CPU-ARC paper reuses LLM reasoning traces on CPU at lower cost
First merged summary.
Researchers describe CPU-ARC caches that match frontier models on hard math and science tasks at 16–291x lower cost.
Researchers introduce CPU-ARC, agentic reasoning caching that turns LLM reasoning traces into CPU-executable caches so a solved task can be applied to new instances with little language-model involvement. They say the caches match frontier LLMs on difficult mathematical and scientific tasks, stay explainable and continually trainable, and resist input perturbations. Invocation is reported as 16–291x cheaper than the source model. The work addresses models that recompute every problem from scratch.
- ARCs are built from LLM reasoning traces and run on CPU.
- Reported invocation cost is 16–291x lower than the source model.
- Authors say results match frontier LLMs on hard math and science tasks.
- Caches are described as explainable, continually trainable, and robust to perturbations.
- A solved task can be reused with little language-model involvement.
What each outlet adds
beyond the shared facts- Hugging Face daily papersA solved task is reused with little language-model involvement.
- arXivFrames the problem as LLMs recomputing every task from scratch.
Coverage
oldest first · one row per article · UTC- · 6d ago
Hugging Face daily papersFirst report
CPU-ARC caches LLM reasoning traces for cheaper CPU reuse
The paper reports on-par results with frontier models on hard math and science tasks at 16–291x lower cost.
- · 5d ago
arXiv cs.AI / cs.LG / cs.CL
CPU-ARC enables efficient on-CPU LLM reasoning reuse
CPU-ARC builds agentic reasoning caches from LLM traces matching frontier models on math tasks at 16-291x lower cost.