{"id":"1fe82d70-9dd3-45dc-92f5-8bee7b1f8e82","arxiv_id":"2607.04755","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"TTA-guided model merging on six TCGA cohorts can improve rehearsal-free continual WSI classification and knowledge retention, but remains sensitive to task order and current-test adaptation.","lead":"This paper benchmarks test-time adaptive model merging methods for continual whole-slide image cancer subtyping without storing old slides. It shows they can retain prior knowledge better than many classic continual-learning baselines while remaining sensitive to task order and current-distribution bias.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the paper's own stability-plasticity caveat.","rationale":"The paper is an empirical benchmark that maps three existing TTA-guided merging methods to continual WSI classification and reports multi-cohort, multi-order, IND/OOD results against standard CL baselines. Its strongest claim is carefully hedged by the Abstract and Discussion. The reader's identified weakest assumption matches exactly the stability-plasticity trade-off the authors themselves flag; the tables already supply the quantitative evidence of that trade-off. Because the claim is presented as a promising direction rather than a finished solution, and because the experimental support is multi-faceted (six cohorts, two IL settings, two domain protocols, two orders, F1 variants), no further load-bearing flaw is required to keep the CONDITIONAL verdict. The suggested freeze-old-variables ablation is a useful verification step that would either reinforce or modestly strengthen the paper's own caveat, but does not alter the current assessment.","tokens_in":11115,"tokens_out":535,"duration_ms":6458,"concrete_test":"Recompute CLASS-IL bACC, FGT and BWT for AdaMerging and Hi-Vec on both task orders after freezing all merge variables associated with tasks 1..t-1 during the TTA step at t (i.e., optimize only the newest coefficient/mask/layer). If the FGT gap between original and reverse orders shrinks by >50% while current-task masked bACC stays within 2 points of the reported numbers, the paper's own caveat is confirmed as the dominant remaining risk; otherwise the claim of improved retention is more robust than the Discussion suggests.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (entropy minimization on the current unlabeled stream as sole surrogate for merge-variable adaptation, without an explicit historical-subspace constraint) is already the paper's central self-identified limitation (Section 2.2 Eq. 3; Discussion). Tables 2-4 and the reversed-order ablation already quantify the resulting sensitivity: AdaMerging/Hi-Vec keep FGT low (~2.2-2.7) under the original order yet rise to ~14 under reverse order, while still beating regularization baselines and limited-replay DER++. The strongest claim is therefore framed as a benchmark finding with remaining order/adaptation sensitivity, not as a solved method. No additional load-bearing inconsistency appears in the experimental design, metrics, or mapping of AdaMerging/AdaRank/Hi-Vec to the rehearsal-free WSI protocol.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"This paper presents a benchmark that transfers three test-time adaptive model-merging methods (AdaMerging, AdaRank, Hi-Vec) to rehearsal-free continual whole-slide image (WSI) classification. Independently fine-tuned task models are consolidated in parameter space, with merge variables (coefficients, singular-component masks, or hierarchical layers) adapted by Shannon entropy minimization on the current unlabeled test stream (Eq. 3). Evaluation uses six TCGA cancer-subtyping cohorts under CLASS-IL and TASK-IL, in-domain and site-held-out OOD splits, two task orders, and standard continual metrics (bACC, Macro/Weighted F1, FGT, BWT), with comparisons to LwF, EWC, DER++ (10/30 WSI buffers), naive fine-tuning, and a fully supervised upper reference. The authors report strong task-specific performance and improved retention relative to regularization and limited-replay baselines without storing historical WSIs, while documenting sensitivity to task order and to adaptation driven only by the current distribution.","tokens_in":11321,"tokens_out":1360,"duration_ms":23958,"significance":"Continual computational pathology is practically important: gigapixel WSIs and institutional constraints make rehearsal and centralized retraining difficult. Mapping multi-task/TTA merging methods into a strict rehearsal-free sequential protocol, with multi-cohort TCGA evaluation, IND/OOD site splits, dual task orders, and both CLASS-IL and TASK-IL, is a useful contribution to the field. The experimental design is relatively strong for a benchmark (10-fold CV, means±std, complementary F1 metrics, forgetting/BWT). The paper is appropriately cautious: it frames TTA-guided merging as promising rather than solved, and the Discussion explicitly identifies the stability–plasticity trade-off arising from entropy minimization on the current stream alone. If the reported retention advantages hold under broader orders and against static merging, this would motivate a concrete research line—adaptive merging with explicit historical-subspace protection—for lifelong pathology models.","major_comments":[{"comment":"The central claim that adapting merge variables at test time improves continual WSI performance is not isolated from static merging. MergeSlide is cited as the fixed-rule continual merging baseline that TTA is meant to extend (Introduction; §2), yet neither MergeSlide nor simple task-arithmetic / fixed-coefficient merging appears in Tables 2–4. Without that control, gains over LwF/EWC/DER++ cannot be attributed to TTA rather than to model merging itself. Please add at least one static merging baseline under the same protocol (same backbone, K-patch sampling, heads, and task orders).","section":"Introduction; Tables 2–4"},{"comment":"Section 1 poses whether repeated TTA update-and-reconstruct of merge variables justifies its computational cost, but the experiments report no wall-clock time, TTA steps, or memory overhead relative to EWC/LwF/DER++ or static merging. For a deployment-oriented pathology benchmark this is load-bearing: if TTA is expensive per test stream, the practical recommendation changes. Please report cost (e.g., seconds per task / per fold on the A100) alongside accuracy.","section":"Section 1; Section 3 Implementation Details"},{"comment":"Only two task sequences are evaluated (B→R→N→E→T→C and its reverse; Tables 2–3). FGT for AdaMerging/Hi-Vec rises from ~2.2–2.7 to ~14 under reversal, which the paper correctly flags as order sensitivity, but two orders are a thin basis for that conclusion with six heterogeneous cohorts (common vs rare, Tab. 1). At least one additional order (e.g., rare-first or interleaved common/rare) would make the sensitivity claim more robust and better support the Discussion’s call for historical-subspace protection.","section":"Tables 2–3; Section 4 Discussion"}],"minor_comments":[{"comment":"Page headers repeatedly show “ZeroSlide,” which does not match the title or content and looks like a leftover template string; please correct.","section":"Running header"},{"comment":"Several spacing/typo issues in the abstract and introduction (e.g., “Recentstate-of-the-art,” “test-timeadaptivemerging,” “withoutstoringhistoricalWSIs,” “Sequentialfine-tuningissusceptible”). A careful copy-edit pass is needed.","section":"Abstract; Section 1"},{"comment":"Fig. 1 is dense; the three panels (regularization, rehearsal, TTA-guided merging) would be clearer with a short caption walkthrough of symbols (R_k, ξ_t, λ/B/ϕ*).","section":"Fig. 1"},{"comment":"Hi-Vec’s continual mapping (dynamic layer selection, δ_OOD, δ_sim, encoder updates) is compressed in §2.2; a short algorithm box or pseudocode for one continual step would improve reproducibility.","section":"Section 2.2 Hi-Vec"},{"comment":"State the value of K (sampled patches) and the TTA optimizer/step budget explicitly in Implementation Details; they are free parameters of the protocol but not numerically listed in the main text.","section":"Section 3 Implementation Details"},{"comment":"Clarify whether task-specific linear heads are ever merged or only concatenated at CLASS-IL inference; a one-sentence statement would remove ambiguity for multi-head setups.","section":"Section 3 Implementation Details"}],"recommendation":"major_revision","confidential_remarks":"Fit is appropriate for a CV/medical-imaging venue that accepts rigorous benchmarks. Novelty is primarily in the mapping and multi-cohort evaluation rather than a new algorithm; that is fine if framed as a benchmark, but the missing static-merging control is the main reason I prefer major_revision over minor. Mild author-lineage overlap with MergeSlide is disclosed via citation and does not appear circular given the external TTA methods and standard CL baselines. No integrity concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a careful domain-transfer benchmark, not a new algorithm. The useful finding is that AdaMerging, AdaRank, and Hi-Vec, when run sequentially on six TCGA subtyping cohorts with no stored slides, can deliver strong current-task accuracy and lower forgetting than EWC/LwF and limited-replay DER++ under both CLASS-IL and TASK-IL, IND and OOD site splits, and two task orders.\n\nWhat they did well: the experimental design is thorough for the subfield. Same TITAN backbone and aggregator, 10-fold CV, fully-supervised and naive fine-tuning references, Macro/Weighted F1 for imbalance, FGT/BWT, and an explicit reverse-order ablation. Tables 2–4 and Figure 2 make the retention story concrete—AdaMerging/Hi-Vec keep FGT near 2 under the original order and still beat regularization baselines when order is reversed—while the Discussion openly flags the stability–plasticity issue that entropy minimization on the current stream can bias earlier task vectors.\n\nSoft spots are real but proportionate. Novelty is moderate: the methods are imported; the contribution is the sequential WSI protocol and the multi-cohort comparison. Free parameters (K, Ne, Hi-Vec thresholds, TTA schedule) are only lightly disclosed, and no code or exact splits are released, so independent re-run is not yet possible. The central claim is correctly framed as a benchmark result with remaining order sensitivity, not as a solved continual learner. That matches the stress-test note; I do not see an unacknowledged load-bearing flaw.\n\nWho it is for: people working on privacy-constrained continual pathology or model merging under distribution shift. They will get a usable map of what works and what still breaks. It deserves a serious referee; I would accept for peer review with the usual requests for code, fuller hyper-parameter tables, and tighter framing as a benchmark. Worth engaging if you care about rehearsal-free medical continual learning.","headline":"Solid rehearsal-free WSI benchmark that maps TTA merging methods into continual pathology and shows real retention gains, with the order-sensitivity caveat already owned by the authors.","tokens_in":11975,"tokens_out":492,"would_cite":true,"duration_ms":5454,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Test-time adaptive model merging retains prior cancer-subtyping knowledge on whole-slide images without storing old slides.","keywords":["continual learning","model merging","test-time adaptation","whole-slide images","computational pathology","catastrophic forgetting","TCGA cancer subtyping"],"falsifier":"A controlled run on the same six TCGA folds in which entropy-driven coefficient or mask updates are replaced by a fixed or history-aware merge rule; if forgetting and current-task accuracy then become no better than (or worse than) plain sequential fine-tuning, the claimed benefit of pure test-time adaptation disappears.","tokens_in":11978,"feed_emoji":"🔬","tokens_out":628,"duration_ms":5360,"temperature":0.7,"pith_summary":"Pathology labs cannot keep every prior whole-slide image when a new cancer-subtyping task arrives, yet sequential fine-tuning forgets earlier cohorts. This benchmark study shows that modern model-merging techniques, which combine independently trained task models in parameter space and then adapt the merge coefficients, masks, or hierarchical layers on unlabeled current test slides, can deliver strong accuracy on the new task while preserving earlier knowledge far better than classic regularization methods and competitively with limited-replay baselines. Across six TCGA cohorts, both in-domain and out-of-domain sites, and two task orders, the approach works under both class-incremental and task-incremental protocols. The practical payoff is a rehearsal-free pipeline that never stores historical patient slides yet still reduces forgetting. The same experiments also reveal that the gains remain sensitive to the order of tasks and to how aggressively the current distribution is allowed to reshape the accumulated merge.","feed_headline":"Merging models at test time keeps old cancer knowledge without old slides","feed_subtitle":"Six TCGA cohorts show lower forgetting than classic continual learning while matching small-buffer replay.","key_machinery":"TTA-guided continual model merging: independently fine-tuned task vectors (or their SVD components or hierarchical heads) are combined with method-specific merge variables that are optimized by Shannon-entropy minimization on the current unlabeled test stream before the final merged weights are used for inference.","core_discovery":"When state-of-the-art test-time adaptive merging methods (AdaMerging, AdaRank, Hi-Vec) are transferred to a rehearsal-free continual stream of whole-slide cancer-subtyping tasks, they achieve strong current-task performance and substantially lower forgetting than regularization-based continual-learning baselines, without retaining any historical WSIs, while remaining competitive with small-buffer replay under both CLASS-IL and TASK-IL protocols.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Test-time merging retains cancer knowledge without old slides","TTA model merging cuts forgetting in continual WSI subtyping","Adaptive merging rivals small-buffer replay without past WSIs","Merged models adapted at test time hold prior TCGA knowledge","Test-time merged models keep task knowledge across cancer cohorts"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That minimizing prediction entropy on only the newest unlabeled test slides is enough to adapt the merge without an explicit safeguard that protects subspaces already learned for earlier tasks.","fun_headline_variants_meta":{"raw":{"variants":["Test-time merging retains cancer knowledge without old slides","TTA model merging cuts forgetting in continual WSI subtyping","Adaptive merging rivals small-buffer replay without past WSIs","Merged models adapted at test time hold prior TCGA knowledge","Test-time merged models keep task knowledge across cancer cohorts"]},"model":"grok-4.5","effort":"low","cost_usd":0.003866,"raw_usage":{"total_tokens":1231,"prompt_tokens":784,"num_sources_used":0,"completion_tokens":64,"cost_in_usd_ticks":38660000,"prompt_tokens_details":{"text_tokens":784,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":383,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":784,"tokens_out":64,"duration_ms":3745,"temperature":1.0,"reasoning_tokens":383,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T14:01:59.501752+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A controlled run on the same six TCGA folds in which entropy-driven coefficient or mask updates are replaced by a fixed or history-aware merge rule; if forgetting and current-task accuracy then become no better than (or worse than) plain sequential fine-tuning, the claimed benefit of pure test-time adaptation disappears.","supporting_citations":[],"review_version":1}