Idea
A training approach enabling AI agents to achieve superior autonomous task execution with minimal curated demonstrations for enterprises.
Research Paper
Core Innovation
This paper introduces LIMI, demonstrating that agentic intelligence emerges from strategically curated minimal data rather than large-scale datasets. It overturns traditional scaling laws by achieving higher agency performance with 128 times fewer samples. The Agency Efficiency Principle guides efficient cultivation of machine autonomy through quality over quantity in training data.
Market Size (TAM)
$20–50B TAM for autonomous AI agents; $2–10B SAM from enterprises in software development, scientific research, and automation. Driven by demand for efficient AI task execution and reduced training costs.
Potential Customers & Pain Points
- Enterprises needing efficient autonomous AI agents
- AI developers seeking data-efficient training methods
- Industries requiring reliable AI task execution
- Research labs focused on agentic intelligence benchmarks
Business Model
Licensing the LIMI training framework and curated datasets to AI developers and enterprises; offering consulting for integration and customization.
Competitive Landscape
- Kimi-K2-Instruct
- DeepSeek-V3.1
- Qwen3-235B-A22B-Instruct
Implementation Challenges
- Adoption resistance to minimal data training paradigms
- Integration with existing AI workflows
- Validation across diverse real-world tasks
Validation Strategy
- Benchmark LIMI on diverse autonomous task datasets
- Pilot deployments with enterprise partners in software and research workflows
- Iterate training samples based on real-world feedback
Research Paper Overview
LIMI: Less is More for Agency
Summary
This paper defines Agency as AI systems' autonomous capacity to discover problems, hypothesize, and execute solutions through self-directed engagement. It challenges the traditional scaling law that more data improves agency, showing that strategic curation of minimal, high-quality demonstrations can yield superior agentic intelligence. LIMI achieves 73.5% on agency benchmarks using only 78 training samples, outperforming state-of-the-art models trained on thousands of samples. The findings establish the Agency Efficiency Principle, emphasizing quality over quantity in training data for machine autonomy.