# TODO List for Paper Refinement

## 2. Methodology & Theoretical Rigor
- [ ] **Theoretic Proof Augmentation (Appendix)**:
    - [ ] Add a formal assumption regarding the **Similarity Kernel** $w_{i,j}$ (e.g., Gaussian/Exponential decay with distance).
    - [ ] Explicitly prove the link between the SMD window $k$ and the support of the kernel.
    - [ ] Add a discussion on how SMD serves as a **first-order greedy proxy** by performing local repulsion without the need for sequential iterations.
- [ ] **Parameter Justification (Sec 3.3)**:
    - [ ] Explicitly explain the discrepancy between the statistical correlation length ($\sim 85$ tokens) and the decision radius ($k=7$).
    - [ ] Define $\gamma=0.1$ as a "nuance preservation" factor, distinguishing SMD from hard NMS.

## 3. Implementation & Complexity
- [ ] **Hardware Overhead Analysis**: Document the GPU latency and memory I/O of the 1D max-pooling operation compared to standard Top-K selection.
- [ ] **Sensitivity Analysis**: Run an ablation study on $k \in \{3, 5, 7, 9, 15, 31\}$ and $\gamma \in \{0, 0.1, 0.2, 0.5\}$ to demonstrate the robustness of the choice.

## 4. Writing & Narrative Flow
- [ ] **Bridge Sentence**: Add a transition in Sec 3.1 or 3.3 to explicitly connect the "Empirical NMS" from observations to the "Submodular SMD" in methodology.
- [ ] **Figure Captions**: Ensure all captions reflect the consolidated terminology (SMD).