TURTLE: Evolutionary Search Auto-Generates Hard Tool-Use Benchmarks

A paper from Natasha Jaques' team introduces TURTLE, an unsupervised environment design method that uses LLM multi-agent evolutionary search to auto-generate challenging tool-use tasks, raising the share of hard AppWorld problems from 19.8% to 52.5%.

2026-10-24 ~ 2026-10-24 · 2 related posts

• TURTLE paper auto-evolves tool-use benchmarks, boosting hard tasks on AppWorld by 165% — MichaelD1729 · 2026-10-24
• Evolutionary search auto-generates agent tool-use tasks, lifting AppWorld hard-task share to 52.5% — minsuk_chang · 2026-10-24