{"id":"b33bb520-e2cf-43cb-84b4-6e63fd075232","arxiv_id":"2501.18984","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Querent selects top-K regions per query patch using min/max region summaries and reports state-of-the-art results on 11 whole-slide image datasets.","lead":"A new method, Querent, lets each patch in a whole-slide image attend only to the regions predicted most relevant to it, instead of attending to all patches at once. The authors report better accuracy and survival predictions than prior methods on over 10 pathology datasets, with a claim of much lower compute.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Appendix H.3 and Algorithm 1 imply O(N^2 d/K) region-importance cost, contradicting the central O(NK) near-linear efficiency claim.","rationale":"The reader's stated weakest assumption focuses on Lemma B.4's claim that the min/max selection score s_i upper-bounds true patch interactions; that is a real and valid concern about the approximation theorem. However, I identify a more fundamental and internally verifiable defect: the algorithm as written and as analyzed in the paper's own appendix is quadratic in the number of patches because the region importance estimation phase computes scores for all N queries against all M = N/K regions. This directly contradicts Section 3.2.3's claim of O(NK) complexity and the abstract's promise of practical efficiency. The consequence is that the central claim fails even if every lemma about the approximation bound were correct. I therefore agree with the reader's REJECT verdict but for a different primary reason, hence 'partial' agreement on the weakest assumption. The empirical results over 11 datasets are not in question here; the issue is that the method does not deliver the complexity reduction that motivates it. A concrete computational test on the released code would settle the matter quickly and is essential before any revision can be assessed.","tokens_in":23917,"tokens_out":2945,"duration_ms":29642,"concrete_test":"Run the released code with fixed region size K=16 on synthetic patch sequences of length N = 4k, 8k, 16k, and 32k, measuring wall-clock time of Phase 2 (region importance estimation) only. If runtime grows roughly quadratically (log-log slope near 2), the O(NK) claim is false. If it grows linearly, inspect whether the code actually computes scores for all N queries against all M regions; if not, document the missing shortcut and update Algorithm 1 and Appendix H accordingly. A simpler analytical check: count multiply-adds in Algorithm 1 Phase 2 — N queries times M regions equals N(N/K) dot products, which is O(N^2/K) for any K independent of N.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim is that Querent reduces full self-attention from O(N^2) to O(NK) while keeping global context, making gigapixel-scale WSI analysis practical. This claim is contradicted by the paper's own complexity analysis and pseudo-code. Algorithm 1 Phase 2 loops 'for each query patch q' and 'for i = 1 to M' computing s_i via dot products between the projected query and every region's metadata. With N query patches and M = N/K regions, this is N * M = N^2/K dot products, each costing O(d), i.e., O(N^2 d / K) — quadratic in N for fixed K. Appendix H.2.2 explicitly reports region importance estimation as O(N R d), and H.3 aggregates the total as O(N d + N R d + N k p d). Since R = N/p, this simplifies to O(N^2 d / p) for constant p and k, not O(N K) as stated in Section 3.2.3. The method is therefore quadratic in N unless Phase 2 is omitted or approximated in a way not described anywhere in the manuscript. Figure 6's near-linear scaling plot is thus unexplained: it either measures only the selective-attention phase or relies on an implementation shortcut that is absent from the algorithm and complexity analysis. Because the entire motivation for the method is avoiding quadratic attention cost, this internal inconsistency invalidates the central practical-efficiency claim regardless of the approximation theorem's status. The theorem itself also rests on the unproven bound in Lemma B.4, but the complexity contradiction is the more directly falsifiable and load-bearing defect.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Querent, a transformer-style MIL framework for whole-slide images. The method partitions patch tokens into spatial regions, summarizes each region by learned projections of element-wise minimum and maximum feature vectors, scores each query patch against all region summaries, and then computes dense self-attention only over the top-K regions for that query. The authors claim this reduces self-attention complexity from O(N^2) to O(NK), provide a theorem stating that the query-aware attention matrix approximates full self-attention with bounded error, and report state-of-the-art accuracy, AUC, F1, and C-index results across 11 public WSI datasets. The appendix contains pseudo-code, a complexity analysis, and the proof of the approximation theorem.","tokens_in":24273,"tokens_out":4447,"duration_ms":42263,"significance":"If the efficiency and approximation claims were valid, Querent would be a practically useful contribution to computational pathology, combining global context with sub-quadratic cost. The paper's strengths are its broad empirical evaluation across biomarker prediction, mutation prediction, subtyping, and survival analysis, its release of code, and its inclusion of a self-contained appendix with pseudo-code and complexity derivations. However, the central efficiency claim is internally contradicted by the paper's own complexity equations, and the theoretical guarantee is stated for random projections rather than for the trained model actually evaluated. These issues are load-bearing: they undermine the stated motivation for the method, independent of the empirical tables.","major_comments":[{"comment":"The claimed reduction to O(NK) is contradicted by the paper's own complexity analysis. Algorithm 1 Phase 2 loops \"for each query patch q\" and, inside that loop, \"for i = 1 to M\", computing a dot product for every query-region pair. Appendix H.2.2 reports this region importance estimation as O(NRd), and Appendix H.3 aggregates the total as O(Nd + NRd + Nkpd). Since R = N/p, this simplifies to O(N^2 d/p), which is quadratic in N for constant p and k, not O(NK) as stated in Section 3.2.3. Consequently, Figure 6's near-linear scaling is not explained by the described algorithm; it either measures only the selective-attention phase or relies on an implementation shortcut that is not described anywhere in the manuscript.","section":"Section 3.2.3, Appendix H.2.2, H.3, Algorithm 1"},{"comment":"The proof of Lemma B.4 uses the element-wise inequality m_min <= x_j <= m_max for raw features, but the importance score in Eq. (3) is computed after learned projections fmin and fmax. There is no guarantee that fmin(m_min) <= fmin(x_j) <= fmax(m_max) holds element-wise, and the maximum of absolute dot products with projected metadata is not an upper bound on |<q, x_j>| for arbitrary learned query features that may have negative components. The first step of Lemma B.4 therefore fails, and the ranking-stability Lemma B.5 and Theorem B.6 inherit this gap. The Lipschitz constant L of the learned projections is also an assumed free parameter that is never specified or justified for the trained networks.","section":"Appendix B.1.2, Lemma B.4, Eq. (3)"},{"comment":"The theorem is stated for \"random projection matrices WQ, WK in R^{d x d}\" and uses Johnson-Lindenstrauss inner-product preservation, but the model in Section 3.2.3 uses a learned Wqkv and learned projections fq, fmin, fmax. The proved object is therefore not the model that is trained and evaluated, so the abstract's claim of \"a theoretically bounded approximation of full self-attention\" is not supported for Querent as implemented. Additionally, the exponential spatial-decay assumption on attention scores is not established for the learned features, so the theorem's applicability to the actual setting is unverified.","section":"Theorem 3.1 / Theorem B.6"}],"minor_comments":[{"comment":"The symbol K is used both for the number of patches per region and for the number of selected regions per query, while Appendix H uses p and k for the same concepts; this notational collision makes the complexity claims difficult to verify.","section":"Section 3.2.1 and 3.2.2"},{"comment":"The phrase \"10, 0002 ~ 100, 0002 pixels\" appears to be a typo for \"10,000^2 to 100,000^2 pixels\" and should be corrected.","section":"Introduction"},{"comment":"Several cells in Tables 1 and 2 contain triplicated values (for example, the Querent row on BCNB-ER and the DTFD row on KIRC), which appears to be a table-generation error that should be fixed.","section":"Tables 1 and 2"},{"comment":"The statement of Theorem 3.1 in the main text contains a dangling clause, \"The hidden dimension satisfies, and the number of selected regions per query satisfies:\", before the numbered conditions; the quantifiers should be stated explicitly.","section":"Theorem 3.1"},{"comment":"Appendix G.2 describes a hybrid loss combining binary classification and ranking components for training the importance estimation module, but Section 3.2 and Algorithm 1 describe only end-to-end training with task losses; this discrepancy should be reconciled.","section":"Appendix G.2 vs. Section 3.2"},{"comment":"The caption reports P<0.005 but does not state which statistical test was used or how many repeated runs were compared; this information should be included in the caption or in Appendix G.1.","section":"Figure 4"}],"recommendation":"reject","confidential_remarks":"The paper's own Appendix H gives O(N^2 d/p) for the complete algorithm, so the central near-linear efficiency claim is internally inconsistent. The theoretical result also concerns random projections rather than the trained model. These are not local presentation issues; they invalidate the main contribution as framed. I did not treat the triplicated entries in Tables 1 and 2 as evidence of misconduct; they look like formatting artifacts, but they should be corrected in any future version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Zhengrui and colleagues have a real idea here: instead of fixed local-global attention for gigapixel WSIs, score each region with learned min/max metadata against the query, and run dense attention only on the top-K regions. That is genuinely new in the MIL/transformer WSI literature they cite, and the min/max scoring mechanism is concrete and ablatable. The experiments are the paper's strongest asset: 11 public datasets, 5-fold CV, standard baselines, ablations against random selection and a side network, and released code. The gains are modest but consistent (average C-Index 0.670 vs 0.647; AUC gains in the low single digits), and the ablation shows the query-aware selection adds value over random region selection, especially on UBC-OCEAN. Nothing about the empirical section feels cooked.\n\nThe soft spots are large and load-bearing. The headline claim—near-linear O(NK) complexity approximating full self-attention—is contradicted by the paper's own complexity analysis. Algorithm 1 Phase 2 scores every query against every region: N queries times M=N/K regions, which is O(N^2/K) dot products, and Appendix H.2.2 states O(NRd) with R=N/p. The total in H.3 is O(Nd+NRd+Nkpd)=O(N^2 d/p) for constants. That is quadratic in N, not near-linear. The 'chunking' discussion and Figure 6's near-linear plot do not account for this phase, so the practical-efficiency motivation collapses. The approximation theorem is also unsupported: Lemma B.4 assumes the element-wise min/max bound carries through the learned projections fmin/fmax, but those projections break the element-wise comparison, and the proof mixes original and projected feature spaces. The theorem may be repairable with a different argument, but as written it does not prove what it claims.\n\nThese are not minor technicalities; they are the paper's central claims. That said, the empirical contribution could still stand if the authors honestly report the quadratic region-scoring cost and compare against other subquadratic alternatives, or find a genuinely subquadratic importance estimator (e.g., hashing or coarse clustering). As written, the paper overclaims.\n\nWho should read it: anyone working on efficient attention for gigapixel pathology, for the experimental design and the query-adaptive idea—skim the theory, distrust the complexity claims. This deserves a serious referee: the idea is novel and the experiments are extensive enough that the issues could be fixed in revision. I would send it out, but with the expectation that the main claims need either a real proof, a real subquadratic algorithm, or a repositioned paper.","headline":"The empirical core is honest and the query-adaptive selection idea is worth a look, but the paper's central complexity claim is contradicted by its own appendix and the approximation theorem does not go through.","tokens_in":24803,"tokens_out":3731,"would_cite":false,"duration_ms":35320,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Querent claims a query-aware selective attention mechanism keeps full self-attention's expressiveness within a bounded approximation error while reducing WSI modeling cost from O(N^2) to O(NK), and reports state-of-the-art results on 11…","keywords":["whole slide image analysis","query-aware attention","self-attention approximation","computational pathology","gigapixel images","multiple instance learning","long sequence modeling","survival prediction"],"falsifier":"On real WSI patch features, compute for many (query, region) pairs the gap si minus the true maximum interaction max_j |<q, x_j>|; if a non-negligible fraction of gaps are negative, the selection rule can miss the truly relevant regions, contradicting Lemma B.4. A simpler end-to-end check is to compare the regions Querent selects to the true top-K regions from full attention on a held-out slide, since low overlap means the bound is not doing the work.","tokens_in":23739,"feed_emoji":"🔬","tokens_out":5438,"duration_ms":47374,"temperature":0.7,"pith_summary":"Querent is a framework for analyzing whole-slide pathology images that tries to keep the modeling power of full transformer self-attention while escaping its quadratic cost. The central move is to let each patch decide, through cheap region-level summaries, which parts of a gigapixel slide are worth attending to, and then run exact attention only on those selected regions. The paper claims this query-aware selection produces a theoretically bounded approximation of full self-attention and reduces complexity from O($N^{2}$) to O(NK), and reports state-of-the-art accuracy, AUC, F1, and C-index on 11 public datasets across biomarker prediction, mutation prediction, subtyping, and survival analysis. If correct, the method would make whole-slide attention practical at 100,000-patch scale with roughly 1 GB memory and 500 GFLOPs, versus 37 GB and 10,000 GFLOPs for full attention.","feed_headline":"Attention for gigapixel slides drops from quadratic to near-linear","feed_subtitle":"Each patch attends only to its top relevant regions, so whole-slide attention runs at 100k patches with a fraction of the memory.","key_machinery":"The load-bearing object is the region-level min/max metadata pair (m_min_i, m_max_i) and the score si defined in Eq. (3). The min and max vectors are meant to bracket all patch features in a region, so that the maximum absolute dot product between a projected query and these two summary vectors upper-bounds the relevance of any patch in the region. That bound is what lets the top-K selection in Eq. (4) act as a cheap proxy for full attention, and it is the first step of Lemma B.4 that feeds Theorem B.6.","core_discovery":"The paper's central claim is that dynamic, query-dependent sparsification does not sacrifice full self-attention's expressiveness. For each query patch q, Querent partitions the slide into regions, computes per-region element-wise min and max summary vectors, projects them with learned networks, and scores every region by si = max(|<q_hat,m_min_hat>|, |<q_hat,m_max_hat>|). The top-K regions are then used to compute dense multi-head self-attention between the query and all patches in those regions, giving each patch its own attention pattern. The paper proves Theorem B.6, a Frobenius-norm bound ||A-B||_F <= (2 + B/sqrt(d)) epsilon between the query-aware attention matrix and the full self-attention matrix, under conditions on region diameter, spatial separation, attention decay exponent, and embedding dimension, and it reports that this bounded approximation outperforms prior MIL and transformer baselines on all evaluated tasks.","pith_inferences":["Editorial: The guarantee is only as strong as the element-wise min/max bracketing after learned projections; if the projections break the inequality, the top-K selection can silently drop relevant regions, so a diagnostic check on si versus true max interaction would settle the practical validity.","Editorial: The same selective-attention idea could transfer to other gigapixel-scale imagery, such as satellite or electron-microscopy data, where local relevance is context-dependent.","Editorial: An even cheaper variant would score regions with the raw min/max vectors before projection; comparing that variant's top-K overlap with Querent's would isolate how much the learned projections help or hurt selection."],"forward_implications":["Whole-slide transformers can process 100k+ patches with near-linear cost, opening gigapixel-scale self-attention to routine use.","The method yields consistent gains over prior attention-based MIL baselines, such as an average C-index of 0.670 versus 0.647 for the second-best method across eight TCGA survival tasks.","Query-aware selection is a general recipe: any task where relevant context varies per query could adopt the region-summarization-and-top-K attention pattern.","The theoretical bound, if it holds, gives users a worst-case guarantee rather than an empirical heuristic for sparse attention."],"supporting_citations":[{"why":"Supplies the Johnson-Lindenstrauss inner-product preservation bound used in Theorem B.6's dimension condition.","marker":"(Kaban, 2015)"},{"why":"Defines the full self-attention mechanism that Querent approximates.","marker":"(Vaswani, 2017)"},{"why":"The linear-attention MIL baseline (TransMIL) that Querent's theoretical section positions against.","marker":"(Shao et al., 2021)"},{"why":"The local-global attention baseline (HIPT) that motivates the need for dynamic query-aware selection.","marker":"(Chen et al., 2022)"},{"why":"PLIP, the pretrained encoder that supplies the 512-dimensional patch features in all experiments.","marker":"(Huang et al., 2023)"},{"why":"ABMIL, the attention-based MIL baseline and a component of the aggregation stage comparison.","marker":"(Ilse et al., 2018)"},{"why":"LongMIL, the local-global transformer baseline compared in all experiments.","marker":"(Li et al., 2024a)"},{"why":"MambaMIL, the state-space baseline that the survival analysis comparison must beat.","marker":"(Yang et al., 2024)"}],"fun_headline_variants":["Gigapixel slides: query-aware attention cuts compute dramatically","Adaptive region selection keeps full attention accuracy at lower cost","Near-linear attention for gigapixel images via dynamic sparsification","Each patch attends to its important regions, making gigapixel feasible"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes the min/max summary of a region is a true upper bound on how relevant any patch in that region is to the query, so that the top-K regions chosen by that score are nearly the right ones; if the learned projections break that element-wise bound, the approximation guarantee no longer follows.","fun_headline_variants_meta":{"raw":{"variants":["Gigapixel slides: query-aware attention cuts compute dramatically","Adaptive region selection keeps full attention accuracy at lower cost","Near-linear attention for gigapixel images via dynamic sparsification","Each patch attends to its important regions, making gigapixel feasible"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000436,"raw_usage":{"total_tokens":2219,"prompt_tokens":949,"completion_tokens":1270,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":1198}},"tokens_in":565,"tokens_out":1270,"duration_ms":11313,"temperature":1.0,"reasoning_tokens":1198,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:45:30.919034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On real WSI patch features, compute for many (query, region) pairs the gap si minus the true maximum interaction max_j |<q, x_j>|; if a non-negligible fraction of gaps are negative, the selection rule can miss the truly relevant regions, contradicting Lemma B.4. A simpler end-to-end check is to compare the regions Querent selects to the true top-K regions from full attention on a held-out slide, since low overlap means the bound is not doing the work.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Johnson-Lindenstrauss inner-product preservation bound used in Theorem B.6's dimension condition."}],"review_version":1}