{"id":"7fa1b012-e2c8-479a-90e6-6e0f27092112","arxiv_id":"2507.02395","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"CoMEL improves continual multiple-instance learning on pathology slides, reporting higher bag-level accuracy and localization overlap with less forgetting than prior continual MIL methods.","lead":"This paper presents CoMEL, a training framework for whole-slide pathology images that learns new organ or tumor tasks one after another while remembering old ones, and also marks where tumors are on the slide. It combines efficient attention, self-generated pseudo-labels, and orthogonal low-rank updates, and reports gains of up to 11 points in slide-level accuracy and 23.4 points in tumor-region overlap over prior continual-learning methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BPPL's pseudo-labels are generated from the same attention maps the method is trying to improve; if initial localization is poor, Eqs. (7)-(8) reinforce the model's own errors, so the reported IoU gains may not transfer outside the tuned setting.","rationale":"The reader's weakest assumption identifies exactly the point on which the central claim rests: the claimed localization improvements come from BPPL, and BPPL's pseudo-labels are produced by the model's own attention. If that attention is not already good, the pseudo-labels are noisy and the instance loss can reinforce the model's mistakes. The paper provides ablations showing BPPL helps on the evaluated benchmarks, but those ablations are performed with the same evaluation metric used to select hyperparameters in Secs. S2.5-S2.6, and Fig. S1(a) does not quantify initial pseudo-label quality. This is a correctness risk, not an internal inconsistency: the method is coherent and the experiments are internally consistent, but the evidential weight of the localization claim depends on an assumption that is stated rather than demonstrated. I agree with the reader that the appropriate verdict is conditional on independent re-implementation with released code, real normal-slide data, and a check of whether the gains persist when initial localization quality is varied.","tokens_in":26197,"tokens_out":8214,"duration_ms":104250,"concrete_test":"Run CoMEL on the same five-organ sequence but initialize the re-embedding module from a model trained on a different organ, or randomly perturb the attention maps used by BPPL for the first 10 epochs, and compare IoU against the reported CoMEL and GDAT-only numbers. If the IoU advantage over GDAT-only shrinks or reverses under this degraded-initialization condition, then the localization claim depends on the assumption that attention is already reasonably accurate, and the self-training loop in Eqs. (7)-(8) is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is in Sec. 3.3: BPPL assumes the MIL attention already localizes positive instances. The pseudo-labels y_hat_m in Eq. (7) are derived from bag prototypes x_bag and x_tilde_bag built from the same attention A that Eq. (8) then trains, so the instance loss is a self-training loop. If attention is initially poor, for example high scores on normal patches in a tumor slide, the positive prototype is contaminated; L_sep in Eq. (6) only separates the two prototypes from each other, and the filtering indicator 1(Y_hat = Y) removes nothing as long as the bag-level prediction is correct. The paper's own Sec. S5 explicitly acknowledges that BPPL 'heavily relies on the accuracy of attention scores as pseudo-labels' and that performance is highly sensitive to hyperparameter selection. The pseudo-label accuracy curve in Fig. S1(a) does not report initial accuracy or compare with a no-BPPL control, so it cannot rule out the possibility that the localization gains come from a favorable initialization plus evaluation-set hyperparameter tuning rather than from a robust mechanism. Since the central claim is enhanced localization with minimal forgetting, this is the place where the argument is least secure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoMEL, a continual multiple instance learning (MIL) framework for whole slide image analysis that jointly addresses bag-level classification and instance-level localization. It introduces three components: Grouped Double Attention Transformer (GDAT) for scalable instance feature refinement, Bag Prototypes-based Pseudo-Labeling (BPPL) for generating instance pseudo-labels from attention and bag prototypes, and Orthogonal Weighted Low-Rank Adaptation (OWLoRA) for rehearsal-free continual learning with orthogonal low-rank bases. The method is evaluated on sequential organ datasets from CAMELYON-16 and PAIP for tumor detection and on TCGA for tumor subtyping. The main empirical claims are that CoMEL outperforms prior continual learning baselines in both bag-level accuracy and localization metrics (e.g., ACCinst 72.64, IoU 41.87, Dice 51.44, ACCbag 62.96 in Table 1) while incurring the least forgetting. The supplementary material includes ablations, efficiency analyses, reversed task order experiments, and an explicit limitation statement about BPPL's reliance on attention quality.","tokens_in":1865,"tokens_out":2114,"duration_ms":96227,"significance":"If the reported results hold, CoMEL fills a genuine gap: continual learning for MIL has mostly focused on bag classification, whereas localization under non-stationary tasks is clinically important and relatively unexplored. The evaluation is extensive, with 10-run comparisons in two task orders, ablations of all major components, and robustness checks across feature extractors. The pipeline is rehearsal-free, which is attractive for privacy-sensitive medical deployments, and OWLoRA is described clearly. The main uncertainty is whether BPPL's self-training loop is reliable beyond the tuned setting; this is a load-bearing issue for the paper's central claim of \"enhanced localization\" and is acknowledged in the paper's own Sec. S5.","major_comments":[{"comment":"The localization gain attributed to BPPL rests on a self-training loop: the pseudo-labels in Eq. (7) are derived from the attention A and the prototypes x_bag and x_tilde_bag, and Eq. (8) then trains the same attention mechanism on those pseudo-labels, filtered only by bag correctness and confidence. The paper's own Sec. S5 states that BPPL \"heavily relies on the accuracy of attention scores as pseudo-labels,\" and Tables S7-S8 show high sensitivity to T, tau, lambda1, and lambda2. The current evidence in Fig. S1(a) reports pseudo-label accuracy during training but gives no initial value and no comparison against a no-BPPL control, so it cannot rule out that the gains in Table 3 (ACCinst 79.27 vs. 68.36, IoU 49.32 vs. 36.31) are driven by favorable initialization plus hyperparameter selection. Please add: (i) pseudo-label accuracy and its agreement with ground truth at task start and at each task boundary; (ii) an ablation that applies the same filtering and loss weighting to attention scores without the prototype-based re-labeling; and (iii) an analysis of how performance changes when initial attention is deliberately degraded, to test whether BPPL corrects or amplifies errors. These analyses are necessary to support the claim of \"reliable instance pseudo-labeling.\"","section":"Sec. 3.3, Eqs. (5)-(9)"},{"comment":"The negative bags for the PAIP organs are constructed by randomly splitting tumor slides into two halves and deleting all tumor regions from the \"normal\" half. Thus, the normal bags are not independent normal WSIs but normal tissue extracted from cancer patients, and the binary tumor/normal split within each organ is artificial. The reported IoU and ACCinst for the PAIP tasks in Table 1 may therefore reflect this synthetic distribution rather than a realistic deployment distribution. Please clarify whether the official PAIP challenge provides truly normal slides, report results on them if available, and otherwise discuss this as an explicit limitation in the main text. This is load-bearing because the combined CM-16+PAIP sequence is the primary evidence for continual instance localization.","section":"Sec. 4.1 and Table S1"},{"comment":"The implementation section states that the same hyperparameters are used for all tasks, but no validation protocol is described. The ablations show that ACCinst ranges from 60.32 to 74.15 as T varies, from 61.23 to 74.15 as tau varies, and from 61.23 to 74.15 as lambda1 varies (Tables S7-S8), yet the chosen values are reported only as final settings. Because several baselines in Table 1 use the same GDAT+BPPL backbone, the margin of CoMEL over InfLoRA (IoU 41.87 vs. 31.87) might partly reflect hyperparameter tuning rather than the continual-learning mechanism. Please specify how (T, tau, lambda1, lambda2, eta, epsilon, d, lambda3) were selected, for example using a held-out validation set disjoint from the test slides, and provide a robustness table over a range of reasonable values for the main continual sequence.","section":"Sec. 4.1 and Tables S7-S8"}],"minor_comments":[{"comment":"The equation labeled Forgetinst is defined using Rbag_{n,l} and is used for bag-level forgetting; it should be labeled Forgetbag. The typo is confusing because the same symbol is also used for instance-level forgetting in Eq. (S3).","section":"Sec. S1.2.2, Eq. (S5)"},{"comment":"It would clarify what \"the overall complexity of Eq. (3) is also O(mM)\" means in terms of the group size F; the current sentence is easy to misread as saying the complexity is independent of M. Please state explicitly that the complexity is linear in M for fixed m, and define m in terms of the grouping window F.","section":"Sec. 3.2, Eq. (3)"},{"comment":"The ablation column header G_hat(V_n; E) is not defined in the main text; Eq. (11) describes gradient projection for U_n only. Please specify how projection is applied to V_n and to the aggregator E, or remove the undefined notation.","section":"Table 5 and Sec. 3.4"},{"comment":"The sentence \"ablating GDAT, leading to the results of ABMIL\" is imprecise: ABMIL is a separate architecture, not GDAT with the grouped attention removed. The supplementary comparison with \"Single Attn\" in Table S11 is the right kind of ablation; the main text should refer to it directly.","section":"Sec. 4.6"},{"comment":"The pseudo-label accuracy curve lacks axis labels and error bars and does not indicate which task it is averaged over. Please add these details and report the first-epoch value, which is needed to assess the starting point of the self-training loop.","section":"Fig. S1(a)"},{"comment":"The claimed \"up to 23.4% in localization accuracy\" is not easy to locate in the tables (Table 1 shows a 10.0-point IoU gap and Table 3 a 12.87-point gap). Please state the exact comparison underlying this claim or revise the wording to avoid ambiguity.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with extensive ablations and a clear presentation, but the central localization claim needs additional evidence beyond the current self-training setup. I recommend major revision rather than rejection. I am also concerned that the PAIP negative bags are synthetic and that hyperparameters may have been selected on the evaluation setup, which could inflate the reported margins. The authors should be asked to address the BPPL self-training concern, the PAIP data construction issue, and the validation protocol in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: CoMEL is a well-built continual MIL paper with unusually thorough benchmarking, but the headline localization gains depend on a self-training pseudo-labeling loop whose starting point is assumed rather than shown, so the IoU numbers are conditional on attention being decent already.\n\nThe genuinely new thing is the problem formulation: evaluating and optimizing instance-level localization under continual task sequences for WSIs, which the prior MIL-CL work (ConSlide, QPMIL-VL) mostly ignored. The three components are combinations of known ideas, but they are cleanly motivated. GDAT's grouped double attention is a sensible scalability fix for ~10^5 instances, OWLoRA's orthogonal low-rank bases with gradient projection is a reasonable way to do rehearsal-free continual learning with arbitrary feature extractors, and the ablations show each piece contributes. Credit where due: the experiments cover two task orders, multiple backbones, single-dataset and joint training, efficiency tables, and the reported gains over InfLoRA and ConSlide are consistent. The reversed-sequence results in the supplement are exactly the kind of robustness check you want to see.\n\nThe soft spot is where the stress-test points: BPPL. The pseudo-labels in Eq. (7)-(8) are built from bag prototypes derived from the same attention map A that the model is being trained to improve. If attention starts out bad, the positive prototype is contaminated, and the filter 1(Y_hat=Y) does nothing as long as the bag prediction is right. The paper's own Sec. S5 admits BPPL 'heavily relies on the accuracy of attention scores as pseudo-labels' and that performance is highly sensitive to hyperparameters. The supplementary pseudo-label accuracy curve doesn't show a no-BPPL control or initial accuracy, so it doesn't answer the bootstrapping question. On top of that, the PAIP 'normal' slides are made by cutting tumor regions out of tumor slides, so the negative distribution is synthetic. No code release either. These are real limitations, but notice they mainly threaten the localization claim; the bag-level accuracy and forgetting gains are less dependent on BPPL and are more likely to transfer.\n\nConclusion: this is a solid engineering contribution for computational pathology and continual learning. It deserves a serious referee. I would send it out with a request for code, a validation-based hyperparameter selection, and ideally an evaluation on real normal slides. It doesn't need to be desk-rejected; it needs independent verification of the mechanism.","headline":"Solid, thoroughly benchmarked continual MIL paper where the localization headline rests on self-training pseudo-labels that deserve independent verification before you trust the IoU gains.","tokens_in":27037,"tokens_out":2232,"would_cite":true,"duration_ms":26119,"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":"Continual multiple instance learning on whole-slide images can keep both bag-level accuracy and pixel-level tumor localization while a model adapts to a sequence of new organs and subtypes, and CoMEL claims to achieve this rehearsal-free.","keywords":["continual learning","multiple instance learning","whole slide image analysis","histopathology","instance localization","pseudo-labeling","low-rank adaptation","catastrophic forgetting"],"falsifier":"Two checks would settle the central claim. First, run CoMEL on a ten-task organ sequence and plot task-1 localization IoU after each new task: OWLoRA's orthogonal basis budget is finite, so if the mechanism is the reason forgetting stays low, forgetting should creep up as the subspace fills. Second, train the first task with a randomly frozen attention head and measure whether BPPL's pseudo-labels and bag-prototype losses still raise IoU over that poor-attention baseline; if IoU stays flat, the method's localization gains depend on initially good attention, exactly the assumption stated in Sec. 3.3.","tokens_in":25928,"feed_emoji":"🔬","tokens_out":10753,"duration_ms":109445,"temperature":0.7,"pith_summary":"This paper goes after an under-explored problem: multiple instance learning (MIL) on histopathological whole slide images is cheap in annotation (only slide-level labels) but is almost always trained once, whereas real hospitals accumulate slides of new organs and tumor subtypes continuously. The authors propose CoMEL, a framework for continual MIL that must keep predicting the right slide class and the right tumor location after every new task. It does so with a grouped double-attention transformer for scaling to roughly $10^{5}$ patches per slide, bag-prototype pseudo-labeling that turns attention into instance-level training signal, and an orthogonal low-rank adaptation that adds per-task subspaces without overwriting earlier ones. On a five-organ sequence (CAMELYON-16 plus PAIP), CoMEL reports the best instance accuracy, IoU, Dice, and bag accuracy among all baselines, and on four TCGA tumor-subtype tasks it reports the lowest bag-level forgetting; if these results hold, continual re-training on hospital data becomes a realistic way to keep pathology models current without storing patient patches.","feed_headline":"Continual MIL lifts tumor localization up to 23%","feed_subtitle":"CoMEL learns new organs and subtypes while preserving slide labels and pixel-level tumor maps — no rehearsal memory.","key_machinery":"The argument is carried by two mechanisms working together. Bag Prototypes-based Pseudo-Labeling (BPPL) converts the MIL attention vector $A$ into a class-probability vector by min-max normalization, builds a positive prototype $x_{\\mathrm{bag}} = \\sum_m a_m x_m$ and a negative prototype from the reversed attention $\\tilde{a}_m = (1-\\alpha_m)/\\sum_m (1-\\alpha_m)$, and pseudo-labels instances by their residual similarity to the two prototypes, discarding labels whenever the bag prediction is wrong or confidence is below the adaptive threshold $\\tau_1$. Orthogonal Weighted Low-Rank Adaptation (OWLoRA) represents each task's weights as a sum of low-rank bases with orthonormal columns, imposing intra-orthogonality through a penalty $L_{\\mathrm{in}} = \\|U_n^T U_n - I_d\\|_F^2$ and inter-orthogonality through gradient projection $\\hat{G}(U_n) = (I_D - \\sum_{k<n} U_k U_k^T)\\, G(U_n)$; because the first task's principal components are frozen, later tasks literally cannot move the subspace the first task occupies. Grouped Double Attention Transformer (GDAT) makes these trainable and scalable: it rewrites the $O(M^2)$ attention as two $O(mM)$ attentions over average-pooled region tokens, then adds the raw values back with weight $\\eta$ so grouped attention does not flatten instance diversity.","core_discovery":"The central claim is that under a continual MIL setup, a model can keep learning new whole slide image tasks (new organs, or new tumor subtypes within an organ) while preserving both the slide-level classification and the instance-level localization that existing continual MIL methods mostly neglect. CoMEL's first task trains full weights, extracts the top principal components by singular value decomposition, and then freezes them; every subsequent task adds low-rank bases $U_k S_k (V_k)^T$ that are forced to be orthonormal and whose gradients are projected onto the orthogonal complement of all earlier bases, so new knowledge cannot displace old. Localization is driven by turning attention scores into per-instance class probabilities, forming positive and negative bag prototypes, and pseudo-labeling instances whose confidence passes an adaptive threshold gated on the bag prediction. Against the prior LoRA-based state of the art (InfLoRA) on the combined organ sequence, CoMEL reports ACCinst 72.64 versus 70.17, IoU 41.87 versus 31.87, Dice 51.44 versus 40.84, and ACCbag 62.96 versus 56.93; on the TCGA subtype sequence it reports ACCbag 87.94, Forgetbag 10.12, and M.ACCbag 92.28, the best of all compared methods.","pith_inferences":["Making BPPL's pseudo-labels conditional on the model's own attention (the paper explicitly assumes the MIL model can successfully localize positive instances in Sec. 3.3) implies the method's gains should shrink for backbones whose first-task attention is poor; measuring per-task pseudo-label accuracy against attention quality would locate where the real dependency on pre-trained features sits.","OWLoRA's orthogonal subspace has a finite rank budget (the paper concedes this in Sec. S5), so on open-ended sequences of ten or more tasks the forgetting curve should eventually rise; a long-sequence experiment would settle whether the mechanism scales beyond the five-task setups tested here.","The positive-versus-reversed-attention prototype contrast is a general mechanism: the same idea could be dropped into other weakly supervised continual settings, such as incremental semantic segmentation from image-level labels, where pseudo-label noise similarly accumulates across tasks.","The paper leaves blurry tasks (mixed organs within a single task, which is what hospital data streams actually look like) as future work; whether BPPL's adaptive threshold and OWLoRA's task-specific bases survive ambiguous task boundaries is an open empirical question."],"forward_implications":["Rehearsal-free continual MIL can beat rehearsal-based methods on bag accuracy while storing no patient patches, addressing the privacy concern that limits memory-based approaches like ER and ConSlide in hospitals.","Instance-level forgetting under continual MIL is substantially reducible: CoMEL reports Forgetinst 14.00 on the five-organ sequence versus 17.89 for InfLoRA and roughly 19-27 for the other baselines, with the pattern reproduced under reversed task order in Table S13.","The localization machinery also helps outside the continual setup: on the merged five-organ data, GDAT+BPPL raises instance accuracy to 79.27, IoU to 49.32, and Dice to 60.68, versus 68.24, 36.45, and 49.33 for the strongest localization baseline smAP.","CoMEL is backbone-agnostic: with ResNet50, PLIP, and CONCH as frozen feature extractors it still beats the continual baselines on Dice, showing the gains do not depend on the UNI foundation model.","On TCGA tumor subtyping, continual bag classification reaches ACCbag 87.94 and Forgetbag 10.12, leaving a small gap to the joint-training upper bound of ACCbag 91.49, which suggests the continual setting can approach the non-continual ceiling."],"supporting_citations":[{"why":"Attention-based MIL aggregator that defines how attention A is produced and used as the localization signal, which CoMEL builds on.","marker":"[24]"},{"why":"ConSlide, the rehearsal-based continual MIL baseline whose TCGA task sequence and masked accuracy metric CoMEL adopts and must beat without rehearsal.","marker":"[22]"},{"why":"QPMIL-VL, the strongest rehearsal-free continual MIL baseline, which OWLoRA is designed to rival without a vision-language model or stored data.","marker":"[17]"},{"why":"InfLoRA, the interference-free LoRA method CoMEL extends, and the closest prior baseline on both bag and instance metrics.","marker":"[39]"},{"why":"RRT-MIL, the feature re-embedding MIL whose region grouping into tokens CoMEL follows to make the double attention computationally feasible.","marker":"[58]"},{"why":"smAP, the MIL method with the strongest prior localization results, which serves as the direct comparison for BPPL's instance metrics.","marker":"[7]"},{"why":"LoRA, the original low-rank adaptation scheme whose subspace decomposition OWLoRA generalizes with orthogonality constraints.","marker":"[21]"},{"why":"UNI, the pathology foundation model used as the frozen feature extractor in all main experiments.","marker":"[13]"},{"why":"TCGA, the dataset that supplies the four-organ tumor subtyping sequence for the continual bag classification experiments.","marker":"[67]"}],"fun_headline_variants":["CoMEL: continual WSI learning without forgetting localization","New continual MIL preserves tumor maps while learning new tasks","CoMEL boosts WSI localization by 23% in continual MIL","Continual MIL for WSI: keep slide labels, keep tumor locations","Minimal-forgetting MIL learns new organs without losing tumor pixels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the MIL model's attention already points at positive instances reasonably well: BPPL builds its prototypes and pseudo-labels from that attention, so if attention is poor at the start, the model trains on its own noisy labels and the claimed localization gains would not appear.","fun_headline_variants_meta":{"raw":{"variants":["CoMEL: continual WSI learning without forgetting localization","New continual MIL preserves tumor maps while learning new tasks","CoMEL boosts WSI localization by 23% in continual MIL","Continual MIL for WSI: keep slide labels, keep tumor locations","Minimal-forgetting MIL learns new organs without losing tumor pixels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000258,"raw_usage":{"total_tokens":1656,"prompt_tokens":1091,"completion_tokens":565,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":707,"completion_tokens_details":{"reasoning_tokens":478}},"tokens_in":707,"tokens_out":565,"duration_ms":5967,"temperature":1.0,"reasoning_tokens":478,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:30:36.501005+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Two checks would settle the central claim. First, run CoMEL on a ten-task organ sequence and plot task-1 localization IoU after each new task: OWLoRA's orthogonal basis budget is finite, so if the mechanism is the reason forgetting stays low, forgetting should creep up as the subspace fills. Second, train the first task with a randomly frozen attention head and measure whether BPPL's pseudo-labels and bag-prototype losses still raise IoU over that poor-attention baseline; if IoU stays flat, the method's localization gains depend on initially good attention, exactly the assumption stated in Sec. 3.3.","supporting_citations":[{"cited_title":"Attention-based deep multiple instance learning","cited_arxiv_id":null,"evidence_quote":"Attention-based MIL aggregator that defines how attention A is produced and used as the localization signal, which CoMEL builds on."},{"cited_title":"Conslide: Asynchronous hier- archical interaction transformer with breakup-reorganize re- hearsal for continual whole slide image analysis","cited_arxiv_id":null,"evidence_quote":"ConSlide, the rehearsal-based continual MIL baseline whose TCGA task sequence and masked accuracy metric CoMEL adopts and must beat without rehearsal."},{"cited_title":"Queryable Prototype Multiple Instance Learning with Vision-Language Models for Incremental Whole Slide Image Classification","cited_arxiv_id":"2410.10573","evidence_quote":"QPMIL-VL, the strongest rehearsal-free continual MIL baseline, which OWLoRA is designed to rival without a vision-language model or stored data."},{"cited_title":"Inflora: Interference-free low-rank adaptation for continual learning","cited_arxiv_id":null,"evidence_quote":"InfLoRA, the interference-free LoRA method CoMEL extends, and the closest prior baseline on both bag and instance metrics."},{"cited_title":"Feature re-embedding: Towards foun- dation model-level performance in computational pathology","cited_arxiv_id":null,"evidence_quote":"RRT-MIL, the feature re-embedding MIL whose region grouping into tokens CoMEL follows to make the double attention computationally feasible."},{"cited_title":"Sm: en- hanced localization in multiple instance learning for medical imaging classification","cited_arxiv_id":null,"evidence_quote":"smAP, the MIL method with the strongest prior localization results, which serves as the direct comparison for BPPL's instance metrics."},{"cited_title":"Lora: Low-rank adaptation of large language models","cited_arxiv_id":null,"evidence_quote":"LoRA, the original low-rank adaptation scheme whose subspace decomposition OWLoRA generalizes with orthogonality constraints."},{"cited_title":"The cancer genome atlas pan-cancer analysis project","cited_arxiv_id":null,"evidence_quote":"TCGA, the dataset that supplies the four-organ tumor subtyping sequence for the continual bag classification experiments."}],"review_version":1}