VentureBeat
Follow
New AI training method creates powerful software agents with just 78 examples
A new study introduces LIMI, a framework demonstrating that less data can lead to more intelligent AI agents. Researchers found that high-quality agentic demonstrations, not massive datasets, are key to developing autonomous AI systems. LIMI's approach focuses on curating strategically chosen examples of AI problem-solving. Experiments showed that a model trained on just 78 carefully selected demonstrations significantly outperformed models trained on thousands of examples. This discovery is crucial for enterprise applications where data collection is challenging and expensive. Agency is defined as AI systems that can autonomously discover problems, formulate hypotheses, and execute solutions. Current LLM training often assumes more data equates to higher agentic intelligence, leading to complex pipelines and resource demands. LIMI's method involves collecting queries and detailed trajectories of AI actions to solve them. The dataset was built using real-world scenarios and synthesized queries, vetted by human experts. Trajectories captured the full problem-solving process, including errors and refinements. LIMI-trained models achieved superior scores on benchmarks for agentic skills, tool use, and coding. The study suggests mastering agency is about understanding its core principles, not just data scaling. This framework offers a sustainable path for developing specialized AI agents for businesses.
