AI researchers release benchmarks for agents and reliability (opens in new tab)

Research posted May 15 focused on whether large language model systems can remain reliable when they retrieve evidence, use tools, remember prior interactions or operate through multi-agent harnesses. Microsoft Research published a clarification of its recent work on AI delegation and long-horizon reliability, saying the paper “LLMs Corrupt Your Documents When You Delegate” aims to develop evaluation methods for delegated workflows and clarify what it does and does not claim. New papers intro...