Agent isn’t the key: RUC's AiScientist shows 23 hours and 74 rounds of long-horizon memory

Agent不是关键！人大AiScientist实现23小时、74轮长程记忆

A Renmin University of China team released AiScientist, which ran 23 hours and 74 experiment loops on MLE-Bench Lite Detecting Insults, raising validation AUC from 0.903 to 0.982 with 18 best-so-far updates. The paper says its core is File-as-Bus, which persists analysis, code, logs, and results in the workspace; removing it drops PaperBench by 6.41 points and MLE-Bench Lite Any Medal by 31.82 points. The real lever here is state continuity, not simply adding more agents.

Why it matters: HKR-H lands because the title flips a live assumption: memory continuity, not more agents. HKR-K lands on the 23h/74-run setup, AUC 0.903→0.982, and ablations; HKR-R lands because builders are debating multi-agent stacks vs durable state.