ughly one-eighth of the training samples used by NKR. To prevent evaluation leakage, we compare the generated supervision against the benchmark evaluation sets and verify that no evaluation questions are included in either adapter training or causal-localization calibration.
We additionally report the per-video construction time measured on four NVIDIA A100 GPUs. For Qwen3-VL-8B, multi-scale data generation, causal localization, visual-supervision generation, and adapter optimization take approximately 30, 20, 10, and 10 minutes, respectively, for approximately 70 minutes in total. For Qwen3-VL-32B, the corresponding stages take approximately 60, 30, 15, and 40 minutes, for approximately 145 minutes in total. Visual-supervision generation includes computing and caching the video-induced teacher deltas used during optimization.
Appendix D Hyperparameter Analysis
We analyze the sensitivity of MIT3 to two key hyperparameters: the number of selected layers KK and the sampling ratio used for visual memory distillation. All experiments are evaluated on LVBench, varying one factor at a time while keeping the remaining configuration fixed.
Number of Selected Layers.
Tables 4 and 5 report the effect of the layer budget KK for Qwen3-VL-8B and Qwen3-VL-32B, respectively. For both backbones, performance generally improves as more causally relevant layers are included, but begins to decline once the selected set becomes overly broad. Qwen3-VL-8B achieves its best performance at K=18K=18, while Qwen3-VL-32B peaks at K=32K=32. This behavior suggests that effective memory internalization benefits from concentrating adaptation on a sufficiently expressive yet selective subset of layers: too few layers provide insufficient capacity for storing video-specific knowledge, whereas expanding the adaptation set beyond the causally salient region may dilute this specialization. We therefore use K=18K=18 and K=32K=32 for the 8B and 32B models, respectively, throughout the main experiments.
Table 4: Effect of the selected-layer budget KK for Qwen3-VL-8B on LVBench.
KK
1
3
6
12
18
24
Accuracy (%)
26.1
24.9
35.6
39.6
43.5
41.7
Table 5: Effect of the selected-layer budget KK for Qwen3-VL-32B on LVBench.
KK
12
16
24
32
48
Accuracy (%)
37.9
42.1
46.7
50.6
48.1
Visual-Supervision Sampling Ratio.
Table 6 studies the fraction of temporally grounded QA pairs used for visual memory distillation with Qwen3-VL-8B. Removing visual supervision reduces accuracy from 43.5 to 38.2, confirming its importance for effective memory internalization. Increasing the sampling ratio from 50% to 100%, however, yields only a marginal improvement of 0.2 points. This saturation indicates that visual memory distillation does not require dense supervision over all local QA pairs; a moderate subset already provides sufficient visual guidance. We therefore use a 50% sampling ratio in the main experiments to retain nearly the full benefit while reducing the cost of teacher-delta generation.
Table 6: Effect of the visual-distillation sampling ratio on LVBench with Qwen3-VL-8B.
Visual QA Sampling Ratio
Accuracy (%)
0%
38.2
50%
43.5
100%
43.7
Appendix E Layer-Wise Distribution with Qwen3-VL-32B
We repeat the layer-distribution analysis with Qwen3-VL-32B, selecting the top-32 adaptation layers for each LVBench video. As in the 8B setting, Figure 11 visualizes the normalized per-video causal utilities together with the cross-video selection frequency. A similar structure emerges at the larger model scale: high-utility layers are concentrated within a shared region across videos, while a smaller subset of selected layers varies with individual video content. The frequently selected layers account for approximately 72% of all top-32 selections, indicating substantial cross-video consistency in the preferred adaptation locations. At the same time, the remaining selections preserve video-specific variation around this shared structure. Overall, the 32B analysis exhibits the same combination of a 