Published on September 23, 2025, the LIMI report says 78 selected demonstrations achieved 73.5% on AgencyBench, outperforming models trained on 10,000 samples.
The post about LIMI, published on September 23, 2025, describes a study of data curation for agent autonomy. According to the report, the system uses 78 selected demonstrations and scores 73.5% on the AgencyBench benchmark, outperforming models trained on 10,000 samples. The text argues that strategic data selection, not just data scale, may drive autonomy. These figures raise a testable hypothesis, but they are not, by themselves, enough to establish a general conclusion about agent training.
To assess the result, consult the original publication and check how demonstrations were selected, which models and comparison conditions were used, and how performance was measured. The available claim does not detail these procedures; do not treat it as proof that smaller datasets are always better. If you use AI to study or reproduce the work, avoid entering internal or confidential data without authorization and follow your organization's data protection policies.