{"id":"de84465b-89a8-4c13-b59d-4021f789845c","arxiv_id":"2607.07083","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"Combining dataset-prior sampling with top-k group active sampling yields more stable optimization and better downstream task performance than prior active subsampling methods across classification, reconstruction, and segmentation tasks.","lead":"This paper improves active subsampling for tasks like MRI and image classification by combining a fixed prior-based sampling pattern with group-based (top-k) active sampling. It matters because faster, smarter data acquisition reduces costs and time in medical imaging and remote sensing.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"Theorem 1 mischaracterizes A-DPS loss as a composition of k task models, but the actual architecture applies f_θ independently at each iteration to progressively sampled inputs — no composition occurs.","rationale":"The reader correctly identified the most load-bearing concern: Theorem 1's modeling of DPS-top-1 as a composition of k task functions does not match the actual A-DPS architecture, where f_θ is applied independently at each iteration to progressively sampled inputs. The product bound ∏L_r arises from this composition assumption, which is structurally incorrect for the implemented system. This weakens the paper's theoretical contribution. However, the empirical results are strong, consistent across four datasets and three task types, and supported by publicly available code. The ablation in Table 3 (Ps=0%) shows that group sampling alone provides gains over A-DPS, and adding prior sampling (Ps>0) provides further improvement, supporting both components of the method. The concern does not invalidate the empirical contribution but does mean the theoretical justification needs revision. CONDITIONAL is appropriate: the method works, but the theory as stated does not accurately describe why. The paper would be substantially strengthened by either (a) correcting Theorem 1 to model the actual loss structure including the sampling network's role in the gradient chain, or (b) reframing the theorem as a simplified illustrative model rather than a direct analysis of the A-DPS architecture.","tokens_in":20405,"tokens_out":3820,"duration_ms":112630,"concrete_test":"Re-derive the Lipschitz bound for the actual A-DPS loss L_total = Σ_j ||t - f_θ(A_j x)||² where A_j depends on A_{j-1} through the sampling network g_j(f_θ(A_{j-1}x)). If the resulting bound on the gradient Lipschitz constant for top-1 does not exceed that of top-k under realistic L_j ≥ 1 assumptions (including the sampling network's Lipschitz constants in the chain), then Theorem 1's conclusion does not hold for the actual architecture and the theoretical claim should be revised or withdrawn.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 1 models the DPS-top-1 loss as ||f(x*) - f_k(x_k)|| where f = f_k∘f_{k-1}∘...∘f_1, yielding a product bound ∏L_r on the Lipschitz constant. But the actual A-DPS architecture (Eqs. 7–8) does not compose task models. At iteration j, the task model f_θ is applied to the partially sampled input A_{j-1}x to produce context t_j = f_θ(A_{j-1}x), which feeds the sampling network g_j to select the next sample. The loss at each iteration is ||t - f_θ(A_j x)||², computed independently against the task target — not against the output of a previous task model application. The total loss is a sum of such terms ('losses are accumulated over all iterations'). There is no chain f_k(f_{k-1}(...f_1(x_1)...)). The gradient path from early samples to later losses does pass through alternating task model and sampling network evaluations, but Theorem 1's product bound ∏L_r captures only task model Lipschitz constants and omits the sampling network entirely. Thus the claimed inequality L_k ≤ ∏L_r does not follow from the actual architecture. The empirical gains (Tables 1–6) are consistent and do not depend on this theorem, but the theoretical contribution — 'a smoother loss landscape and more stable optimization' — rests on a model that doesn't match the implemented system.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes PGA-DPS, an enhancement of Active Deep Probabilistic Subsampling (A-DPS) that combines (i) a deterministic prior-informed sampling pattern derived from training data with (ii) group-based (top-k) active sampling instead of top-1. The central claims are that group sampling yields a smaller effective Lipschitz constant of the loss (Theorem 1), leading to smoother optimization, and that the combination of prior-aware fixed sampling with context-guided active group sampling consistently outperforms A-DPS, DPS, and other baselines. The method is evaluated on MNIST and CIFAR-10 classification, fastMRI knee reconstruction, and AeroRIT hyperspectral segmentation, with ablation studies on the prior/active sampling split and temperature.","tokens_in":20905,"tokens_out":1465,"duration_ms":131256,"significance":"The empirical contribution is substantial: four datasets, three task types, multiple independent baselines (LOUPE, VDS, greedy, RL-based methods), and consistent improvements across all settings. The provision of code, ablation studies on hyperparameters (Tables 3, 7, 8), and a temperature sensitivity analysis (Tables 9, 10) strengthen reproducibility. The practical idea of blending a fixed prior mask with a reduced number of active group-sampling iterations is simple, well-motivated, and shown to be effective, particularly in low-measurement regimes. However, the theoretical contribution (Theorem 1) does not accurately model the implemented A-DPS architecture, as detailed below.","major_comments":[{"comment":"§3.3, Theorem 1 (and Appendix A): The theorem models the DPS-top-1 loss as a composition f = f_k(f_{k-1}(...f_1(x_1)...)), yielding a product bound ∏L_r on the Lipschitz constant. However, the actual A-DPS architecture (Eqs. 7–8) does not compose task models in this manner. At iteration j, the task model f_θ is applied to the partially sampled input A_{j-1}x to produce context t_j = f_θ(A_{j-1}x), which feeds the sampling network g_j. The loss at each iteration is computed independently against the task target, and 'losses are accumulated over all iterations' (§3.2). There is no chain f_k∘f_{k-1}∘...∘f_1. The gradient path from early samples to later losses does pass through alternating task model and sampling network evaluations, but the product bound ∏L_r captures only task model Lipschitz constants and omits the sampling network g_j entirely. Thus the claimed inequality L_k ≤ ∏L_r and","section":null},{"comment":"the conclusion that 'DPS-top-k exhibits a smaller effective Lipschitz constant than DPS-top-1' do not follow from the actual architecture. The empirical gains (Tables 1–6) are consistent and do not depend on this theorem, but the theoretical contribution as stated is not valid. The authors should either (a) revise Theorem 1 to model the actual A-DPS loss accumulation structure (sum of independent per-iteration losses with gradient backpropagation through shared parameters) and re-derive the Lipschitz bound accordingly, or (b) reframe the theorem as a motivating heuristic rather than a formal result, making clear that it describes an idealized composition not present in the implementation.","section":null},{"comment":"§3.3 and Table 3: The optimal (P_s, A_s) configuration varies across tasks: (60%, 20%) for MNIST, (10%, 20%) for CIFAR-10, (30%, 30%) for MRI, and (80%, 20%) for HSI segmentation. The paper provides a heuristic recommendation in §6 (use the DPS vs. A-DPS performance gap as an indicator of Lipschitz characteristics), but this is not validated quantitatively. Given that P_s ranges from 10% to 80% across tasks, the practical guidance for selecting these hyperparameters remains underspecified. A sensitivity analysis showing performance degradation as a function of deviation from the optimal P_s (beyond the partial results in Table 7) would strengthen the claim that the method is robust to hyperparameter choice.","section":null}],"minor_comments":[{"comment":"§3.2, Eq. (8): The notation ϕ_j = g_j(t_j) uses j as both the iteration index and a subscript for the sampling network g_j, but it is unclear whether g_j denotes a distinct network per iteration or the same network with different inputs. Clarification needed.","section":null},{"comment":"Table 3: The entry for P_s=70%, A_s=40% is marked '—' with no explanation. Please state why this configuration was not evaluated.","section":null},{"comment":"Figure 1 caption: 'DPS picks 31 samples in one step, A-DPS over 31 iterations, and PGA-DPS in just 3 iterations' — the number 31 appears task-specific (MNIST at ~4% sampling) but is not defined in the caption context. Consider clarifying.","section":null},{"comment":"§4.1.2: The sentence 'the proportions of prior (deterministic) sampling and active sampling are fixed to 60 and 20 %' is missing a unit or context — should read '60% and 20% of the target samples' for consistency with later sections.","section":null},{"comment":"Appendix B.4: 'tow downsampling operations' should be 'two downsampling operations'.","section":null},{"comment":"Table 2: A-DPS shows a performance decrease beyond r=14% (e.g., 70.4 at r=10% vs. 68.3 at r=20%), which the authors attribute to 'an inflated Lipschitz constant.' This explanation relies on Theorem 1, which as noted above does not accurately model the architecture. An alternative empirical explanation (e.g., overfitting of the sampling network at higher sampling ratios) should be discussed.","section":null},{"comment":"The paper states (§3.2) that 'the DPS-top-k approach showed improved performance' over DPS-top-1, citing Huijben et al. (2020a). Since this is a central motivation, a brief explanation of why top-k outperforms top-1 in the original DPS work would help readers unfamiliar with that result.","section":null}],"recommendation":"major_revision","confidential_remarks":"The empirical results are solid and the method is practical, but Theorem 1 is the paper's stated theoretical contribution and it does not hold as written. The skeptic's concern is well-founded: the composition structure assumed in the theorem does not match the A-DPS architecture described in Eqs. 7–8. If the authors can either fix the theorem to match the actual loss structure or honestly reframe it as a heuristic motivation, the paper would be acceptable. The empirical work alone (four datasets, three tasks, consistent gains, ablations, code) meets the bar for publication, but the theoretical claim as currently stated is incorrect and must not stand uncorrected."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The paper you should know about is PGA-DPS, which takes the existing A-DPS active subsampling framework and makes two changes: (1) it prepends a fixed, prior-informed sampling mask learned from training data, and (2) it replaces top-1 iterative sampling with top-k group sampling. The empirical work is genuinely thorough—four datasets, three task types (classification, reconstruction, segmentation), and comparisons against DPS, A-DPS, LOUPE, VDS, greedy methods, and RL-based approaches. The gains are consistent across all settings, code is public, and the ablations on hyperparameters and temperature are responsible. This is a well-executed applied paper.","headline":"Solid empirical improvement to active subsampling; Theorem 1 doesn't match the actual architecture","tokens_in":21199,"tokens_out":201,"would_cite":true,"duration_ms":72622,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Group sampling plus dataset priors beats one-at-a-time active subsampling","keywords":[],"falsifier":"If one could show that the one-at-a-time active sampling pipeline does not actually compose k separate task models with independently large Lipschitz constants (for instance, because parameters are shared across iterations or the effective composition is closer to a single function), then the product-of-Lipschitz-constants bound in Theorem 1 would not apply, and the theoretical justification for group sampling's smoother loss landscape would collapse. The empirical gains would then need an alternative explanation.","tokens_in":20679,"feed_emoji":"🎯","tokens_out":693,"duration_ms":122214,"temperature":0.7,"pith_summary":"When a machine learning model must decide which measurements to keep from a large signal (an MRI scan, a hyperspectral image, a digit), it can either pick a fixed pattern that works well on average or adaptively pick samples one at a time based on what it has already seen. The adaptive approach is powerful but has two flaws: it ignores general knowledge about what the training dataset looks like, and picking samples one-by-one compounds the nonlinearity of the task model across iterations, making optimization unstable. This paper proposes fixing both flaws simultaneously. First, it reserves a fixed fraction of the sampling budget for a learned deterministic pattern that captures dataset-wide priors. Second, it selects the remaining samples in groups rather than one at a time. The central theoretical claim is that group selection avoids a multiplicative stacking of Lipschitz constants that plagues one-at-a-time selection, yielding a smoother loss landscape. The method, called PGA-DPS, consistently outperforms prior adaptive and fixed sampling methods across image classification, MRI reconstruction, and hyperspectral segmentation.","feed_headline":"Group sampling plus dataset priors beats one-at-a-time active subsampling","feed_subtitle":"Selecting measurements in batches rather than one-by-one smooths the optimization landscape, and adding a fixed prior mask captures global规律","key_machinery":"Gumbel-Softmax reparameterization for differentiable discrete sampling; Gumbel top-k trick for selecting multiple samples simultaneously; a deterministic (fixed) learned sampling mask derived from training-data priors; an LSTM-based sampling network that encodes contextual information from previously selected samples; and a Lipschitz-constant comparison between the compositional loss of top-1 sampling and the single-function loss of top-k sampling.","core_discovery":"The paper identifies a structural reason why selecting samples one at a time in active subsampling is harder to optimize than selecting them in groups: under the assumption that each iteration's task model has Lipschitz constant at least one, the one-at-a-time approach composes k models whose effective Lipschitz constant is the product of k constants, while the group approach uses a single model whose Lipschitz constant is just one term. Because neural network task models typically have Lipschitz constants well above one, this product grows rapidly with the number of iterations, creating a rougher optimization landscape. By combining this group-selection insight with a deterministic prior-in","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Group sampling avoids Lipschitz scaling in active subsampling","Batch selection and fixed priors outperform one-at-a-time sampling","Dataset priors and group sampling improve active subsampling","Group sampling smooths the optimization landscape for subsampling","Selecting measurements in groups smooths subsampling optimization"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The theoretical advantage of group sampling over one-at-a-time sampling rests on modeling the one-at-a-time pipeline as a composition of k independent task models whose Lipschitz constants multiply. If the same network with shared parameters is reused across iterations, this multiplicative bound may not hold and the theoretical gap could dissolve.","fun_headline_variants_meta":{"raw":{"variants":["Group sampling avoids Lipschitz scaling in active subsampling","Batch selection and fixed priors outperform one-at-a-time sampling","Dataset priors and group sampling improve active subsampling","Group sampling smooths the optimization landscape for subsampling","Selecting measurements in groups smooths subsampling optimization","Fixed priors and group sampling enhance active probabilistic subsampling"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":1749,"prompt_tokens":541,"completion_tokens":1208,"prompt_tokens_details":null},"tokens_in":541,"tokens_out":1208,"duration_ms":36536,"temperature":1.0,"reasoning_tokens":1203,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T20:22:24.974830+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"If one could show that the one-at-a-time active sampling pipeline does not actually compose k separate task models with independently large Lipschitz constants (for instance, because parameters are shared across iterations or the effective composition is closer to a single function), then the product-of-Lipschitz-constants bound in Theorem 1 would not apply, and the theoretical justification for group sampling's smoother loss landscape would collapse. The empirical gains would then need an alternative explanation.","supporting_citations":[],"review_version":1}