Stanford University and Google DeepMind developed TURTLE, a system that turns existing tool-use tasks into harder versions while checking that they remain solvable. The system combines a proposer agent based on AlphaEvolve, a solver model, and a feedback generator that guides successive changes. A task is considered solvable when the solver completes it at least 50% of the time across a minimum of 10 attempts. Tests on AppWorld, Ï„-Bench, and AutomationBench notably raised the share of hard tasks from 19.8% to 165% in the evolved AppWorld set, which included 735 tasks evaluated with Gemma 4 31B-it.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
