LIMI: Less Is More for Agency (arxiv.org)

LIMI (Less Is More for Intelligent Agency) introduces a new take on building autonomous AI: instead of chasing scale with massive training corpora, the authors show that strategically curated, high-quality demonstrations of autonomous behavior can produce strong agency. They define Agency as the ability of systems to autonomously discover problems, hypothesize solutions, and execute actions using tools and environments—moving AI from reasoning to actual work. LIMI targets collaborative software development and scientific research workflows and formalizes the "Agency Efficiency Principle": autonomy emerges from targeted examples, not sheer data volume. Technically, LIMI is trained on just 78 carefully designed agentic demonstrations and achieves 73.5% on comprehensive agency benchmarks, far outperforming state-of-the-art comparators (Kimi-K2-Instruct 24.1%, DeepSeek-V3.1 11.9%, Qwen3-235B-A22B-Instruct 27.5%, GLM-4.5 45.1%). It also improves 53.7% over models trained on 10,000 samples while using 128× fewer examples. The paper’s result shifts emphasis toward dataset curation, task-relevant interaction traces, and protocol design for training agents—suggesting major practical gains in development cost, sample efficiency, and faster iteration for agentic systems. For AI/ML practitioners, this reframes research priorities from raw scale to the quality and structure of agentic demonstrations.