40, …\ldots seconds) up to the full video length.
Short-scale windows are non-overlapping, while longer-scale windows overlap to retain events crossing segment boundaries.
For each window, a VLM generates complementary low-, mid-, and high-level descriptions covering visible details, actions and transitions, and the overall content.
For windows up to 40 seconds, overlapping information across the three levels is fused into one coherent, redundancy-reduced description; for windows of 80 seconds or longer, the three descriptions are retained separately.
Further details on overlap and related configurations are provided in Appendix A.
Using the generated descriptions, we construct local and global QA supervision without another video pass.
Local QA pairs are generated from individual window descriptions and cover entities, actions, attributes, temporal order, and clip-level reasoning.
Each local question includes the timestamp or temporal range of its source window, allowing the adapter to encode temporal position and order while identifying the supporting video segment for visual memory distillation in Section 3.4.
At inference, questions retain the native format of each evaluation benchmark.
To capture video-wide knowledge, we merge recurring entities and selected long-window descriptions into a global memory that summarizes the event timeline, state changes, and major stages of the video.
This global memory conditions global-QA generation and is also included as one next-token prediction example.
Each generation round selects a different group of focus clips and produces NN timestamp-free global QA pairs, cycling through the groups to distribute supervision across the video.
These global questions target cross-clip retrieval and reasoning.
The resulting descriptions and QA pairs form the video-specific training set:
𝒟V=𝒟desc∪𝒟localQA∪𝒟globalQA.\mathcal{D}_{V}=\mathcal{D}_{\mathrm{desc}}\cup\mathcal{D}_{\mathrm{localQA}}\cup\mathcal{D}_{\mathrm{globalQA}}.
(3)
During adapter optimization, 𝒟desc\mathcal{D}_{\mathrm{desc}} provides next-token prediction supervision, while 𝒟localQA\mathcal{D}_{\mathrm{localQA}} and 𝒟globalQA\mathcal{D}_{\mathrm{globalQA}} provide fine-tuning examples.
Temporally grounded local QA pairs associate each example with its supporting video segment for visual memory distillation, while description and global-QA examples remain text-only.
Our description and entity construction follow the multi-level generation design of NKR (Guan et al., 2026), while local-QA generation adapts its evidence format to open-ended answering and the global-memory and full-video QA components provide video-wide supervision introduced by MIT3.
Appendix B provides the corresponding construction details.
3.3 Causal Localization of Video Memory
The second stage localizes where to write knowledge from VV in the frozen backbone.
Adapting every model layer distributes memory parameters throughout the backbone, while a fixed subset assumes shared storage locations across videos.
MIT3 instead uses delta-injection causal tracing to measure each model layer’s video-specific causal memory utility.
For a candidate layer ll, we ask: if the hidden-state change induced by the video is injected during question-only inference, how strongly can the remaining computation convert it into a higher likelihood for the paired answer?
The resulting answer gain ranks model layers by utility as video-specific memory-writing locations.
Video-induced representations.
For each calibration triple (Vi,Qi,Ai)(V_{i},Q_{i},A_{i}), we run the frozen model with the relevant video segment and question, then with the question alone.
Let qiq_{i} denote the aligned positions of identical question tokens in both forward passes.
Let Hqi,lV​QH^{VQ}_{q_{i},l} and Hqi,lQH^{Q}_{q_{i},l} denote their hidden states at model layer ll under video-conditioned and question-only inference, respectively.
We define the video-induced representation change on these aligned question to