sequential-tuning

This paper introduces Sequential Agent Tuning (SAT), a coordinator-free training paradigm for multi-LLM teams that provides monotonic improvement guarantees and plug-and-play invariance, enabling smaller models to outperform larger ones.