AnyS1 Trains One 4B System-1 Across 1,000+ Skills, Cuts Agent Cost 75-91% on Three Benchmarks
Summary
USTC and NTU researchers release AnyS1, a 4B pointer System-1 trained on grounded enumerable decisions mined from 1,000+ public skills and run inside isolated containers. Across three unseen benchmarks and multiple System-2 model families, AnyS1 cuts episode cost 75-91% and model-call time 63-92% versus an uncompiled System-2 agent, while preserving task reliability under zero-shot threshold transfer.
Originally reported by arxiv.org
Read the original article →Original headline: AnyS1 Trains One 4B System-1 Across 1,000+ Skills, Cuts Agent Cost 75-91% on Three Benchmarks