CPU-ARC Paper: Cached Agentic Reasoning Traces Match Frontier LLMs at 16-291x Lower Cost

The CPU-ARC paper introduces Agentic Reasoning Caches, CPU-executable artifacts distilled from LLM reasoning traces that solve problem families without re-invoking the source model. The authors report on-par performance with frontier LLMs across math and science tasks at 16-291x lower invocation cost, with ARCs being fully explainable, continually trainable and robust to input perturbations — reframing inference as cache-hit reuse for recurring reasoning shapes.