A New Agency-Focused Supervision Approach Scales Software AI Agents With Only 78 Examples

Researchers from Shanghai Jiao Tong University and SII Generative AI Research Lab have introduced LIMI ("Less Is More for Agency"), a supervised fine-tuning approach that significantly enhances the capabilities of large language models (LLMs) for software and research tasks using only 78 carefully curated, tool-grounded demonstrations. Unlike traditional methods that rely on vast amounts of instruction data, LIMI emphasizes the importance of data quality and structure, demonstrating that agentic competence scales more effectively with dense, multi-turn trajectories that encompass reasoning, tool use, and environment interactions. Fine-tuning models like GLM-4.