Thinking Fast, Anywhere: One System-1 for the Decision Points of Any Agent Workflow
Summary
- Researchers introduce AnyS1, a framework that uses human-authored procedural skills to create a domain-general source of decision points for LLM agent workflows.
- AnyS1 mines enumerable decisions from over 1,000 public skills and trains a single 4B pointer System-1 model to handle routine choices.
- At run time, compiled workflows resolve routine decisions in a single forward pass and fall back to System-2 when confidence drops below a threshold.
- Across three unseen benchmarks, AnyS1 reduces episode costs by 75-91% and model-call times by 63-92% while preserving task reliability.
Why it matters
AnyS1 significantly reduces the high cost and latency of autonomous LLM agents by delegating routine intermediate steps to a lightweight System-1 model without sacrificing reliability.