Evolving existing tasks creates harder, more diverse tool-use benchmarks

TURTLE recombines tasks and tool sequences to generate evaluations calibrated to a model’s current success rate.

Big Tech
Hyunji Alex Nam · Michael Dennis · Natasha Jaques · Minsuk Chang

Google Deepmind · Stanford University · University of Washington

Research Digest··3 min read
Nam and colleagues introduce TURTLE, a multi-agent system that evolves existing benchmark tasks rather than constructing new tools and environments.

TURTLE uses three language-model agents.

§