Published on July 5, 2026, the entry describes HASTE, a hierarchical multi-agent system. In an ablation with 159 skills across eight competitions, hierarchical loading achieved a 100% medal rate, compared with 62.5% for flat loading.
HASTE organizes reusable skills for machine-learning engineering agents into three levels: global, domain, and competition-specific. In the reported ablation, involving 159 skills across eight competitions, hierarchical loading achieved a 100% medal rate; flat loading achieved 62.5%. The entry does not provide the medal distribution or other quantitative results here.
The authors suggest that how accumulated skills are delimited and loaded can affect competition performance and token use. To assess the comparison, consult the original publication associated with the entry and check the ablation methodology, competition conditions, and measurement criteria; the reported figures alone do not show that the approach will have the same effect on other tasks.