WEBVTT

1
00:00:00.220 --> 00:00:04.760
Consider a neutral timeline with Phase A, Event B, and Phase C.

2
00:00:05.440 --> 00:00:11.440
The question asks what happened immediately before Event B, so
finding B is only the first step.

3
00:00:12.240 --> 00:00:17.540
Inspecting only the frames inside B may miss the earlier
interval that actually contains the answer.

4
00:00:18.200 --> 00:00:24.080
The system needs to localize the event, inspect the preceding
window, and connect the evidence across time.

5
00:00:24.940 --> 00:00:30.320
This original diagram contains no patient details, operative
images, or recommended medical actions.

6
00:00:30.880 --> 00:00:35.380
It explains the retrieval problem MedClaw studies in long
surgical-video question answering.

7
00:00:36.027 --> 00:00:39.527
MedClaw separates retrieval planning from visual perception,

8
00:00:40.067 --> 00:00:44.227
using a text-only orchestrator and frozen vision-language
sub-agents.

9
00:00:45.087 --> 00:00:50.227
The orchestrator never sees the video directly; it decides which
evidence to request through tools.

10
00:00:51.047 --> 00:00:57.067
Visual sub-agents inspect the specified pixels and return
textual observations that the orchestrator can reason about.

11
00:00:57.927 --> 00:01:01.807
No model weights are updated during this harness's inference or
heuristic skill-learning loop.

12
00:01:02.527 --> 00:01:08.467
Its adjustable part is the external library of retrieval
strategies, not a newly fine-tuned visual backbone.

13
00:01:09.267 --> 00:01:14.467
Keeping those roles separate turns an implicit attention
decision into a visible request for evidence.

14
00:01:14.873 --> 00:01:22.013
The tools expose progressively finer evidence: a full recording,
a chosen temporal segment, and individual frames.

15
00:01:22.873 --> 00:01:28.473
A full-video pass produces a rough timeline and then revisits a
candidate window at higher resolution.

16
00:01:29.273 --> 00:01:34.613
A segment tool reads the localized interval or an adjacent
interval needed for a temporal question.

17
00:01:35.593 --> 00:01:41.753
Independent frame inspection can check a specific visible
detail, including a smaller spatial crop when needed.

18
00:01:42.613 --> 00:01:47.292
A knowledge-retrieval tool supplies general background terms,
rather than answers from the benchmark.

19
00:01:48.053 --> 00:01:53.453
Background knowledge and the evidence from this particular
recording remain different inputs to the reasoning process.

20
00:01:54.053 --> 00:02:00.073
Each tool call names its arguments, such as a time range, and
returns an observation into the running context.

21
00:02:00.893 --> 00:02:05.513
The orchestrator then decides whether the evidence is sufficient
or another look is needed.

22
00:02:06.413 --> 00:02:11.493
The action-observation sequence is recorded until the model
stops and produces its final answer.

23
00:02:12.353 --> 00:02:18.133
In the original before-B example, an event-window observation
should trigger inspection of the preceding interval.

24
00:02:18.673 --> 00:02:24.293
That makes the requested re-look visible, instead of hiding the
retrieval decision inside one compressed pass.

25
00:02:25.393 --> 00:02:31.573
A recorded trace supports inspection, but recording a decision
alone does not guarantee its interpretation is correct.

26
00:02:32.180 --> 00:02:37.500
Heuristic Skill Distillation starts from one hundred labeled
questions in the separate SVU training split.

27
00:02:38.140 --> 00:02:44.140
Sixty questions support candidate mining, and forty form the
validation set used to gate those candidates.

28
00:02:44.880 --> 00:02:50.700
The agent first answers blind, then compares its answer and
retrieval trace against the reference during mining.

29
00:02:51.540 --> 00:02:58.640
An operational miss can suggest a tighter crop or adjacent
window; an unreachable answer offers no retrieval lesson.

30
00:02:59.320 --> 00:03:05.600
A skill must beat the empty-library mean by over zero point zero
five, with no answer losing two or more points.

31
00:03:06.340 --> 00:03:11.340
The retained lesson changes the external skill library, while
the model parameters remain frozen.

32
00:03:11.927 --> 00:03:15.147
The learned library prominently includes directed re-look,

33
00:03:15.667 --> 00:03:19.647
a strategy triggered by questions about what precedes or follows
an event.

34
00:03:20.327 --> 00:03:25.207
After localizing the event, the agent inspects the following
window for an after question.

35
00:03:25.667 --> 00:03:30.447
For a before or preparation question, it returns to the
preceding window instead.

36
00:03:31.327 --> 00:03:33.967
The strategy addresses a specific retrieval failure:

37
00:03:34.507 --> 00:03:38.227
reading the named event without reading the context the question
asks about.

38
00:03:38.967 --> 00:03:44.167
Our neutral timeline illustrates the idea; its phases are not
observations from a real surgical recording.

39
00:03:44.907 --> 00:03:50.887
The mechanism is evidence seeking across time, rather than
assuming the customary next step from background knowledge.

40
00:03:51.553 --> 00:03:56.133
MedClawBench combines two hundred fifty-nine long neurosurgery
questions

41
00:03:56.613 --> 00:03:59.953
with eight hundred sixty-four held-out lecture-video questions.

42
00:04:00.733 --> 00:04:03.693
The total is one thousand one hundred twenty-three questions,

43
00:04:04.133 --> 00:04:07.233
with filtering designed to remove specified answer shortcuts.

44
00:04:08.193 --> 00:04:13.113
Blind and local evidence tests check whether a question requires
watching its targeted interval.

45
00:04:14.133 --> 00:04:20.053
Lecture questions also face an audio-only filter, reducing
answers obtainable by listening to narration alone.

46
00:04:21.253 --> 00:04:23.973
Answers receive four one-to-five rubric scores:

47
00:04:24.553 --> 00:04:29.353
correctness, detail, contextual understanding, and temporal
understanding.

48
00:04:30.373 --> 00:04:32.953
These benchmark answer scores are research measurements,

49
00:04:33.293 --> 00:04:36.433
not patient outcomes or a test of clinical deployment readiness.

50
00:04:36.940 --> 00:04:43.060
A controlled ablation wraps the same Qwen two point five V L
seven-billion-parameter visual backbone in the harness.

51
00:04:43.900 --> 00:04:48.860
On the neurosurgery split, its one-shot four-dimensional average
is two point zero four.

52
00:04:50.160 --> 00:04:53.860
With the harness, the same visual-backbone row reports two point
five eight.

53
00:04:55.000 --> 00:04:59.900
The added orchestrator and tool loop are part of the comparison;
the visual model itself is not retrained.

54
00:05:01.120 --> 00:05:06.580
A separate component ablation reports a zero point three eight
score reduction when directed re-look is removed.

55
00:05:07.600 --> 00:05:13.340
Read these as scoped rubric-score improvements, not
probabilities that an agent can safely interpret any clinical
video.

56
00:05:13.960 --> 00:05:19.760
Return to Event B: locating the event is useful, but the
question may point to another nearby interval.

57
00:05:20.340 --> 00:05:25.580
A text planner can choose that interval, while a visual tool
inspects what the recording actually contains.

58
00:05:26.340 --> 00:05:30.420
The observations and decisions remain in a trace that can be
inspected and revisited.

59
00:05:31.320 --> 00:05:36.320
A validated retrieval lesson can later guide the next question
without changing the underlying model weights.

60
00:05:37.300 --> 00:05:40.920
MedClaw studies that harness design under a defined long-video
benchmark,

61
00:05:41.300 --> 00:05:43.400
with evidence gathering as the central mechanism.

62
00:05:44.260 --> 00:05:49.040
The idea is to improve how a system looks for support before
trusting the answer it produces.