WEBVTT

1
00:00:00.360 --> 00:00:03.420
Imagine someone asking: help me plan a team dinner.
想象有人提出：帮我安排一次团队聚餐。

2
00:00:04.019 --> 00:00:07.140
The request sounds simple, until you remember the background.
这个请求听起来很简单，直到你考虑它的背景。

3
00:00:07.640 --> 00:00:12.460
Suppose their team includes a vegetarian colleague, and one
member needs step-free access.
假设团队有素食者，还有一位成员需要无台阶通道。

4
00:00:12.920 --> 00:00:16.700
An attractive restaurant list could still miss what makes the
plan useful.
即使餐厅列表看起来很不错，也可能遗漏让计划真正有用的条件。

5
00:00:17.500 --> 00:00:21.080
This is an illustrative request, rather than an example of a
real participant.
这是一个用于说明的请求，而非真实参与者的案例。

6
00:00:22.080 --> 00:00:27.220
xDailyBench studies this difference between following the words
and satisfying the task.
xDailyBench 研究遵循文字要求与满足实际任务之间的差异。

7
00:00:27.833 --> 00:00:32.533
The benchmark starts from requests people have used AI for, or
genuinely intended to complete.
基准从人们实际使用 AI 完成、或真正想完成的请求出发。

8
00:00:33.093 --> 00:00:38.013
Each task supplies an instruction, optional workspace materials,
and relevant user context.
每个任务提供指令、可选的工作区材料，以及相关的用户背景。

9
00:00:38.693 --> 00:00:44.673
Some requirements are explicit. Others reflect basic
expectations, such as accuracy or usable formatting.
有些要求是明确的，另一些则体现准确性或可用格式等基本期待。

10
00:00:45.373 --> 00:00:49.073
A third group depends on the particular user and the context
supplied with the task.
还有一类要求取决于具体用户，以及任务中提供的背景。

11
00:00:49.713 --> 00:00:53.833
The agent needs to connect those pieces, rather than invent a
background it was never given.
智能体需要连接这些信息，而不是虚构并未获得的背景。

12
00:00:54.553 --> 00:00:58.513
The final collection includes two hundred forty-eight tasks
across fifty-one scenarios.
最终任务集包含二百四十八个任务，覆盖五十一个场景。

13
00:00:59.127 --> 00:01:04.967
An open-ended task can have several good answers. The rubric
defines what those answers must accomplish.
开放任务可以有多个好答案；评分准则定义它们必须完成什么。

14
00:01:05.507 --> 00:01:11.507
For the dinner illustration, a plan might satisfy dietary needs,
access needs, and a practical schedule.
在聚餐示例中，计划可以满足饮食需求、通行需求，以及可执行的时间安排。

15
00:01:12.047 --> 00:01:16.307
These are separate criteria, so a judge can inspect each
requirement individually.
这些是独立的条件，因此评估者可以逐项检查要求。

16
00:01:17.087 --> 00:01:20.627
Each criterion receives a binary judgement: met or not met.
每项条件接受二元判断：满足或未满足。

17
00:01:21.087 --> 00:01:25.427
The criteria carry weights, reflecting their relative importance
in completing the task.
各项条件带有权重，反映它们对完成任务的相对重要性。

18
00:01:26.127 --> 00:01:31.567
The resulting score can credit useful progress without
pretending the whole task was completed perfectly.
由此得到的分数可以认可有用进展，而不会把它当成完美完成整个任务。

19
00:01:32.133 --> 00:01:37.753
The paper uses assisted screening, then manual cross-review, to
refine the tasks and their criteria.
论文先进行辅助筛选，再通过人工交叉审阅完善任务及条件。

20
00:01:38.393 --> 00:01:44.413
Reviewers check whether each criterion is atomic, clear, and
suitable for a binary judgement.
审阅者检查每项条件是否独立、明确，并适合二元判断。

21
00:01:45.193 --> 00:01:49.133
A pilot compares automatic and human judgements on the same
outputs and rubrics.
试验阶段使用相同输出和评分准则，比较自动判断与人工判断。

22
00:01:49.893 --> 00:01:53.633
Criteria that miss the required agreement threshold are reviewed
and revised.
未达到所需一致性门槛的条件会被重新审阅和修改。

23
00:01:54.333 --> 00:01:59.513
The final experiments compare eleven models with matched
instructions, tools, and execution budgets.
最终实验在匹配的指令、工具与执行预算下比较十一个模型。

24
00:02:00.293 --> 00:02:05.493
The important object is the full response or deliverable,
assessed against the task requirements.
关键评估对象是完整的回答或交付物，并对照任务要求判断。

25
00:02:06.173 --> 00:02:11.193
In the reported results, two models tie at a task-level score of
seventy-five point six percent.
报告的结果中，两个模型并列达到百分之七十五点六的任务级得分。

26
00:02:11.753 --> 00:02:17.433
That score averages task scores. It is not the percentage of
tasks completed without any missing criterion.
这一指标平均各任务得分，而非毫无遗漏地完成任务的比例。

27
00:02:18.473 --> 00:02:22.093
Now inspect one model, Seed two point one Evolving, at the
rubric level.
现在查看 Seed-2.1-Evolving 这一模型的评分条件级结果。

28
00:02:22.733 --> 00:02:28.273
Its explicit criteria score is eighty-one percent; its implicit
criteria score is sixty-five percent.
它在明确条件上的得分为百分之八十一，在隐含条件上为百分之六十五。

29
00:02:28.793 --> 00:02:34.253
Those are pooled criterion scores, a sixteen-percentage-point
gap within this reported comparison.
这是汇总条件后的得分，在这一比较中相差十六个百分点。

30
00:02:34.853 --> 00:02:39.293
Looking only at the overall number would hide which part of
understanding the task needs improvement.
如果只看总体数字，就会掩盖任务理解中哪些部分仍需要改进。

31
00:02:39.860 --> 00:02:46.240
Return to the dinner plan. A polished answer can still overlook
a need that was present in the supplied context.
回到聚餐计划：一个精美答案仍可能忽略所提供背景中的需求。

32
00:02:47.080 --> 00:02:52.940
xDailyBench makes those requirements explicit in evaluation,
even when they were implicit in the request.
xDailyBench 在评估中明确这些要求，即使它们在请求中只是隐含的。

33
00:02:53.760 --> 00:03:00.600
Its contribution is a way to ask whether an assistant connects
instructions, background, and a usable final result.
它提供了一种方式，检验助手能否连接指令、背景与可用的最终结果。

34
00:03:01.260 --> 00:03:06.840
The lesson is to read the whole task before deciding that the
visible words have been satisfied.
这里的启发是：在认定文字要求已经满足之前，先理解完整的任务。