TLDR: The research paper “SII-GAIR LIMI: Less is More for Agency” introduces a novel approach to developing AI agents that can autonomously discover problems, formulate hypotheses, and execute solutions. Challenging the traditional belief that more data leads to better AI, LIMI demonstrates that sophisticated agentic intelligence can emerge from a minimal, strategically curated dataset of just 78 training samples. This method significantly outperforms state-of-the-art models trained on vastly larger datasets, establishing an “Agency Efficiency Principle” where quality and strategic curation of demonstrations are more critical than data volume for cultivating machine autonomy.
The field of Artificial Intelligence is rapidly moving beyond systems that merely process information to those that can act autonomously. This shift, termed the “Age of AI Agency,” highlights the growing demand for AI that doesn’t just think, but actively works—discovering problems, formulating hypotheses, and executing solutions through self-directed engagement with environments and tools.
Traditionally, the development of advanced AI capabilities, particularly in large language models, has followed a scaling law: more data generally leads to better performance. This often translates into increasingly complex training pipelines and substantial resource requirements. However, a groundbreaking new research paper, “SII-GAIR LIMI: Less is More for Agency”, fundamentally challenges this paradigm.
Introducing LIMI: A New Principle for AI Agency
Authored by a team including Yang Xiao, Mohan Jiang, Jie Sun, Keyu Li, and Pengfei Liu, the paper introduces LIMI (Less Is More for Intelligent Agency). This approach demonstrates that agentic intelligence can emerge from minimal, yet strategically curated, demonstrations of autonomous behavior, rather than from data abundance.
LIMI focuses on two critical domains: collaborative software development (vibe coding) and scientific research workflows. These areas collectively represent a vast majority of knowledge work scenarios and demand the full spectrum of agentic capabilities, such as autonomous task execution, multi-step reasoning, and collaborative problem-solving.
How LIMI Achieves More with Less
The success of LIMI is built upon three core innovations in data curation:
-
Novel Agentic User Query Synthesis: The researchers pioneered methods to create authentic user queries. This involved human-AI collaborative query collection from real-world scenarios and systematic synthesis of queries from GitHub Pull Requests (PRs) using advanced LLMs like GPT-5. This ensures that the training demonstrations accurately reflect genuine agentic behavior.
-
Systematic Trajectory Collection Protocol: For each curated query, a complete multi-turn interaction sequence, or “trajectory,” is captured. These trajectories record the entire collaborative workflow, from initial task understanding through iterative model reasoning, tool utilization, and environmental feedback, all the way to successful task completion. This provides high-quality, realistic examples of sophisticated agentic behavior.
-
The Agency Efficiency Principle: Through this meticulous process, LIMI reveals that sophisticated agentic intelligence arises from the strategic curation of minimal, high-quality demonstrations, rather than from large-scale data accumulation. This directly challenges traditional scaling paradigms in AI development.
Remarkable Performance with Minimal Data
The experimental results are striking. Using only 78 carefully designed training samples, LIMI achieved an impressive 73.5% performance on AgencyBench, a comprehensive benchmark for agentic capabilities. This dramatically outperforms state-of-the-art models like Kimi-K2-Instruct (24.1%), DeepSeek-V3.1 (11.9%), Qwen3-235B-A22B-Instruct (27.5%), and GLM-4.5 (45.1%).
Most notably, LIMI demonstrated a 53.7% improvement over models trained on 10,000 samples, achieving superior agentic intelligence with 128 times fewer training examples. This highlights the profound impact of strategic data curation over sheer data volume.
Generalization and Environmental Impact
LIMI’s advantages aren’t confined to specific benchmarks. Its superiority extends across diverse generalization benchmarks, including tool use, coding, and scientific computing tasks. This indicates that the strategic data curation approach leads to broad improvements in model capabilities.
Furthermore, the research explored the impact of the execution environment. While LIMI showed intrinsic improvements even without direct tool access, its integration with the SII CLI environment significantly amplified these benefits for complex agentic tasks. This suggests that foundational reasoning improvements, combined with sophisticated tool coordination skills, are key to effective agentic AI.
Also Read:
- Foundation Models Navigate Virtual Worlds: New Strategies for Reinforcement Learning
- Understanding Limits in AI Alignment: A Capacity-Based Perspective
Reshaping AI Development
The findings from the LIMI research establish a new “Agency Efficiency Principle.” This principle suggests that machine autonomy emerges from the strategic curation of high-quality agentic demonstrations, not from data abundance. As industries transition from AI that merely thinks to AI that actively works, LIMI provides a sustainable paradigm for cultivating truly agentic intelligence. It underscores that mastering agency requires understanding its essence and carefully crafting learning experiences, rather than simply scaling up training data.
