目标：连续发起“总结论文、全文重点、解释选区、自定义提问”后，可以知道哪些结果完成但还没看；点击后直接跳转到对应回答并高亮。
While most robotic learning systems are evaluated in environments that closely match the training data...
a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
We study generalization across new homes, objects, and task instructions.
While most robotic learning systems are evaluated in environments that closely match the training data...
a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
We study generalization across new homes, objects, and task instructions.
While most robotic learning systems are evaluated in environments that closely match the training data...
a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
We study generalization across new homes, objects, and task instructions.
While most robotic learning systems are evaluated in environments that closely match the training data...
a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
We study generalization across new homes, objects, and task instructions.
While most robotic learning systems are evaluated in environments that closely match the training data...
a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
We study generalization across new homes, objects, and task instructions.
While most robotic learning systems are evaluated in environments that closely match the training data...
a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
We study generalization across new homes, objects, and task instructions.
While most robotic learning systems are evaluated in environments that closely match the training data...
a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
We study generalization across new homes, objects, and task instructions.
While most robotic learning systems are evaluated in environments that closely match the training data...
a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
We study generalization across new homes, objects, and task instructions.
While most robotic learning systems are evaluated in environments that closely match the training data, a number of prior works have explored broader open-world generalization.
Training such generalist policies requires new modeling approaches to handle scale and diversity.
Visual-language-action models provide an attractive solution by fine-tuning pretrained vision-language models for robot control.
We study generalization across new homes, objects, and task instructions.
先做“未读 N + 查看跳转 + 高亮”，再做完整队列弹窗；这样改动小，能快速验证体验。
任务阅读状态建议本地保存，不走坚果云；聊天记录仍按现有逻辑保存，避免多设备阅读状态互相干扰。
必须全部已读且没有回答中任务才可点击；点击后直接清空，不需要确认。
排队中任务直接移除；回答中任务先 abort 当前请求，再标记为已取消，可单独移除。
解释选区、翻译选区、手动选中文字提问都记录 attachmentID、pageIndex、rects/sortIndex、原选区文本；查看时同时跳转 PDF 和聊天回答。