WEBVTT

1
00:00:00.320 --> 00:00:06.040
Imagine a profile form with a text field, a Save button, and a
working stored-value update.

2
00:00:06.780 --> 00:00:12.060
The confirmation message never appears, so the user cannot tell
whether the requested action finished.

3
00:00:12.700 --> 00:00:17.880
A repair could add that message, yet accidentally remove the
already working Save behavior.

4
00:00:18.560 --> 00:00:22.780
A prettier screenshot would not reveal that regression without
trying the control again.

5
00:00:23.640 --> 00:00:28.600
This is an original schematic example, with no real user data or
measured repair outcome.

6
00:00:29.240 --> 00:00:35.940
It introduces WebWorld's central question: what evidence should
allow a proposed web-code repair to survive?

7
00:00:36.560 --> 00:00:43.620
WebWorld treats the browser as an executable simulator of how an
HTML artifact behaves under user actions.

8
00:00:44.500 --> 00:00:50.180
The runtime renders the current page, applies interaction
probes, and records evidence of what happened.

9
00:00:51.000 --> 00:00:58.000
That evidence includes screenshots, document state, console
information, probe outcomes, and an action trace.

10
00:00:59.020 --> 00:01:05.640
A vision-language model reads those observations and proposes a
critique, rather than certifying its own repair.

11
00:01:06.540 --> 00:01:11.340
A planner converts the critique into something the browser can
check after the candidate edit.

12
00:01:12.220 --> 00:01:18.800
Here, world model means the web execution platform; it does not
mean a physical-world prediction model.

13
00:01:19.380 --> 00:01:25.780
An interaction contract states what must improve, how to test
it, and what existing behavior must survive.

14
00:01:26.580 --> 00:01:31.860
In our form example, the target is a visible confirmation after
pressing the Save button.

15
00:01:32.520 --> 00:01:37.920
The replay specifies a stable control, an input action, and the
state to inspect afterward.

16
00:01:38.740 --> 00:01:45.680
The preserve set carries previously verified predicates, so
fixing one target cannot silently erase prior progress.

17
00:01:46.560 --> 00:01:51.180
For a local issue, the contract also limits the edit's impact to
the relevant region.

18
00:01:51.800 --> 00:01:58.000
These are bounded claims about specified interactions, rather
than a promise that every possible action will work.

19
00:01:58.827 --> 00:02:03.707
A typed repair skill produces a constrained patch inside the
contract's impact scope.

20
00:02:04.627 --> 00:02:10.707
Candidate filters can discard unrelated, duplicate, or
preservation-risk edits before they reach the browser.

21
00:02:11.687 --> 00:02:15.227
That narrows the search, but it does not establish that the page
now works.

22
00:02:16.087 --> 00:02:21.247
The candidate must be executed again, with evidence tied to its
contract and recorded proof level.

23
00:02:22.147 --> 00:02:28.347
For the interactive form, replay checks the new confirmation and
rechecks the previously working Save behavior.

24
00:02:29.107 --> 00:02:34.867
If either fails, our illustrative repair should remain a
rejected candidate rather than become the next baseline.

25
00:02:35.560 --> 00:02:43.060
WebWorld distinguishes safe runtime steps from
certificate-backed quality steps that can be exported for
training.

26
00:02:44.040 --> 00:02:49.360
A probe run, partial fix, or rejected edit can still help decide
what to try next.

27
00:02:50.080 --> 00:02:54.640
But those steps do not enter the supervised training pool simply
because the agent was active.

28
00:02:55.620 --> 00:03:02.480
Only admitted transitions carry the artifact, the failure it
addressed, and the evidence that justified acceptance.

29
00:03:03.700 --> 00:03:10.080
Previously verified behavior becomes part of later preservation
checks, creating the proposed quality ratchet.

30
00:03:11.020 --> 00:03:15.140
The learning example therefore describes verified progress
between states,

31
00:03:15.620 --> 00:03:18.200
rather than only a final preferred screenshot.

32
00:03:18.840 --> 00:03:22.140
The experiments compare raw HTML supervision

33
00:03:22.200 --> 00:03:27.020
with supervision selected from thirty-two thousand eight hundred
accepted transitions.

34
00:03:27.920 --> 00:03:34.600
Within each model size, the backbone family, prompt template,
optimizer, and evaluation runner remain fixed.

35
00:03:35.440 --> 00:03:39.900
This makes the source of supervision the central difference in
the matched training comparison.

36
00:03:41.220 --> 00:03:47.560
HTMLBench four hundred is the primary evaluation, with
MiniAppBench validation used to study transfer.

37
00:03:48.640 --> 00:03:53.760
The construction pipeline reports separation from benchmark
prompts and their deterministic test cases.

38
00:03:54.740 --> 00:04:01.240
The following chart compares the two
twenty-seven-billion-parameter rows, not unrelated
frontier-model configurations.

39
00:04:01.780 --> 00:04:09.040
The raw-supervision twenty-seven-billion-parameter row reports
an HTMLBench composite score of forty-seven point four.

40
00:04:09.680 --> 00:04:14.740
The WebWorld row at the same model size reports fifty-two point
seven under that benchmark.

41
00:04:15.340 --> 00:04:19.520
That is a gain of five point three score points in this matched
training setup.

42
00:04:20.280 --> 00:04:24.880
The composite score is distinct from the separately reported
test-case pass percentage.

43
00:04:25.520 --> 00:04:31.280
The same model-size pair improves the MiniAppBench validation
average by fourteen point nine points.

44
00:04:32.120 --> 00:04:35.680
These results concern the reported interactive HTML benchmarks,

45
00:04:36.060 --> 00:04:38.660
rather than a guarantee of correctness for every website.

46
00:04:39.240 --> 00:04:45.800
A separate nine-billion-parameter ablation keeps critique and
repair routing while loosening the certificate gate.

47
00:04:46.660 --> 00:04:52.400
Without the certificate, its HTMLBench gain over Raw is only
zero point four score points.

48
00:04:53.420 --> 00:04:58.680
The full WebWorld admission rule gives four point eight points
over Raw at that model size.

49
00:04:59.720 --> 00:05:04.480
The experiment therefore supports more than the value of asking
a model to critique its own page.

50
00:05:05.100 --> 00:05:10.800
It tests whether the loop's selected training transitions are
backed by the specified verification rule.

51
00:05:11.760 --> 00:05:17.440
The effect size belongs to this matched ablation; it should not
be generalized to every browser-checking system.

52
00:05:18.087 --> 00:05:24.847
Return to our original form: a confirmation matters, but so does
the Save action that worked before.

53
00:05:25.527 --> 00:05:31.407
A candidate change should be described as a concrete target,
then tested under the relevant interactions.

54
00:05:32.107 --> 00:05:38.907
The working parts become explicit preservation obligations,
rather than assumptions hidden behind a polished screenshot.

55
00:05:39.867 --> 00:05:45.147
WebWorld makes that distinction part of both the repair loop and
the training-data admission rule.

56
00:05:45.867 --> 00:05:52.327
A proposed improvement is a hypothesis; the recorded contract
evidence determines whether the system should keep it.

57
00:05:53.407 --> 00:06:00.147
That is the idea to carry forward: self-improvement needs a
definition of progress that survives execution.