Abstract
Reasoning caches built from LLM traces can efficiently apply previously derived solutions to new problem instances with robustness and much lower invocation cost than source models.
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.
Community
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.
This made me grin: LLMs reinventing the wheel every time is so true, and caching the solved reasoning to run on a CPU is a lovely fix. How do the caches cope when a new problem looks similar but breaks one assumption?
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Phronix: The Capability-Centric Learning Paradigm for Agents (2026)
- Agent Behavior as Code: Efficient and Robust LLM Agents with Programmatic Specifications (2026)
- Sapien: A Stateful Policy Engine for Autonomous AI Agents (2026)
- Agent in a Bottle: Can LLM Agents Turn Their Capabilities Into Cheap, Scalable Artifacts? (2026)
- Whose turn is it? Context-guided model routing for stateful agents (2026)
- Thinking Before Thinking: Scaling Agentic Inference Through Meta-Reasoning (2026)
- AgentReplay: Token-Wise Trace Replay Is Essential for Fair Serving System Performance Benchmarking (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment: @librarian-bot recommend
Lovely idea to separate finding a solution from executing it, and the compiled form that drops the embedding engine and runs as plain Python with a finetuned Llama 3.2 1B as target makes this feel very deployable on edge or homelab machines. The AIME '25 result, where the cache scores 93.3% against 83.3% for the source, is striking. Since the root function is judged against both the transcript and the golden label during construction (Section 2.2.1), I wonder how much of that gain comes from deterministic execution versus the label check, and how often entry-point search produces a confident hit on a query that looks similar but is not actually equirational?
Get this paper in your agent:
hf papers read 2611.09891 Don't have the latest CLI?
curl -LsSf https://hf.co/cli/install.sh | bash Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper
Collections including this paper 0
No Collection including this paper