实测:选项顺序会主导 Jev 的判断
作者跑 eval 发现证据模糊时选项排列顺序会变成隐式策略,证据足够区分后顺序效应消失,问题不在幻觉而在状态信息不足。
One of the evals I ran looked at the effect of choice position on Jev, the long and short of it is that: choice order dominates when evidence is ambiguous In the below example, it's clear that the state contains very little decision-relevant evidence for distinguishing between information and bug report. We know that the user is interested in the annual plan and wants to know whether it supports 5 users, but we also know the page isn't loading. Both choices are therefore supported (bug + information) but nothing in the provided state tells JEV which signal should dominate Once you add enough evidence to clearly discriminate between the choices, the order effect disappears. Overall, the issue isn't that JEV is hallucinating, it's that the state doesn't contain enough decision-relevant information to support the distinction you're asking it to make otherwise choice order becomes an implicit policy I explain this in more detail here: https://x.com/parcadei/status/2101077277622595599
这条是引用着另一条写的:x.com/neural_avb/status/2101736546391244854