Meta 论文 ProPerMem：主动式 AI 助手卡在'何时开口'而非'说什么'
Meta 拿 AI 眼镜场景做了个新基准，GPT-5.3-Chat 有答案也才 38 分，问题出在自己挑时机说话上，方法部分挺有参考价值。
Meta 发布 ProPerMem 基准，用第一视角视频配平均 618K token 的年度记忆库，测试主动式助手自主决定打断时机的能力。所有基线模型得分不超过 15%，包括带检索增强记忆的前沿模型。GPT-5.3-Chat 在有人工给定证据时达 96%，但自主选时机降到 38%，用检索进一步降到 15%。论文提出同时把控内容和时机、少触发的策略，把准确率从 10% 以下提升到 42.4%。
New Meta paper shows that deciding on its own when to interrupt, not just what to say, is where proactive assistants fail.
Proactive assistants, like AI glasses, get no query. They must decide when to speak and what to say by finding the justifying fact in a year of personal memory, not handed to them.
They built ProPerMem: egocentric video paired with a distractor-rich, year-long memory averaging 618K tokens, with human-verified trigger windows. A clip counts only if the assistant speaks just when due inside the window with grounded content.
Every baseline scored at most 15%, even frontier models with retrieval-augmented memory. With oracle evidence in the window GPT-5.3-Chat reached 96%, but choosing its own timing dropped it to 38%, and using retrieval dropped it to 15%.
For proactive assistants, get both the content and the timing right and fire sparingly, which lifted accuracy from below 10% to 42.4%.
