WEBVTT

1
00:00:00.480 --> 00:00:04.720
Imagine an agent repeatedly turning a measurement table into a
short report.
设想智能体反复将测量表格整理为短报告。

2
00:00:05.320 --> 00:00:08.360
Its registry also contains many unrelated tools.
其工具库也包含许多无关工具。

3
00:00:09.140 --> 00:00:14.700
Each request pays for that menu in context, and the model must
decide which tools matter.
每个请求都要为这份工具菜单占用上下文，模型还必须判断哪些工具有用。

4
00:00:15.800 --> 00:00:22.040
LOTS asks whether earlier output experience can teach the system
which tools to keep for this recurring task.
LOTS 研究系统能否从此前的输出经验中学习，为这类重复任务保留哪些工具。

5
00:00:22.880 --> 00:00:28.020
It changes a persistent task-specific tool space while keeping
the model parameters fixed.
它改变的是持久的任务专属工具空间，而模型参数保持不变。

6
00:00:28.740 --> 00:00:35.060
The method's name is Likelihood-Only Tool Scoring. Its evidence
comes from what tools actually produced.
这个方法叫作“仅基于似然的工具评分”，它依据的是工具实际产生的输出。

7
00:00:35.720 --> 00:00:41.480
We will follow one illustrative report, then see how a
single-request measurement becomes a reusable deployment choice.
我们先看一份示意性报告，再看看单次请求的测量如何转化为可复用的部署选择。

8
00:00:42.427 --> 00:00:48.047
Suppose a table reader returns twenty degrees Celsius and a
conversion tool returns sixty-eight degrees Fahrenheit.
假设表格读取工具返回二十摄氏度，而换算工具返回六十八华氏度。

9
00:00:48.527 --> 00:00:54.067
The agent writes that the reading is sixty-eight degrees
Fahrenheit. LOTS now freezes that realized answer.
智能体写道，这个读数为六十八华氏度。此时，LOTS 固定这份已生成的回答。

10
00:00:54.587 --> 00:00:59.527
It records the context that supported the answer, including tool
calls and observed outputs.
它记录支持这份回答的上下文，包括工具调用和观察到的输出。

11
00:01:00.167 --> 00:01:05.107
Then it removes the evidence attributable to one tool and scores
the very same answer again.
然后，它移除归属于某个工具的证据，并对完全相同的回答再次评分。

12
00:01:05.767 --> 00:01:11.167
The comparison does not ask the agent to generate a different
answer or judge whether the first answer was correct.
这个对比并不要求智能体重新生成另一份回答，也不评判原回答是否正确。

13
00:01:11.707 --> 00:01:16.787
It asks how strongly the fixed model supports those answer
tokens under the changed context.
它检查的是，上下文改变之后，固定模型对这些回答词元的支持有多强。

14
00:01:17.400 --> 00:01:21.360
The score is the answer's mean log-likelihood with the recorded
evidence,
分数是保留记录证据时回答的平均对数似然，

15
00:01:21.860 --> 00:01:24.700
minus its mean log-likelihood after the intervention.
减去干预之后回答的平均对数似然。

16
00:01:25.280 --> 00:01:32.000
For a purely illustrative calculation, suppose those values are
minus zero point four and minus zero point nine.
以一个纯粹的示意计算为例，假设这两个值分别为负零点四和负零点九。

17
00:01:32.460 --> 00:01:37.480
The difference is positive zero point five: removing that
evidence made the same answer less likely.
两者之差为正零点五：移除那份证据后，相同回答的似然降低了。

18
00:01:38.320 --> 00:01:40.260
Other tools remain part of the scoring context.
其他工具仍在评分上下文中。

19
00:01:40.920 --> 00:01:44.680
Benchmark-specific interventions also account for downstream
tool dependencies.
针对不同基准的干预也处理后续工具依赖。

20
00:01:45.300 --> 00:01:49.000
A positive contribution means support for this realized output.
正贡献表示支持这份已生成输出。

21
00:01:49.420 --> 00:01:51.960
It is not a correctness certificate for the tool or answer.
它不是对工具或回答正确性的认证。

22
00:01:52.440 --> 00:01:57.280
The model's reliance on evidence is the signal being collected,
with no correctness labels used by this scoring step.
所收集的信号是模型对证据的依赖，这个评分步骤不使用正确性标签。

23
00:01:57.893 --> 00:02:04.613
One execution is only one observation. LOTS aggregates
contributions across scoreable requests of the same task type.
一次执行只是一次观察，LOTS 会汇总同类任务中可评分请求的贡献。

24
00:02:05.493 --> 00:02:11.233
Each tool's priority is its average contribution over requests
where that tool supplies usable scoring evidence.
每个工具的优先级，是它在提供可用评分证据的请求上的平均贡献。

25
00:02:12.753 --> 00:02:16.373
A tool that was not exposed on a request contributes no evidence
from that request.
如果某个工具没有出现在一次请求中，这次请求就不会为它提供证据。

26
00:02:17.933 --> 00:02:22.793
The tools are ranked, and a top-budget subset becomes the stored
tool space for later requests.
工具被排序后，预算范围内排名靠前的子集成为后续请求使用的已存储工具空间。

27
00:02:24.253 --> 00:02:26.933
There is a second decision here: how many tools to keep.
还需另一个决定：保留多少工具。

28
00:02:28.073 --> 00:02:31.253
In the reported experiments, fitting labels help select that
budget.
报告实验使用拟合标签帮助选择这个预算。

29
00:02:31.833 --> 00:02:38.573
So likelihood ranking is label-free; the experimental
budget-selection procedure is not entirely label-free.
因此，似然排序不需要标签，但实验中的预算选择程序并非完全不使用标签。

30
00:02:39.513 --> 00:02:45.793
The selected space is deployment state attached to a recurring
task, rather than a fresh search for each request.
选中的工具空间是附属于重复任务的部署状态，而不是每次请求都重新搜索。

31
00:02:46.593 --> 00:02:51.433
Later requests reuse that menu, avoiding another round of
selection over the full registry.
后续请求复用这份菜单，从而避免再对整个工具库进行一次选择。

32
00:02:52.353 --> 00:02:57.673
If new output traces arrive, the same measurement and
aggregation procedure can update the stored priorities.
如果有新的输出轨迹，相同的测量和汇总程序可以更新已存储的优先级。

33
00:02:58.513 --> 00:03:04.733
A newly arriving task gets its own space; creating that space
does not overwrite the earlier task's configuration.
新出现的任务会获得自己的工具空间；创建它不会覆盖已有任务的配置。

34
00:03:05.653 --> 00:03:09.393
The paper also tests rankings fitted with one model and served
by another.
论文还测试了由一个模型拟合排序，再交给另一个模型使用的情况。

35
00:03:10.093 --> 00:03:14.733
On BFCL, transferred rankings outperform the serving model's
full-menu baseline
BFCL 中，迁移排序优于执行模型的全菜单基线，

36
00:03:14.813 --> 00:03:17.153
in eighteen of twenty cross-model pairs.
二十组跨模型配对中有十八组如此。

37
00:03:17.893 --> 00:03:22.433
The controlled TGB benchmark contains four thousand requests,
ten task families,
受控 TGB 基准有四千请求与十个任务族，

38
00:03:22.553 --> 00:03:24.853
and fifteen tools, including distractors.
以及十五个工具，其中包括干扰工具。

39
00:03:25.593 --> 00:03:29.593
Each family uses eighty fitting requests and three hundred
twenty evaluation requests.
每个任务族使用八十个拟合请求和三百二十个评估请求。

40
00:03:30.353 --> 00:03:35.813
For Qwen three, the eight-billion-parameter model, serving all
tools gives fifty-six point five percent accuracy;
Qwen3-8B 使用所有工具时准确率为百分之五十六点五；

41
00:03:36.373 --> 00:03:38.293
LOTS reaches eighty-one point five.
LOTS 达到百分之八十一点五。

42
00:03:38.873 --> 00:03:45.153
That is a twenty-five-point gain in this controlled setting. All
eight evaluated TGB models improve over the full menu.
在这个受控设置中，提升为二十五个百分点；八个被评估的 TGB 模型都优于完整菜单。

43
00:03:46.033 --> 00:03:48.613
BFCL measures complete multi-turn call sequences,
BFCL 衡量完整多轮调用序列，

44
00:03:48.993 --> 00:03:52.213
while GTA-Atomic evaluates executable ReAct tool use.
GTA-Atomic 评估可执行 ReAct 工具使用。

45
00:03:53.053 --> 00:03:56.053
On GTA-Atomic, LOTS improves six of seven models,
GTA-Atomic 中，LOTS 改善七模型中的六个，

46
00:03:56.513 --> 00:03:58.613
so the measured gains are broad but not universal.
所以收益较广，但非普遍成立。

47
00:03:59.393 --> 00:04:05.793
A persistent ranking separates construction work from serving
later requests. Both costs matter when judging efficiency.
持久排序把构建工作与后续请求的执行分开；判断效率时，两部分成本都很重要。

48
00:04:06.413 --> 00:04:12.873
On TGB, task-specific pruning reduces serving prompt tokens by
thirty-three to forty-six percent
TGB 任务剪枝减少执行提示词元百分之三十三至四十六，

49
00:04:13.193 --> 00:04:14.733
relative to all fifteen tools.
相比使用全部十五个工具。

50
00:04:15.413 --> 00:04:19.253
For one GTA-Atomic task, the paper counts tool-selection tokens
对一个 GTA-Atomic 任务，论文将工具选择词元，

51
00:04:19.313 --> 00:04:22.293
together with tokens used to serve the evaluation requests.
与执行评估请求的词元一并计入成本。

52
00:04:22.893 --> 00:04:28.213
Under that accounting, LOTS uses roughly one-quarter to one-half
as many tokens as TTO,
按此统计，LOTS 词元约为 TTO 的四分之一至二分之一，

53
00:04:28.613 --> 00:04:30.713
with still larger savings over beam search.
相比束搜索，节省幅度更大。

54
00:04:31.533 --> 00:04:35.933
These are scoped token comparisons, not a measured universal
wall-clock speedup.
这些是在特定范围内的词元成本对比，并不是对所有情况测得的运行时间加速。

55
00:04:36.433 --> 00:04:41.313
The practical benefit comes from carrying useful tool-selection
evidence forward across recurring work.
实际收益来自把有用的工具选择证据保留下来，服务于重复出现的工作。

56
00:04:41.940 --> 00:04:47.500
LOTS turns observed tool outputs into priorities for a
persistent task-specific action space.
LOTS 把观察到的工具输出，转化为持久任务专属行动空间中的优先级。

57
00:04:48.080 --> 00:04:53.420
The fixed-answer comparison makes the mechanism inspectable:
what changes when one tool's evidence disappears?
固定回答的对比让机制可以被检查：当一个工具的证据消失时，什么发生了变化？

58
00:04:54.420 --> 00:04:58.640
Aggregation asks whether that evidence keeps supporting later
examples of the same kind of task.
汇总检查的是，这份证据是否持续支持同类任务的后续示例。

59
00:04:59.240 --> 00:05:04.360
Deployment reuses the smaller menu, and new traces can revise it
without updating the underlying model.
部署阶段复用更小的菜单，新轨迹可以修改菜单，而不更新底层模型。

60
00:05:05.460 --> 00:05:10.880
Likelihood scoring, budget selection, and task correctness
remain distinct decisions throughout the explanation.
在整个解释中，似然评分、预算选择，以及任务正确性，始终是不同的判断。

61
00:05:11.560 --> 00:05:13.680
The result is a concrete route to adaptation:
结果提供具体适应路径：

62
00:05:14.120 --> 00:05:18.840
retain the tools whose outputs support recurring work, then test
the resulting configuration.
保留支持重复工作输出的工具，再测试最终配置。