NVIDIA's CPU-ARC: reasoning caches 16–291x cheaper with frontier-level performance
nvidia · hf · 2026-11-10
NVIDIA's paper 'CPU-ARC: Agentic Reasoning Caching for on-CPU Reasoning Reuse' (Belcak, Heinrich, Muralidharan, Diao, Molchanov; part of Project Moonshot: Local Frontier Intelligence) argues LLMs conflate task-solving with executing solutions, re-deriving everything from scratch. CPU-ARC builds CPU-executable agentic reasoning caches (ARCs) from LLM reasoning traces, applying previously derived solutions to new problem instances. ARCs reportedly match frontier LLMs on difficult math and science tasks while being fully explainable, continually trainable, robust to input perturbations, and 16–291x cheaper to invoke than the source model.
More from Infra
- FlashState inference system doubles vLLM throughput on GLM 5.3 Flash, cuts latency 3.7x — zainhas · 2026-11-10
- UC Berkeley open-sources FreeToken, an edge-native engine running frontier MoE models on consumer hardware — petrusenko_max · 2026-11-10
- Zero data retention is why enterprises self-host: Azure ZDR requires approval, no ZDR for fable — HanchungLee · 2026-11-10
- One slide: when to use Agent Sandbox vs Agent Substrate for agent infrastructure — davemccollough · 2026-11-10
- Godzilla and Sonic on silicon: the chip industry's lost die-art easter eggs — thehiphopswami · 2026-11-10
- MLCommons lays out SC26 MLPerf agenda with Training v6.1 results and roadmap session — TheKanter · 2026-11-10