 multi-scale supervision introduced in Section 3.2, together with representative training examples. To balance temporal coverage and redundancy, the 5-, 10-, and 20-second windows are non-overlapping, while windows of 40 seconds or longer use 50% overlap with a stride equal to half the window length. For windows up to 40 seconds, overlapping content from the low-, mid-, and high-level descriptions is fused into a single coherent, redundancy-reduced training example; for windows of 80 seconds or longer, the three descriptions are retained as separate examples. Table 3 summarizes the resulting configuration. Figure 5 illustrates the corresponding training formats using a 5-second fused-description example, an 80-second separate-description example, and representative local and global QA pairs. Ellipses indicate omitted content for compact presentation.
Table 3: Per-video multi-scale training-data construction. “Merged” indicates that overlapping information from the three description levels is fused into one coherent, redundancy-reduced training example.
Window
Stride
Overlap
Descriptions
Local QA
5 s
5 s
0%
Merged
1
10 s
10 s
0%
Merged
1
20 s
20 s
0%
Merged
3
40 s
20 s
50%
Merged
5
80 s
40 s
50%
Separate
5
≥\geq160 s
Half-window
50%
Separate
10
Short window (5 seconds): fused description example
Time range: 00:00:00–00:00:05
Fused example:
White serif text on a dark background explains that unique wooden masks are carved to honor the dead and placed in sacred shrines, where they gradually return to dust and symbolize the spirit’s peaceful journey into the next life. Subtle blue light orbs drift in the background …
Long window (80 seconds): separate description examples
Time range: 00:00:00–00:01:20
[High] A young girl and two small green-haired children interact in a mystical forest …
[Mid] The video opens with text about wooden masks and spirit guides; the girl then observes the children nearby …
[Low] White text appears on screen; the girl has dark hair and wears a blue top, while the children have green hair with colorful ornaments …
Local QA
Q: From 00:00:00 to 00:02:40, what does the girl do while interacting with the creatures?
A: She moves through the forest and interacts closely with the two creatures …
Global QA
Q: What does the text in the video say about the purpose of the wooden masks?
A: They are carved to honor the dead and symbolize the spirit’s peaceful journey …
Figure 5:
Representative training examples for the multi-scale supervision pipeline.
For short windows (≤40\leq 40 s), overlapping information from the high-, mid-, and low-level descriptions is fused into one coherent, redundancy-reduced training example, whereas for long windows (≥80\geq 80 s), the descriptions are retained as three separate examples.
Local QA pairs include explicit temporal grounding, while global QA pairs are timestamp-free.
Ellipses denote omitted content.
Appendix B Data Generation Prompts
This appendix documents the prompts used for data generation in MIT3. We provide the newly introduced prompts in full, while components inherited from or adapted from NKR (Guan et al., 2026) are summarized at the level of their structure and modifications.
B.1 NKR-Based Components
Our data-generation pipeline retains several components from NKR (Guan et al., 2026). Specifically, we preserve its three-level captioning format within a single model call and its two-stage entity-processing procedure. For local clip-level QA generation, we follow the same caption-and-entity evidence organization, but adapt the task from multiple-choice to open-ended QA and require each question to be temporally grounded.
Multi-level clip-description prompt
Instructions. You are a professional video annotator. Based on the input video clip, generate three levels of natural, fluent, and descriptive captions in English.
Output format. Return a JSON object with the fields
caption_high_level, caption_mid_level, and
caption_low_level.
Guidelines.
•
Capture temporal dynamic