1. Presentation clarity.
I found the paper considerably difficult to follow throughout. Several technical terms, such as "persistent shares," "nested allocation," "reference scale," and "trace-matched scale," are introduced without sufficient context, particularly before readers are familiar with the proposed method. The motivation and significance of the theoretical contributions, including the uniqueness and quantitative bounds of the proposed allocation, are also difficult to understand when first introduced. Even in the method and experimental sections, the relationships between individual components and their roles in the overall framework are not always clearly explained. I believe the presentation needs substantial improvement to make the main ideas and contributions accessible to readers.
2. The performance improvements are relatively modest.
Performance improvements over the baselines are modest in several experiments. For example, Table 1 reports improvements from 4.99 to 4.79 in SfM depth error (×10⁻³), 39.31 to 39.36 dB in low-light PSNR, and 17.22 to 17.51 dB in reflective-pixel PSNR compared with FisherRF. Although larger improvements are reported in the task-specific focus regions, it is still unclear whether these gains are sufficiently meaningful. Furthermore, the qualitative results in Figure 3 do not show clearly distinguishable improvements over the baselines. Additional quantitative evidence supporting the significance of these improvements would strengthen the evaluation.
3. Insufficient comparison and differentiation from related methods.
The main evaluation in Table 1 primarily compares against baselines that do not explicitly consider source-level attribution or task-specific objectives, while Table 5 mainly provides internal allocation ablations. The paper would benefit from comparisons with related methods, particularly those considering source-specific uncertainty (e.g., OUGS) and task-specific information gain (e.g., GO-PRE). In addition, the authors should more clearly distinguish their approach from existing methods that address source-specific uncertainty (e.g., [1]), task-specific information gain (e.g., [2]), or transient/dynamic content during acquisition (e.g., [3]). Including relevant baselines where applicable, along with a more detailed discussion of these differences, would help establish the contributions of the proposed method.
4. Insufficient experimental coverage.
Although the paper evaluates several tasks, the number of test scenes for each task is relatively small (Table 4), particularly for low-light and reflective scenes. The authors should consider expanding the evaluation to more diverse environments or justify whether the current scale is sufficient to demonstrate generalizability across environments. Furthermore, the task-aware experiments exclude nuisance uncertainty, while the nuisance-aware acquisition experiment uses uniform task weights and only one scene. Therefore, quantitative evaluation of the combined effectiveness of task-aware attribution and nuisance handling in closed-loop acquisition remains insufficient.
5. Insufficient detailed analysis.
Section 4.4 appears intended to provide a detailed analysis of the proposed method, but its purpose is not sufficiently clear. The section seems to conflate two different questions: (i) whether the proposed parameter categorization appropriately captures the information sources relevant to the target evaluation scenarios, and (ii) whether the source-level attributions are accurately computed despite the approximations involved. The authors should clarify which aspects the analysis is intended to validate and explain the importance of each.
6. Novelty should be stated more clearly.
The paper should more clearly articulate its novel contributions. In particular, the source-level information attribution applies the existing Shapley-value framework to relatively straightforward categories of 3DGS parameters, making the novelty of this component appear limited. The authors should better clarify what distinguishes the proposed method from existing approaches and where its main novelty lies.
<references>
[1] Li, Haiyi, et al. "OUGS: Active View Selection via Object‐aware Uncertainty Estimation in 3DGS." Computer Graphics Forum. 2026.
[2] Zhu, Jingsen, Silvia Sellán, and Alexander Terenin. "A Bayesian Approach for Task-Specific Next-Best-View Selection with Uncertain Geometry." arXiv preprint arXiv:2605.05095 (2026).
[3] Duan, Hongbo, et al. "DynActiveGS: Active Gaussian Splatting for Dynamic Scene Reconstruction." arXiv preprint arXiv:2608.01178 (2026).
Your trivial decomposition is the right control for what the allocation adds, so we answer it first.
W2. We tested this trivial decomposition directly on PRISM's posterior: computing the log-determinant information gain F_v({b}; R) from only the target block's covariance (the other blocks held as known, with no Shapley averaging) under identical one-hot weights. Across all three static testbeds (Table 6, submitted Table 5) and all three tasks of the new dynamic closed-loop test with real distractors (Section 4.5, Table 2), PRISM outperforms single-block scoring:
On all three static testbeds, single-block scoring is no better than unweighted total information (4.88 vs. 4.85, 39.02 vs. 39.06 dB, 17.38 vs. 17.44 dB), consistent with standalone gains treating the other uncertain blocks as known and double-counting shared pixel evidence (summing to 1.5x I_pers on 3DGS; Figure 7c). Rescaled singletons—normalizing standalone gains to sum to I_pers—also improve on total information (Table 6: 4.83, 39.09 dB, 17.46 dB). PRISM is ahead of them on every static column (4.79, 39.36 dB, 17.51 dB) and in 7 or 8 of 9 dynamic runs (Table 2), though the static margins over rescaled singletons are small.
Running single-block scoring on PRISM's own posterior, Jacobians, and prior isolates the allocation rule cleanly, whereas restricting FisherRF to one parameter block would simultaneously change the uncertainty model (a diagonal Hessian approximation instead of a block log-determinant). A block-restricted FisherRF baseline was not evaluated; this is stated as a limitation in Section 5.
Q1. In Equation (5), P = {G, B, V}, Pi(P) is the set of the six permutations (orderings) of P, and A_b^pi is the set of blocks preceding b in ordering pi. That was our omission, not only a rendering issue: Pi(P) was not defined in the submitted text and is now defined immediately below Equation (5) in Section 3.2.
W1 / Q2. The impression that alpha was tuned came from inaccurate wording in our submitted Reproducibility Statement, which listed alpha among the quantities fixed on validation data. In all task experiments, alpha is simply the un-tuned one-hot vector matching the task definition: alpha^(G) = (1, 0, 0), alpha^(B) = (0, 1, 0), and alpha^(V) = (0, 0, 1), with uniform weights (1/3, 1/3, 1/3) as the total-information control (now written explicitly in Section 3.3, and corrected in the Reproducibility Statement; only the evidence parameters tau, pi_T and the gate constants were fixed on validation data). For sensitivity, the paper reports the uniform and one-hot endpoints on all testbeds (Table 1, including off-diagonal mismatched weights) and one-hot weights under nuisance (Table 2); intermediate alpha values were not run.
Q3. No dataset-level transient mask IoU and no comparison against RobustNeRF's masks were computed. In Figure 3 (bottom row), one selected clutter frame prints IoU 0.65 (with 1% of static pixels flagged) under the persistent-covariance reference c_r versus IoU 0.39 (10% static flagged) under noise only, which is a single-frame illustration rather than a segmentation benchmark. Across probe frames at matched 0.80 transient recall (Section 4.2, Table 4), persistent-region false flags drop from 0.40 (noise only) to 0.39 (trace-matched c_r) and 0.35 (full 3x3 covariance). Here residual normalization supplies Psi for next-view scoring before fitting rather than training masks: in the reported runs Psi enters the selection score only, the posterior is accumulated with sigma_0^2 (not nuisance-aware), and images are trained on unmodified. Removing residual normalization in Table 2 degrades means across Tasks G (29.5 vs. 28.0, 7/9), B (19.7 vs. 20.0 dB, 7/9), and V (16.6 vs. 16.9 dB, 6/9).
Summary of changes: single-block vs. PRISM static (Table 6) and dynamic (Section 4.5, Table 2) comparisons detailed; Pi(P) defined under Eq. (5); one-hot alpha clarified in Section 3.3 and Reproducibility Statement. Thank you for the baseline suggestion; it sharpened what the allocation does and does not buy.