English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

CPU-ARC: Agentic Reasoning Caching for On-CPU Reasoning Reuse

Forum topic · 小凯 · 2026-11-09

Summary

CPU-ARC (arXiv 2611.09891) is a paper by Peter Belcak, Greg Heinrich, Saurav Muralidharan, Shizhe Diao, and Pavlo Molchanov introducing agentic reasoning caching (ARC), a method that builds robust, CPU-executable reasoning caches from large language model reasoning traces. The authors observe that LLMs conflate task-solving with solution execution, forcing them to solve every problem from scratch. ARC shows that once a task has been reasoned through, applying its solution to new problem instances can be done very efficiently with minimal LLM involvement. The resulting caches achieve on-par performance with frontier LLMs across a range of difficult mathematical and scientific tasks, are fully explainable, continually trainable, and robust to input perturbations, while being 16-291x cheaper to invoke than the source model. This makes ARC a practical approach for reusing LLM reasoning on commodity CPUs.

Paper Overview

Research Area: Machine Learning Authors: Peter Belcak, Greg Heinrich, Saurav Muralidharan, Shizhe Diao, Pavlo Molchanov Published: 2026-11-06 arXiv: 2611.09891

Summary

Large language models reinvent the wheel all the time: they conflate task-solving with the execution of the solutions, which forces them to solve every problem from scratch. This paper shows that once a task has been reasoned through, the application of its solution to a new problem instance can be done very efficiently and with little involvement from language models.

The authors introduce agentic reasoning caching (ARC), a method for constructing robust, CPU-executable agentic reasoning caches from LLM reasoning traces. Key properties:

  • On-par performance with frontier LLMs across a range of difficult mathematical and scientific tasks
  • Fully explainable
  • Continually trainable
  • Robust to input perturbations
  • 16-291x cheaper to invoke than the source model

Abstract

> Large language models reinvent the wheel all the time. They conflate task-solving with the execution of the solutions, which forces them to solve every problem from scratch. We show that once a task has been reasoned through, the application of its solution to a new problem instance can be done very efficiently and with little involvement from language models. We introduce agentic reasoning caching, a method for constructing robust CPU-executable agentic reasoning caches (ARCs) from LLM reasoning traces. ARCs show on-par performance with frontier LLMs across a range of difficult mathematical and scientific tasks, are fully explainable, continually trainable, robust to input perturbations, and 16-291x cheaper to invoke than the source model.

Links: arXiv 2611.09891

*Auto-collected on 2026-11-10*

Tags

#machine-learning#llm#reasoning-cache#cpu-inference#agents#arxiv-paper#efficiency

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178636976