CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse
The authors report that CPU-executable reasoning caches match frontier models on tested tasks at 16–291 times lower invocation cost.
CPU-ARC turns language-model reasoning traces into reusable programs that run on CPUs. Once a task has been reasoned through, the cached solution can handle new instances with little language-model involvement. The paper reports comparable performance across a range of difficult mathematical and scientific tasks. The authors describe the caches as explainable, continually trainable and robust to input perturbations.
ArXiv · AI/CL/LG's note
CPU-ARC turns language-model reasoning traces into reusable programs that run on CPUs. Once a task has been reasoned through, the cached solution can handle new instances with little language-model involvement. The paper reports comparable performance across a range of difficult mathematical and scientific tasks. The authors describe the caches as explainable, continually trainable and robust to input perturbations.
ArXiv · AI/CL/LG's note
score 6