AnyS1: a 4B System-1 model cuts agent workflow costs by 75–91% without losing accuracy

kalyan_kpl · x · 2026-11-22

The AnyS1 paper trains a single lightweight System-1 model to handle routine decision points across diverse agent workflows. Key contributions: skill-based decision mining automatically extracts enumerable decision points from 1,072 public agent skills; the Any-DP dataset builds 17,865 typed decision points and 25,986 execution-grounded states to train a general-purpose 4B pointer model; a hybrid System-1/System-2 scheme handles high-confidence bounded decisions in one forward pass while delegating uncertain ones to System-2. Result: 75–91% lower episode costs and 63–92% lower model-call time on SOPBench, ALFWorld, and WebShop without sacrificing task success.

Original post →

More from coding & agent

coding & agent channel →