t 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 common memory-writing core and content-dependent refinements observed with Qwen3-VL-8B.
Figure 11: Layer-wise causal utility and top-32 selection frequency for Qwen3-VL-32B on LVBench.
Appendix F Deployment and Resource Analysis
MIT3 offers practical advantages in settings where long videos serve as persistent information sources and are revisited repeatedly. We compare its serving footprint with persistent KV-cache reuse using Qwen3-VL-32B under the same one-hour, 768-frame setting as Table 1. KV reuse avoids repeated video prefill by retaining the video-conditioned KV cache, whereas MIT3 represents each video with a compact LoRA adapter and performs subsequent inference from text alone.
Table 7: Persistent-state comparison for Qwen3-VL-32B.
Setting
KV Reuse
MIT3
Per-video persistent state
26.9 GB
0.59 GB
Backbone + 1 video
89.0 GB
62.7 GB
Backbone + 10 videos
331.1 GB
68.0 GB
Reduction vs. KV reuse
–
45.6×45.6\times smaller
The resource advantage becomes increasingly pronounced as more long videos are retained for subsequent access. Under persistent KV reuse, each additional video introduces a 26.9-GB cache, whereas MIT3 adds only a 0.59-GB adapter. As a result, maintaining ten one-hour videos requires approximately 331 GB including the backbone under KV reuse, compared with only 68 GB for MIT3. This compact per-video state substantially reduces the memory required to maintain and switch among multiple persistent video memories.
Such a deployment pattern is particularly relevant to repeated analysis of long-video collections, including retrospective surveillance review, recorded meetings, lecture archives, and instructional or maintenance videos. In these settings, multiple video memories can share the same backbone and be stored or switched through lightweight adapters, without keeping the original frames or a large video-dependent KV cache resident for every video.
Appendix G Additional Benchmark Results
G.1 Video-MME
Video-MME evaluates general video understanding across short-, medium-, and long-duration videos. Table 8 reports overall accuracy with the Qwen3-VL-8B backbone, providing an additional evaluation beyond the dedicated long-video benchmarks. MIT3 consistently improves over the matched NKR-style baseline, while remaining competitive with direct video-conditioned inference despite operating without video input at query time.
Table 8: Video-MME results. †\dagger denotes methods trained with the same data as MIT3.
Method
Input
Accuracy (%)
Qwen3-VL-8B (Bai et al., 2025)
<768<768 fr.
69.1
Qwen3-VL-8B + NKR-style†
N/A
62.9
Qwen3-VL-8B + MIT3 (Ours)
N/A
66.7
Appendix H Limitations
Our study focuses on test-time training, where each video is internalized through lightweight, video-specific adaptation of a pretrained model. This setting allows us to isolate and evaluate the proposed memory-writing mechanisms, but computational constraints limit our exploration of how these mechanisms scale to large-scale pretraining. Extending them to the pretraining stage may support more persistent and reusable video knowledge across videos and tasks, which we leave for future work.
Experimental support, please
view the build logs
for errors. Generated by
L
A
T
E
xml
.
Instructions for reporting errors
We are continuing to improve HTML versions of papers, and your feedback helps enhance accessibility and mobile
support. To report errors in the HTML that will help us improve conversion and rendering, choose any of the
methods listed below:
Click the "Report Issue" (
) button, located in the page header.
Tip: You can select the relevant text f