New benchmark tests AI agents' ability to learn tools by trial and error

Original: ScrambleToolBench: Agents Search Exhaustively Even When Their Own Map Points to the Next Step

Source: arxiv.org ↗