AI agents can't research; they only confirm what they already know
a new study finds top models like kimi-k2.6 rely on memory instead of tools, failing when forced to research truly new info. this exposes a critical flaw in current agent benchmarks that confuses knowledge retrieval with research ability.
a study on eleven frontier models revealed that high performance on the browsecomp benchmark stems from 'intrinsic knowledge dependence,' where models use the web to confirm pre-existing answers rather than conduct research. when deprived of answer-supporting documents, every model performed worse than if it had no search tools at all, and mini maxim m2.5 saw accuracy plummet from 44.5% to 8.0%. to test real agent capabilities, researchers built livebrowsecomp with questions on recent events; here, models failed completely without tools and saw massive score drops with them, reshuffling leaderboards and exposing that current benchmarks measure knowledge retention, not research skill.