TLDR: A new study, dubbed ‘LIMI’ (Less Is More for Intelligent Agency) by researchers from China, suggests that highly capable autonomous AI agents can be built with a remarkably small dataset of just 78 carefully selected training examples. This challenges the conventional wisdom that massive datasets are essential for advanced AI, demonstrating that strategic data curation can lead to superior performance.
A groundbreaking study, ‘LIMI’ (Less Is More for Intelligent Agency), conducted by researchers from several Chinese institutions, is poised to redefine the development paradigm for autonomous AI agents. Published on September 28, 2025, the research posits that instead of relying on vast, extensive datasets, a mere 78 meticulously chosen training examples can be sufficient to construct highly effective autonomous agents.
Traditionally, the advancement of AI models, particularly in complex domains, has been synonymous with the ingestion of colossal amounts of data. However, the LIMI paper introduces a novel approach, emphasizing the quality and strategic curation of training data over sheer volume. The study defines ‘agency’ as the capacity of AI systems to operate independently, encompassing abilities such as identifying problems, formulating hypotheses, and resolving tasks through self-directed interactions with various environments and tools.
The LIMI methodology diverges significantly from standard AI training protocols. It utilizes a compact set of 78 handpicked examples, each derived from real-world software development and research tasks. These examples are designed to encapsulate the entire human-AI collaborative process, from the initial query and tool utilization to problem-solving and successful task completion. The overarching goal of this focused training is to imbue models with the capabilities required to function as truly autonomous agents.
In terms of performance, the LIMI model has demonstrated impressive results. On the challenging AgencyBench benchmark, it achieved a success rate of 73.5 percent using only its limited set of 78 training samples. This performance is particularly noteworthy when compared to existing open-weight models. LIMI has been shown to outperform models such as GLM-4.5, Deepseek-V3.1, and Kimi-K2 in critical areas like software development, scientific workflows, and coding benchmarks. It achieved an overall success rate of 74.6 percent, significantly surpassing GLM-4.5’s 47.4 percent, despite utilizing substantially less training data.
Also Read:
- Agentic AI Transforms Network Operations with Autonomous Self-Healing Capabilities
- Hugging Face Unveils Smol2Operator: An Open-Source Framework for Developing Agentic GUI Coders from VLMs
These findings lend considerable weight to recent arguments made by Nvidia researchers, who have suggested that many contemporary AI agents employ unnecessarily large language models. Nvidia’s research indicated that models with fewer than 10 billion parameters might be adequate for many agentic tasks. The empirical evidence provided by the LIMI study strongly supports this perspective, underscoring that intelligent data curation can indeed triumph over brute-force scaling in AI development. The study’s implications are profound, suggesting a more efficient and resource-conscious pathway for the creation of advanced autonomous AI agents, potentially accelerating their integration into various industries.
