多智能体出错,根子常在状态没对齐。缺上下文,或每步都对却拼不成整体。论文里 harness 管预算,超支 5/5 降到 0/5。
Agent teams often fail on state alignment. Missing context, or valid steps that do not fit one world. Harness-owned budget: 5/5 to 0/5.
https://t.co/clEd85uwYp
同一病例换成患者口吻,12个开源临床模型全部掉分:1,000个改写案例平均降9.6个百分点。
In the patient's voice, all 12 open clinical models get less accurate on the same cases. Clinician-voice benchmarks miss this. Test phrasing robustness first.
https://t.co/5nAt1CJL4w