EvasionBench is a new, very large dataset that helps computers spot when company leaders dodge questions during earnings call Q&A.
This paper shows how to make home-helper robots better at long, multi-step chores by smart training on diverse tasks and by polishing the model after training using its own best attempts.