CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse
CPU-ARC turns prior reasoning traces into reusable CPU-executable procedures, reporting frontier-level performance on math and science tasks at 16–291× lower invocation cost. It suggests a way to avoid rerunning a model for every similar instance.
The paper introduces **agentic reasoning caches (ARCs)** built from LLM reasoning traces and executable on CPUs. It reports performance on par with frontier models across difficult math and science tasks, at **16–291× lower invocation cost** than the source model.
Builders can look for repeated, structured reasoning in agent workflows and consider whether a cached procedure could handle new instances with less model involvement. The authors also describe the caches as explainable, trainable, and robust to input perturbations.
The paper introduces **agentic reasoning caches (ARCs)** built from LLM reasoning traces and executable on CPUs. It reports performance on par with frontier models across difficult math and science tasks, at **16–291× lower invocation cost** than the source model. Builders can look for repeated, structured reasoning in agent workflows and consider whether a cached procedure could handle new instances with less model involvement. The authors also describe the caches as explainable, trainable, and robust to input perturbations. The material does not specify implementation details, deployment requirements, or how the results vary across tasks. The reported performance and cost claims come from the paper’s abstract.
The paper reports executable CPU reasoning caches that reuse LLM reasoning traces and claims frontier-comparable performance on difficult math and science tasks at substantially lower invocation cost. It adds a reuse strategy for repeated reasoning, but the abstract leaves implementation and task-specific variation unclear.