedup in execution, marking a significant advancement in cognitive computing capabilities.