CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse
The authors report CPU-executable reasoning caches that cost 16–291 times less to invoke than the model that produced them.
The method turns LLM reasoning traces into reusable programs for applying an already-reasoned solution to new problem instances, with little further model involvement. The authors report performance on par with frontier LLMs across a range of difficult mathematical and scientific tasks. They also describe the caches as fully explainable, continually trainable and robust to input perturbations.
HF Daily Papers' note
The method turns LLM reasoning traces into reusable programs for applying an already-reasoned solution to new problem instances, with little further model involvement. The authors report performance on par with frontier LLMs across a range of difficult mathematical and scientific tasks. They also describe the caches as fully explainable, continually trainable and robust to input perturbations.
HF Daily Papers' note
score 5