, demonstrating the effectiveness of video-specific layer selection and visual memory distillation for parameterized video memory.
Despite performing video-free, text-only inference after one-time internalization, MIT3 remains close to direct video-conditioned inference on the full benchmarks, while requiring no video frames, visual tokens, retrieval, or video-dependent KV cache at query time.
More importantly, with the 32B backbone, MIT3 reaches 50.1 on LVBench >1>1h and 55.9 on LongVideoBench-Long, surpassing direct video-conditioned inference on both long-video subsets.
Together with the small variation between the full LVBench and its >1>1h subset, these results highlight the effectiveness and stability of MIT3 for long-video understanding.
This makes MIT3 particularly suitable for scenarios that repeatedly revisit long videos, such as archived surveillance review, recorded meetings, and lecture or instructional video analysis.
We further analyze the resource advantages of MIT3 over alternative inference paradigms in Appendix F.
4.3 Layer-Wise Distribution
We analyze the layer-wise causal utilities of Qwen3-VL-8B on LVBench to examine whether video-specific localization selects arbitrary layers or reveals a consistent organization of video memory.
For each video, we normalize Rl​(V)R_{l}(V) across layers and identify the top-18 layers selected for adaptation. Figure 3(a) shows the resulting video-layer heatmap, while Figure 3(b) reports the selection frequency of each layer across videos.
Figure 3: Layer-wise causal utility and top-18 selection frequency for Qwen3-VL-8B on LVBench.
The selected layers show substantial cross-video agreement.
A compact group of layers is repeatedly assigned high causal memory utility, forming a shared memory-writing core that accounts for 65% of all top-18 selections.
At the same time, a smaller set of selected layers varies across videos, introducing content-dependent refinements around this common structure.
These results suggest that causal localization captures both a consistent cross-video organization of memory and video-specific adaptation needs, rather than reducing to either a fixed layer subset or arbitrary per-video selection.
The corresponding analysis for Qwen3-VL-32B is provided in Appendix E.
4.4 Ablation Study
Component ablations.
Table 2 isolates the contributions of layer selection and visual memory distillation on LVBench using Qwen3-VL-8B.
All component ablations apply LoRA to the attention projections and compare all-layer, random-KK, and causal top-KK placement with or without the visual loss.
We further examine the choice of LoRA target by applying the complete MIT3 configuration to either the attention projections or the FFN while keeping all other settings fixed.
The two choices yield comparable performance, we therefore follow the NKR configuration and use attention-targeted LoRA in all main experiments.
Table 2: Component and LoRA-target ablations on LVBench with Qwen3-VL-8B.
Configuration
Layer Selection
LoRA Target
Visual Loss
Accuracy (%)
Baseline
All
Attention
–
36.7
Random-KK
Random KK
Attention
–
34.1
All + Visual
All
Attention
✓
41.2
Causal Top-KK
Causal Top-KK
Attention
–
39.2
MIT3
Causal Top-KK
Attention
✓
43.5
MIT3 (FFN)
Causal Top-KK
FFN
✓
43.9
Data efficiency.
We further compare NKR-style and MIT3 under matched supervision budgets using Qwen3-VL-8B and Qwen3-VL-32B.
The budgets in Fig. 4, 0.5k, 2k, 5.5k, 10k, 15k, and 18k examples per video, are estimated from a one-hour video, with 5.5k corresponding to our standard configuration.
MIT3 consistently outperforms the matched NKR-style baseline across all supervision budgets and both backbone sizes, indicating that the proposed memory-writing strategy improves the utilization of synthesized supervision rather than relying on a larger training set.
Notably, the performance of MIT3 approaches saturation once the supervision reaches a moderate scale, and further increasing the number of generated examples yie