LIMI: Less is More for Agency

We define “Agency” as the emergent capacity of AI systems to function as autonomous agents—actively discovering problems, formulating hypotheses, and executing solutions through self-directed engagement with environments and tools. This fundamental capability marks the dawn of the “Age of AI Agency”, driven by a critical industry shift: the urgent need for AI systems that don’t just think, but work. While current AI excels at reasoning and generating responses, industries demand autonomous agents that can execute tasks, operate tools, and drive real-world outcomes. As agentic intelligence becomes the defining characteristic separating cognitive systems from productive workers, efficiently cultivating machine autonomy becomes paramount. Current approaches assume that more data yields better agency, following traditional scaling laws from language modeling. We fundamentally challenge this paradigm. LIMI (Less Is More for Intelligent Agency) demonstrates that agency follows radically different development principles. Through strategic focus on collaborative software development and scientific research workflows, we show that sophisticated agentic intelligence can emerge from minimal but strategically curated demonstrations of autonomous behavior. Using only 78 carefully designed training samples, LIMI achieves 73.5% on AgencyBench, dramatically outperforming state-of-the-art models: Kimi-K2-Instruct (24.1%), DeepSeek- V3.1 (11.9%), Qwen3-235B-A22B-Instruct (27.5%), and GLM-4.5 (45.1%). Most strikingly, LIMI demonstrates 53.7% improvement over models trained on 10,000 samples—achieving superior agentic intelligence with 128 times fewer samples.

Our findings establish the Agency Efficiency Principle: machine autonomy emerges not from data abundance but from strategic curation of high-quality agentic demonstrations. This discovery fundamentally reshapes how we develop autonomous AI systems, suggesting that mastering agency requires understanding its essence, not scaling training data. As industries transition from thinking AI to working AI, LIMI provides a paradigm for sustainable cultivation of truly agentic intelligence.

However, the development of such agentic systems faces critical challenges. Current approaches assume that more data yields better agentic intelligence, following traditional scaling laws from language modeling (Kaplan et al., 2020; Rae et al., 2021; Chowdhery et al., 2023; Scao et al., 2022; Zhang et al., 2022). This paradigm leads to increasingly complex training pipelines and substantial resource requirements, yet this fundamental assumption remains largely untested: do agentic capabilities truly require exposure to vast amounts of training data, or could they emerge more efficiently through strategic approaches? Emerging evidence from adjacent domains suggests a compelling alternative paradigm. LIMA (Zhou et al., 2023) achieved effective model alignment with only 1,000 carefully curated examples, while LIMO (Ye et al., 2025) demonstrated that complex mathematical reasoning can emerge from just 817 strategically selected training samples, achieving a remarkable 45.8% absolute improvement with only 1% of the data typically required. These convergent findings suggest that strategic data curation may be fundamentally more powerful than dataset scale for developing sophisticated AI capabilities, naturally leading us to investigate whether agentic intelligence follows similar efficiency principles.

We introduce LIMI (Less Is More for Intelligent Agency), which demonstrates that agency follows radically different development principles from traditional scaling approaches. Through strategic focus on collaborative software development and scientific research workflows–domains that collectively span the majority of knowledge work scenarios–we show that sophisticated agentic intelligence can emerge from minimal but strategically curated demonstrations of autonomous behavior. Our approach is grounded in three core innovations: (i) First, we pioneer novel agentic user query synthesis methodologies, including human-AI collaborative query collection from realworld scenarios and systematic GitHub pull request-based query synthesis using advanced LLMs, ensuring that our training demonstrations capture authentic patterns of agentic behavior while maintaining ecological validity; (ii) Second, we develop a systematic trajectory collection protocol that captures complete multi-turn interaction sequences for each curated query, recording the full collaborative workflow from initial task understanding through iterative model reasoning, tool utilization, and environmental feedback to successful task completion, providing high-quality training demonstrations of sophisticated agentic behavior in realistic operational contexts; (iii) Third, we reveal the data efficiency principle for AI agency cultivation, demonstrating that sophisticated agentic intelligence emerges from strategic curation of minimal high-quality demonstrations rather than large-scale data accumulation, fundamentally challenging traditional scaling paradigms in agentic AI development.