.
3.
Questions must not contain timestamps or time references, although answers may mention times.
4.
Answers must be definitive rather than speculative.
5.
Answers must contain concise factual content (30–100 tokens).
6.
Do not generate multiple-choice questions, refusals, or answers requiring knowledge beyond the provided memory.
Figure 10: MIT3 prompt for open-ended full-video QA.
Appendix C Additional Implementation Details
Datasets and Baselines
LVBench (Wang et al., 2025a) contains 103 long videos and 1,549 question–answer pairs covering entity recognition, event understanding, key-information retrieval, temporal grounding, reasoning, and summarization. We report results on both the full benchmark and the subset containing videos longer than one hour. For LongVideoBench (Wu et al., 2024), we evaluate on the full benchmark and the LongVideoBench-Long subset adopted by NKR (Guan et al., 2026), which contains 564 questions over 188 videos of 900–3,600 seconds. Subtitle and audio inputs are excluded from both benchmarks. We additionally evaluate on Video-MME (Fu et al., 2025), which covers short-, medium-, and long-duration videos across diverse perception and reasoning tasks.
Published baselines include VideoLLaMA3 (Zhang et al., 2025), AdaReTaKe (Wang et al., 2025b), VideoRAG (Ren et al., 2026), VCA (Yang et al., 2025), Deep Video Discovery (DVD-o3) (Zhang et al., 2026), and NKR (Guan et al., 2026). For direct Qwen3-VL inference, frames are sampled at 2 fps and capped at 768 frames. Since the per-video training data of NKR are unavailable, we follow its data-generation procedure and train the controlled NKR-style baseline and MIT3 using exactly the same generated supervision. The NKR-style baseline optimizes LoRA parameters with only ℒNTP+ℒSFT\mathcal{L}_{\mathrm{NTP}}+\mathcal{L}_{\mathrm{SFT}}, allowing the effects of causal layer selection and visual memory distillation to be evaluated independently of differences in synthetic supervision.
Optimization Details
MIT3 is implemented in PyTorch (Paszke et al., 2019). Following the video-NKR configuration, we use LoRA with rank 256 and scaling factor 512 and optimize the video-specific parameters with AdamW. Unless otherwise specified, each adapter is trained for 5 epochs with a per-device batch size of 4, an initial learning rate of 2×10−42\times 10^{-4} with cosine annealing, and visual-loss weight α=1\alpha=1.
For causal localization, we sample at most 20 calibration QA pairs from each temporal scale. For each candidate layer, the intervention gain Gi,l​(λ)G_{i,l}(\lambda) is evaluated on
Λ={0.25,0.5,0.75,1.0},\Lambda=\{0.25,0.5,0.75,1.0\},
(16)
and approximate the equivalent weighted integral in Equation 7 by averaging over these four sampled intervention strengths:
Ri,l≈1|Λ|​∑λ∈Λ(1−λ)​Gi,l​(λ).R_{i,l}\approx\frac{1}{|\Lambda|}\sum_{\lambda\in\Lambda}(1-\lambda)G_{i,l}(\lambda).
(17)
Thus, bb serves only as an auxiliary variable in the continuous formulation and is not separately discretized during implementation. For visual memory distillation, we randomly sample 50% of the temporally grounded QA pairs at each temporal scale and precompute their teacher deltas before adapter optimization. Sensitivity analyses of the top-KK layer budget and visual-supervision sampling ratio are provided in Appendix D.
Training Data
For a one-hour video, the data-generation pipeline produces approximately 2k multi-scale description examples, 3k temporally grounded local QA pairs, and 500 timestamp-free global QA pairs, yielding approximately 5.5k training examples in total. This corresponds to roughly 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 