WEBVTT

1
00:00:00.320 --> 00:00:06.820
A steering knob can make a language model express a trait more
strongly. But what does a setting of one actually mean?
一个调节旋钮可以让语言模型更强烈地表达某种特质。但设置为一，究竟代表什么？

2
00:00:07.320 --> 00:00:11.160
The same internal strength can produce different behavior across
traits and models.
相同的内部干预强度，在不同特质和模型上可能产生不同的行为。

3
00:00:11.700 --> 00:00:14.880
PersonaDose asks for something more useful: choose a trait,
PersonaDose 提供更有用的控制方式：选择一种特质，

4
00:00:15.140 --> 00:00:17.360
then request a measured average intensity.
再指定可测量的平均表达强度。

5
00:00:18.060 --> 00:00:23.080
We will use the paper's sycophancy experiment, which measures a
model's tendency to agree with the user.
我们将使用论文中的迎合实验，它衡量模型赞同用户的倾向。

6
00:00:23.700 --> 00:00:29.240
A request of sixty means a target mean on the study's behavioral
rubric, not sixty percent of answers.
请求六十分，是指该研究行为量表上的目标平均分，而不是百分之六十的回答。

7
00:00:29.860 --> 00:00:34.740
The challenge has two parts: learn a useful behavioral range,
then calibrate access to it.
这个挑战有两个部分：先学习一个有用的行为范围，再校准对这个范围的控制。

8
00:00:35.300 --> 00:00:41.460
The base language model stays frozen. A learned controller
transforms its hidden activations during generation.
基础语言模型保持冻结，学习得到的控制器在生成过程中变换模型的隐藏激活。

9
00:00:42.040 --> 00:00:47.440
A description tells the controller which trait to express; flow
time determines the intervention strength.
特质描述告诉控制器要表达哪种特质；流时间决定干预强度。

10
00:00:48.100 --> 00:00:51.980
Flow time is a controller parameter, not the time the viewer
waits for an answer.
流时间是控制器的参数，而不是用户等待回答的时长。

11
00:00:52.800 --> 00:00:56.700
Within one base model, the seven evaluated traits share the same
controller.
同一基础模型中的七种被评估特质共享一个控制器。

12
00:00:57.300 --> 00:00:59.100
Different base models have separate controllers.
不同基础模型使用各自的控制器。

13
00:00:59.840 --> 00:01:05.760
The controller learns from trait-expressing responses, without
pairing each response with a requested intensity label.
控制器从表达特质的回答中学习，但不会把每个回答与指定强度标签配对。

14
00:01:06.280 --> 00:01:09.780
That separation matters: response training supplies the
behavior,
这种分离很关键：回答训练提供行为能力，

15
00:01:09.980 --> 00:01:12.800
while calibration later assigns meaningful score units.
后续校准再赋予控制参数有意义的分数单位。

16
00:01:13.487 --> 00:01:19.527
First, the authors try a grid of controller settings on
calibration questions and measure mean trait expression.
首先，作者在校准问题上尝试一组控制器设置，并测量平均特质表达分数。

17
00:01:20.247 --> 00:01:24.027
A monotone fitted curve summarizes how measured expression
changes with strength,
单调拟合曲线概括实测表达分数如何随强度变化，

18
00:01:24.307 --> 00:01:26.347
with interpolation between sampled settings.
并在采样设置之间进行插值。

19
00:01:27.227 --> 00:01:32.207
To serve a target score, the method works backward through that
curve to select an internal setting.
要实现目标分数，方法会沿着这条曲线反向查找，选择对应的内部设置。

20
00:01:32.927 --> 00:01:37.187
If several settings share a fitted score, it chooses the
smallest strength on that plateau.
如果多个设置对应同一个拟合分数，它会选择该平台区间上最小的强度。

21
00:01:37.987 --> 00:01:41.067
A target outside the fitted range is marked unreachable before
testing.
拟合范围之外的目标会在测试前被标记为不可达。

22
00:01:41.887 --> 00:01:44.247
A reachable setting is fixed for held-out questions.
可达目标的设置固定用于留出问题。

23
00:01:44.967 --> 00:01:50.887
This avoids selecting the best-looking setting after seeing the
test responses. The mapping is established beforehand.
这样可以避免看过测试回答后，再挑选看起来最好的设置。映射关系是在测试前建立的。

24
00:01:51.413 --> 00:01:57.933
The study examines seven traits in Qwen. Here is one calibrated
request: sycophancy at sixty.
这项研究考察 Qwen 的七种特质。这里来看一个校准请求：迎合程度六十分。

25
00:01:58.773 --> 00:02:03.673
The fitted map selects flow time one point four eight six four
for the initial controller.
对于初始控制器，拟合映射选择了流时间一点四八六四。

26
00:02:04.333 --> 00:02:09.393
With that setting fixed, held-out responses average fifty-six
point two five on the trait scale.
固定这个设置后，留出回答在特质量表上的平均分为五十六点二五。

27
00:02:10.093 --> 00:02:16.373
The table reports an absolute error of three point seven five.
This describes a measured average, not every answer.
表格报告的绝对误差为三点七五分。这描述的是实测平均值，而不是每个回答。

28
00:02:17.153 --> 00:02:22.033
Each plotted cell averages ten held-out questions, with ten
sampled responses per question.
每个绘制出来的单元格都对十个留出问题求平均，每个问题采样十个回答。

29
00:02:22.813 --> 00:02:27.513
A calibrated request therefore connects an internal intervention
to a behavioral measurement
因此，校准请求将内部干预与行为测量连接起来，

30
00:02:27.553 --> 00:02:29.173
that can be checked on new questions.
这种测量可以在新问题上检验。

31
00:02:29.953 --> 00:02:36.573
The same Qwen controller serves four requested intensities:
twenty, forty, sixty, and eighty.
同一个 Qwen 控制器提供四种请求强度：二十、四十、六十和八十。

32
00:02:37.433 --> 00:02:41.733
Their held-out means are twelve point nine two, thirty-seven
point one three,
对应的留出平均分分别为十二点九二、三十七点一三，

33
00:02:42.353 --> 00:02:45.053
fifty-six point two five, and eighty point five.
五十六点二五和八十点五。

34
00:02:45.673 --> 00:02:50.493
Only the selected steering setting changes. These requests use
the same controller weights.
变化的只有选定的干预设置，这些请求使用相同的控制器权重。

35
00:02:51.413 --> 00:02:56.953
Mean coherence in this example ranges from eighty-eight point
nine two to ninety-two point nine one.
这个示例中的平均连贯性分数介于八十八点九二和九十二点九一之间。

36
00:02:57.853 --> 00:03:00.293
Coherence measures intelligibility under the rubric.
连贯性按照量表衡量回答是否可理解。

37
00:03:00.753 --> 00:03:04.033
It is separate from how closely the requested trait score is
reached.
它与特质分数接近请求值的程度是不同指标。

38
00:03:04.853 --> 00:03:10.313
The pattern illustrates graded access, while the gaps between
requested and realized values remain visible.
这个变化趋势展示了分级控制，同时也保留了请求值与实际值之间的差距。

39
00:03:10.947 --> 00:03:15.487
Across three models, the study tests seven trained traits and
four targets,
研究在三个模型上测试七种已训练特质和四个目标，

40
00:03:15.707 --> 00:03:18.007
giving twenty-eight possible requests per model.
因此每个模型有二十八个可能的请求。

41
00:03:18.706 --> 00:03:21.727
Llama reaches twenty-two requests, Qwen twenty,
Llama 有二十二个可达请求，Qwen 有二十个，

42
00:03:22.067 --> 00:03:24.667
and Gemma fourteen within their calibration ranges.
Gemma 有十四个，均在各自校准范围内。

43
00:03:25.287 --> 00:03:31.227
Mean targeting errors over those reachable requests are six
point one, six point two, and four point seven.
在这些可达请求上，平均目标误差分别为六点一、六点二和四点七分。

44
00:03:31.727 --> 00:03:34.907
Gemma's smaller error accompanies a narrower supported range.
Gemma 的较小误差对应着更窄的支持范围。

45
00:03:35.407 --> 00:03:38.187
The error alone cannot tell you how many requests are available.
仅看误差，无法知道有多少请求可用。

46
00:03:39.027 --> 00:03:44.547
The requests also meeting the held-out coherence floor number
nineteen, twenty, and thirteen, respectively.
同时满足留出连贯性下限的请求数，分别为十九、二十和十三。

47
00:03:45.307 --> 00:03:50.947
Useful dosing needs these three checks together: reachability,
targeting error, and coherent expression.
有用的分级控制需要同时检查三个方面：可达性、目标误差，以及表达的连贯性。

48
00:03:51.500 --> 00:03:57.860
The map belongs to a specific controller. Continued training can
change the relationship between flow time and behavior.
映射属于某个特定控制器。继续训练可能改变流时间与行为之间的关系。

49
00:03:58.540 --> 00:04:04.420
In the paper's separate Qwen follow-up, the base language model
still stays frozen while the controller is updated.
在论文另行开展的 Qwen 后续实验中，控制器被更新，而基础语言模型仍保持冻结。

50
00:04:05.100 --> 00:04:09.600
For sycophancy at sixty, the initial map selects one point four
eight six four;
对于六十分的迎合目标，初始映射选择一点四八六四；

51
00:04:10.120 --> 00:04:13.780
the continued controller's map selects zero point eight seven
four five.
继续训练后控制器的映射选择零点八七四五。

52
00:04:14.480 --> 00:04:19.020
The same behavioral request therefore uses a different internal
strength after the update.
因此，更新后，同一个行为请求会使用不同的内部强度。

53
00:04:19.860 --> 00:04:23.480
The authors remeasure the curve before serving requests with the
updated controller.
作者会在使用更新后的控制器提供请求之前，重新测量这条曲线。

54
00:04:24.300 --> 00:04:29.220
Calibration makes the interface interpretable, but it must
follow the system whose behavior it measures.
校准让控制接口具有可解释性，但它必须与所测量行为的系统保持对应。

55
00:04:29.800 --> 00:04:34.360
PersonaDose separates learning a trait from specifying how much
of it to express.
PersonaDose 将学习一种特质，与指定表达多少这种特质分开处理。

56
00:04:34.900 --> 00:04:41.240
A shared controller provides the behavior. A measured map
translates requested score units into its internal strength.
共享控制器提供行为能力，实测映射则把请求的分数单位转换为内部干预强度。

57
00:04:41.940 --> 00:04:46.540
The Qwen example makes the contract concrete: request sixty,
choose the calibrated setting,
Qwen 示例让控制关系具体化：请求六十分，选择校准设置，

58
00:04:46.920 --> 00:04:49.820
then measure fifty-six point two five on held-out responses.
再在留出回答上测得五十六点二五分。

59
00:04:50.520 --> 00:04:55.300
That is control you can inspect through its range, its targeting
error, and its coherence.
这种控制可以通过范围、目标误差和连贯性来检查。

60
00:04:55.940 --> 00:05:02.320
The study concerns judged mean behavior on trained traits and
short responses; individual answers can still differ.
这项研究针对已训练特质和短回答上的评判平均行为；单个回答仍可能有所不同。

61
00:05:02.860 --> 00:05:07.980
A meaningful steering interface starts when an internal knob
becomes a measurable behavioral request.
当内部旋钮变成可测量的行为请求时，干预控制接口才开始具有明确意义。