WEBVTT

1
00:00:00.280 --> 00:00:07.820
Which moved first, the blue circle or the yellow square? The
question alone does not specify either event order.

2
00:00:08.760 --> 00:00:14.820
With both frames hidden, an answer could be a guess; a correct
guess would not establish that anything was observed.

3
00:00:15.980 --> 00:00:22.400
Now reveal two original event views: the blue circle moves in
the first, before the yellow square changes position.

4
00:00:23.420 --> 00:00:29.460
The visible sequence supports blue-circle first, linking the
answer to evidence rather than object expectations.

5
00:00:30.680 --> 00:00:36.480
These are neutral geometric events, not footage of real people
or a recreated question from a research benchmark.

6
00:00:37.500 --> 00:00:42.480
When a model answers correctly, did the video supply information
necessary to reach that answer?

7
00:00:43.100 --> 00:00:49.320
Compare that event-order question with asking what color a shape
becomes after it is painted with red paint.

8
00:00:50.200 --> 00:00:56.340
The wording suggests red without video, so success may reflect
text rather than visual understanding.

9
00:00:57.400 --> 00:01:03.140
A video may show the same answer, but its presence does not make
the question depend on visual evidence.

10
00:01:04.260 --> 00:01:09.800
Our first question leaves the order open, because either object
could move first in an otherwise plausible scene.

11
00:01:10.760 --> 00:01:16.760
That separates evidence required for an answer from evidence
merely present while the answer is already inferable.

12
00:01:17.820 --> 00:01:24.160
The paper calls text-based success linguistic shortcutting,
motivating a video-removal test of answerability.

13
00:01:24.860 --> 00:01:30.500
The model receives only the question and options, with video
content and subtitles entirely withheld.

14
00:01:31.480 --> 00:01:37.780
Correct answers define text-only answerable items for that
model; other questions are treated as visually grounded.

15
00:01:38.840 --> 00:01:45.040
These labels describe an operational test, so the distinction
depends on which model performs the text-only assessment.

16
00:01:46.180 --> 00:01:53.620
Shortcuts include wording, world knowledge, implausible-option
elimination, and imagined content that happens to match.

17
00:01:54.660 --> 00:02:00.000
A higher score can include stronger language-only reasoning,
without equal gains in interpreting video.

18
00:02:00.840 --> 00:02:06.620
Testing without visuals reveals this contribution, rather than
crediting every correct answer to visual comprehension.

19
00:02:07.260 --> 00:02:13.200
The paper summarizes roughly forty to sixty percent text-only
answerability on the examined benchmarks.

20
00:02:13.800 --> 00:02:19.460
The exact proportion depends on the benchmark and assessing
model, not a universal set of video questions.

21
00:02:20.440 --> 00:02:28.520
Table one reports GPT-five text-only accuracy of forty-eight
point two on VideoMME, with video input withheld.

22
00:02:29.380 --> 00:02:36.340
For multiple-choice MMVU, the score is fifty-seven point one;
chance accuracy differs across benchmark designs.

23
00:02:37.460 --> 00:02:43.620
These correct-answer rates without video show why accuracy alone
cannot isolate the added value of watching.

24
00:02:44.420 --> 00:02:50.160
Post-training data has similar shortcut questions, which can
reward language priors instead of visual reasoning.

25
00:02:50.800 --> 00:02:57.700
VidGround curates Video-R1 post-training data, containing both
video and image question-answer examples.

26
00:02:58.439 --> 00:03:05.100
GPT-five mini sees questions and options; selection keeps those
it cannot correctly answer without visuals.

27
00:03:05.860 --> 00:03:12.100
The retained pool has one hundred eighty-one thousand seven
hundred ten examples, sixty-nine point one percent.

28
00:03:12.840 --> 00:03:18.680
The full pool has two hundred sixty-three thousand seventy-one
examples, mixing video and image questions.

29
00:03:19.600 --> 00:03:25.500
This training-data ratio differs from the benchmark diagnosis,
which measures answerability on separate datasets.

30
00:03:26.480 --> 00:03:33.120
Text-only failure defines selection; it does not guarantee every
retained question has no possible linguistic shortcut.

31
00:03:33.653 --> 00:03:40.353
The approach combines data selection with RL post-training,
targeting examples that require visual evidence.

32
00:03:41.253 --> 00:03:47.833
Qwen two point five VL seven B Instruct is trained with sixteen
uniformly sampled frames per video.

33
00:03:48.493 --> 00:03:54.893
Training uses seven hundred steps with the GRPO-based
optimization and temporal rewards stated in the source.

34
00:03:55.993 --> 00:04:05.093
VideoMME, VideoMMMU, and multiple-choice MMVU are evaluated with
sixteen, thirty-two, or sixty-four video frames.

35
00:04:06.092 --> 00:04:11.693
Full benchmark and visually grounded subset averages use
separate populations, which must not be mixed.

36
00:04:12.533 --> 00:04:18.173
The ablation compares full-pool and selected-pool training,
isolating the effect of data composition.

37
00:04:18.780 --> 00:04:26.040
At sixty-four frames, unfiltered-pool GRPO reports a fifty-three
point three Full average in Table three.

38
00:04:27.160 --> 00:04:33.320
Training on the selected visually grounded pool reaches
fifty-nine point five on that same full-benchmark average.

39
00:04:34.340 --> 00:04:41.160
The six-point-two-percentage-point difference belongs to this
model, data comparison, frame count, and population.

40
00:04:42.160 --> 00:04:48.380
The visually grounded subset instead compares forty-two point
nine with forty-seven point nine: five points.

41
00:04:49.260 --> 00:04:55.860
Thus the reported up-to-six-point-two improvement is not the
gain for every frame setting, task, or subset.

42
00:04:56.760 --> 00:05:02.860
The result supports checking which examples require visual
evidence, rather than assuming more data must be better.

43
00:05:03.467 --> 00:05:09.807
Event order needs the displayed sequence, while a leading color
question can already suggest its answer.

44
00:05:10.887 --> 00:05:16.187
VidGround screens text-only answerability before selecting
examples to guide post-training.

45
00:05:17.007 --> 00:05:23.107
Selected examples provide visual-dependent supervision under the
filter, alongside an existing RL approach.

46
00:05:24.047 --> 00:05:30.867
Measured gains are configuration-specific; the deeper lesson
concerns the evidence that training examples reward.

47
00:05:31.907 --> 00:05:38.227
Evaluation should separate language-only success from the extra
contribution of watching the actual video.

48
00:05:39.087 --> 00:05:44.647
For a model meant to watch before answering, the training
question should make the visual evidence matter to the answer.