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?

2
00:00:07.320 --> 00:00:11.160
The same internal strength can produce different behavior across
traits and models.

3
00:00:11.700 --> 00:00:14.880
PersonaDose asks for something more useful: choose a trait,

4
00:00:15.140 --> 00:00:17.360
then request a measured average intensity.

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.

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.

7
00:00:29.860 --> 00:00:34.740
The challenge has two parts: learn a useful behavioral range,
then calibrate access to it.

8
00:00:35.300 --> 00:00:41.460
The base language model stays frozen. A learned controller
transforms its hidden activations during generation.

9
00:00:42.040 --> 00:00:47.440
A description tells the controller which trait to express; flow
time determines the intervention strength.

10
00:00:48.100 --> 00:00:51.980
Flow time is a controller parameter, not the time the viewer
waits for an answer.

11
00:00:52.800 --> 00:00:56.700
Within one base model, the seven evaluated traits share the same
controller.

12
00:00:57.300 --> 00:00:59.100
Different base models have separate controllers.

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.

14
00:01:06.280 --> 00:01:09.780
That separation matters: response training supplies the
behavior,

15
00:01:09.980 --> 00:01:12.800
while calibration later assigns meaningful score units.

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.

17
00:01:20.247 --> 00:01:24.027
A monotone fitted curve summarizes how measured expression
changes with strength,

18
00:01:24.307 --> 00:01:26.347
with interpolation between sampled settings.

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.

20
00:01:32.927 --> 00:01:37.187
If several settings share a fitted score, it chooses the
smallest strength on that plateau.

21
00:01:37.987 --> 00:01:41.067
A target outside the fitted range is marked unreachable before
testing.

22
00:01:41.887 --> 00:01:44.247
A reachable setting is fixed for held-out questions.

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.

24
00:01:51.413 --> 00:01:57.933
The study examines seven traits in Qwen. Here is one calibrated
request: sycophancy at sixty.

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.

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.

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.

28
00:02:17.153 --> 00:02:22.033
Each plotted cell averages ten held-out questions, with ten
sampled responses per question.

29
00:02:22.813 --> 00:02:27.513
A calibrated request therefore connects an internal intervention
to a behavioral measurement

30
00:02:27.553 --> 00:02:29.173
that can be checked on new questions.

31
00:02:29.953 --> 00:02:36.573
The same Qwen controller serves four requested intensities:
twenty, forty, sixty, and eighty.

32
00:02:37.433 --> 00:02:41.733
Their held-out means are twelve point nine two, thirty-seven
point one three,

33
00:02:42.353 --> 00:02:45.053
fifty-six point two five, and eighty point five.

34
00:02:45.673 --> 00:02:50.493
Only the selected steering setting changes. These requests use
the same controller weights.

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.

36
00:02:57.853 --> 00:03:00.293
Coherence measures intelligibility under the rubric.

37
00:03:00.753 --> 00:03:04.033
It is separate from how closely the requested trait score is
reached.

38
00:03:04.853 --> 00:03:10.313
The pattern illustrates graded access, while the gaps between
requested and realized values remain visible.

39
00:03:10.947 --> 00:03:15.487
Across three models, the study tests seven trained traits and
four targets,

40
00:03:15.707 --> 00:03:18.007
giving twenty-eight possible requests per model.

41
00:03:18.706 --> 00:03:21.727
Llama reaches twenty-two requests, Qwen twenty,

42
00:03:22.067 --> 00:03:24.667
and Gemma fourteen within their calibration ranges.

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.

44
00:03:31.727 --> 00:03:34.907
Gemma's smaller error accompanies a narrower supported range.

45
00:03:35.407 --> 00:03:38.187
The error alone cannot tell you how many requests are available.

46
00:03:39.027 --> 00:03:44.547
The requests also meeting the held-out coherence floor number
nineteen, twenty, and thirteen, respectively.

47
00:03:45.307 --> 00:03:50.947
Useful dosing needs these three checks together: reachability,
targeting error, and coherent expression.

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.

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.

50
00:04:05.100 --> 00:04:09.600
For sycophancy at sixty, the initial map selects one point four
eight six four;

51
00:04:10.120 --> 00:04:13.780
the continued controller's map selects zero point eight seven
four five.

52
00:04:14.480 --> 00:04:19.020
The same behavioral request therefore uses a different internal
strength after the update.

53
00:04:19.860 --> 00:04:23.480
The authors remeasure the curve before serving requests with the
updated controller.

54
00:04:24.300 --> 00:04:29.220
Calibration makes the interface interpretable, but it must
follow the system whose behavior it measures.

55
00:04:29.800 --> 00:04:34.360
PersonaDose separates learning a trait from specifying how much
of it to express.

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.

57
00:04:41.940 --> 00:04:46.540
The Qwen example makes the contract concrete: request sixty,
choose the calibrated setting,

58
00:04:46.920 --> 00:04:49.820
then measure fifty-six point two five on held-out responses.

59
00:04:50.520 --> 00:04:55.300
That is control you can inspect through its range, its targeting
error, and its coherence.

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.

61
00:05:02.860 --> 00:05:07.980
A meaningful steering interface starts when an internal knob
becomes a measurable behavioral request.