{"id":"2393ee9c-8d24-4286-a93b-43b4bf1edd64","arxiv_id":"2411.16064","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GROTO is a source-free class-incremental domain adaptation method that uses multi-granularity class prototypes and prototype topology distillation, achieving state-of-the-art final accuracy on three benchmark variants.","lead":"This paper presents GROTO, a method for adapting a pretrained source model to unlabeled target data that arrive in batches of new classes, without accessing source images. It combines prototype-based pseudo-labeling with topology distillation to reduce forgetting, and reports state-of-the-art accuracy on three domain adaptation benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA claim rests on single-run comparisons with test-set hyperparameter selection; small margins need variance checks.","rationale":"I read the paper as making a first-solution claim for CI-SFUDA with a plausible mechanism. The positive-class mining threshold is a genuine algorithmic risk, and the reader is right that Eq. 1-2 can fail on imbalanced or overconfident batches. However, the supplementary TCD/PCD tables expose this failure mode and the method still performs well, so that weakness alone is not the decisive threat to the headline. The decisive threat is evaluation: the main tables are single-run, and Table 5 reports hyperparameter ablations on the same test benchmark used for the SOTA claim. The margins over ProCA-B and LEAD-B are 1-4 points, which is within typical seed variance for a ViT-B backbone. The provided code makes a seed/variance check feasible. If the check passes, the central claim is supported; if not, the paper should remain conditional on additional statistical evidence. My recommendation is unchanged: CONDITIONAL, pending the variance and validation-based selection check.","tokens_in":26279,"tokens_out":10500,"duration_ms":105771,"concrete_test":"Run the released GROTO code on all three benchmarks with fixed hyperparameters (γ=10, μ0_c=0.5, β=1e-4, nr=10) for 10 random seeds, and re-run the strongest baselines under the same backbone and protocol. Compute mean±std final accuracy and paired differences. Also select hyperparameters on a held-out validation split instead of the test tables. If for any benchmark GROTO's mean minus 2σ does not exceed the best baseline's mean, the SOTA claim is not established at current evidence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: GROTO is state-of-the-art on three CI-SFUDA benchmarks. What has to be true is that the reported margins over strong baselines are reproducible and not artifacts of test-set selection. The paper reports one run per cell, no error bars, and Table 5 selects default hyperparameters (μ0_c=0.5, β=1e-4, nr=10, γ=10) using the same ImageNet-Caltech-CI test accuracy that the SOTA claim is based on. The margins are small: +1.0 percentage points over ProCA-B on ImageNet-Caltech-CI, +1.3 on Office-Home-CI average, and the Office-31 gain is dominated by a few sub-tasks. Without variance or a validation-based selection protocol, a single-seed difference of about one point is not evidence of SOTA. This concern is directly testable because the code is provided.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GROTO, a method for Class-Incremental Source-Free Unsupervised Domain Adaptation (CI-SFUDA). The target data arrive in sequential sessions with disjoint class subsets of the source label space, and no source data are available during adaptation. GROTO consists of two main modules: (1) multi-granularity class prototype self-organization, which mines positive classes via source-similarity and target-probability accumulation distributions, constructs coarse- and fine-grained prototypes, and trains with cross-entropy and contrastive losses; and (2) prototype topology distillation, which aligns target classifier prototypes to source prototypes through compactness and separability losses to mitigate forgetting. The paper reports state-of-the-art final and session accuracies on Office-31-CI, Office-Home-CI, ImageNet-Caltech-CI, and adds DomainNet-126-CI results in the supplement. Code is released.","tokens_in":26482,"tokens_out":6812,"duration_ms":62371,"significance":"If the reported results are reproducible, GROTO would be a strong baseline for CI-SFUDA, a setting that combines source-free adaptation, class-incremental shifts, and unlabeled target streams. The paper contributes a new problem instantiation and a modular pipeline, with extensive benchmark comparisons, ablations, mining-accuracy tables, and complexity analyses. The release of code is a concrete practical asset. The significance is currently tempered by missing implementation details in the topology distillation loss and by the lack of statistical evidence for the claimed state-of-the-art margins.","major_comments":[{"comment":"The target proportion p(f_i) is never defined. The text states only that p(·) is the target proportion, and Algorithm 4 lists it as an input, but no estimation procedure, no source (e.g., normalized cumulative probabilities from Eq. (2)), and no uniformity assumption are provided. Because p(f_i) weights every term in L_com and L_sep, the loss is not implementable as written. Please specify p(·) explicitly or, if the method is intended to be uniform, state that directly.","section":"Sec. 3.4, Eqs. (14)-(15)"},{"comment":"The construction of the source and target prototypes used in L_ptd is incomplete. The source classifier weights are introduced as μ=[μ_1,...,μ_N] for N positive classes, but the paper does not state how the N positive-class columns are selected from the K-class source classifier, nor how the target classifier weights f=[f_1,...,f_N] are initialized at session 1 and expanded when new sessions add classes. This matters because L_ptd compares f against μ; without this specification, the update of the target classifier head is ambiguous.","section":"Sec. 3.4 and Algorithm 4"},{"comment":"The state-of-the-art claim rests on single-run comparisons with small margins (e.g., +1.0 percentage points on ImageNet-Caltech-CI and +1.3 on Office-Home-CI average), and the default hyperparameters in Table 5 (μ0_c=0.5, β=1e-4, nr=10) are selected using the same ImageNet-Caltech-CI average final accuracy that is reported in Table 1. No error bars or multiple seeds are reported. Please report mean±std over at least three seeds for the main comparisons, and either select hyperparameters on a validation split or state clearly that the reported numbers themselves involve test-set selection.","section":"Sec. 4.2 and Table 5"},{"comment":"The HKPCM threshold compares each class's source-similarity and cumulative probability to the mean over all K source classes. The supplementary tables show that false positives are not negligible (e.g., Table 7, I→C sessions 3, 4, 6, 7, and 8 have GROTO TCD Acc. of 90.9, 75.0, 64.3, 66.7, and 71.4 versus ProCA-B 100.0, 83.3, 69.2, 83.3, and 81.8). Since every downstream component consumes the mined positive classes, the paper should analyze when the mean-threshold assumption can fail (e.g., imbalanced target batches, overconfident source models, or classes with very different priors) and, at minimum, report the sensitivity of final accuracy to mining errors.","section":"Sec. 3.3, Eqs. (1)-(2), and Supplementary Tables 7-9"}],"minor_comments":[{"comment":"The notation P_k = Σ_i C_s(G_s(x_i)) is ambiguous: C_s(G_s(x_i)) is a vector, and the equation should refer to the class-k component C_s^k(G_s(x_i)).","section":"Sec. 3.3, Eq. (2)"},{"comment":"There is an inconsistency in the contrastive loss denominator: the main text uses exp(ϕ(G_t(x_i), G_t(x_b))/κ), while Supplementary Algorithm 3 line 43 writes exp(ϕ(G_t(x_i), G_t(x'_i))/κ). Please unify the formula.","section":"Eq. (11) and Algorithm 3"},{"comment":"Office-31-CI is described as having 31 shared classes, but the described split covers 30 classes (three disjoint subsets of 10). Clarify whether one source class is intentionally held out as a permanent negative class, and similarly clarify the number of unused source classes in Office-Home-CI and ImageNet-Caltech-CI.","section":"Sec. 4.1"},{"comment":"The sentence 'With the pseudo label assigned in subsequent stages, unique pseudo-labels evolve as the positive classes' is unclear and should be rewritten to state how pseudo-label sets are constructed and expanded across sessions.","section":"Sec. 3.3"},{"comment":"Reference [35] is incomplete (missing venue and full title), and the captions of Figures 3 and 4 should define the plotted quantities, particularly 'average accuracy of 0 to 9 classes' and the histogram values shown in Figure 4.","section":"References and figure captions"}],"recommendation":"major_revision","confidential_remarks":"The paper is closest to ProCA-B, and the reported gains over that baseline are modest; the missing p(f_i) definition and the unspecified classifier expansion are easy to fix and should be required. The supplementary mining-accuracy tables are useful and should be kept. I would also ask the authors to report multi-seed statistics before any acceptance decision, given the one-point margins."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this is a real new problem formulation—CI-SFUDA—and the method is coherent, but the SOTA claim is built on small margins and the evaluation has two fixable problems. If those get fixed, this is a solid paper; right now I'd read the results as promising but not yet proven.\n\nWhat's new: the setup (source-free, class-incremental, unsupervised) is not covered by ProCA (needs source data), CIDA (not source-free), or SF-UniDA (all target data available). The paper makes a legitimate contribution by defining this scenario and proposing a specific pipeline: hybrid positive class mining, multi-granularity prototypes for pseudo-labeling, and prototype topology distillation against forgetting. The writing is clear, the ablations are useful, and code is provided. The supplementary adds DomainNet results and per-session old-class accuracy, which helps.\n\nThe soft spots are real but addressable. First, the target proportion p(f_i) in Eq. 14–15 is never defined; Algorithm 4 takes it as input but the main text doesn't say how it's estimated. Second, the expansion of the target classifier weights across sessions (how new class weights are initialized and added) is unspecified. That's a reproducibility gap. Third, Table 5 selects hyperparameters (μ0_c, β, nr) using the same ImageNet-Caltech-CI test accuracy that supports the SOTA claim, and there are no error bars. The margins over ProCA-B are around one point on that benchmark and on Office-Home; without variance or a validation-based selection protocol, a single-run one-point gap is not strong evidence. The positive class mining threshold (above-average similarity/probability) is a heuristic that could fail on imbalanced streams, but the supplementary PCD/TCD tables suggest it behaves reasonably on these benchmarks, so I'd weight that as a secondary concern.\n\nOverall: the problem formulation alone justifies a serious look. The method is plausible, the empirical work is extensive, and the code is a plus. But the paper needs a revision that defines p(f_i), specifies classifier weight handling, and adds multiple seeds or a validation-based hyperparameter protocol. I'd send it to review, not desk-reject, and I'd want a referee to push on those points.","headline":"A genuine new problem setup and a plausible method, but the SOTA claim needs variance estimates and a defined p(f_i) before it convinces.","tokens_in":27012,"tokens_out":2837,"would_cite":true,"duration_ms":26130,"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":"GROTO claims a single algorithm handles class-incremental source-free domain adaptation by mining present classes and distilling source topology, reporting state-of-the-art accuracy on three benchmarks.","keywords":["class-incremental learning","source-free unsupervised domain adaptation","prototype learning","topology distillation","pseudo-labeling","catastrophic forgetting","contrastive learning"],"falsifier":"Build a synthetic target session with ten classes where one class has only a few samples, such as five percent of the batch, while another absent source class receives high confidence from the frozen source model; run the positive-class mining and measure its precision and recall against the ground-truth present classes, and if the rare class is missed or the absent class is flagged as positive, the above-mean accumulation rule has failed and the downstream pseudo-labels, prototypes, and topology distillation inherit the error.","tokens_in":26097,"feed_emoji":"🎯","tokens_out":10683,"duration_ms":89002,"temperature":0.7,"pith_summary":"The paper tackles class-incremental source-free unsupervised domain adaptation (CI-SFUDA): a model trained on labeled source data arrives at a target domain where unlabeled images of new class subsets appear session by session, with no access to source images or previously seen target images. Its claim is that a single algorithm, GROTO, can keep learning these increments without collapsing old knowledge, by first deciding which source classes are actually present in each session, then pulling target features of the same class together, and finally distilling the source feature topology into the target classifier. This matters because it would let a deployed model be updated on new categories in a privacy-preserving, streaming way instead of being retrained from scratch. The paper reports state-of-the-art final and per-session accuracy on Office-31-CI, Office-Home-CI, and ImageNet-Caltech-CI, with ablations attributing gains to each module.","feed_headline":"GROTO beats prior methods on 3 incremental adaptation benchmarks","feed_subtitle":"Mining each session's real classes and distilling source topology curbs old-class drift without source data.","key_machinery":"Positive classes are first mined by modeling the per-class accumulation distributions of source similarity and target prediction probability, keeping only classes whose values exceed the mean in both signals. The load-bearing objects are multi-granularity class prototypes and the topological structures formed from them. Coarse-grained prototypes are the source classifier weight vectors plus target features closer than the mean cosine-distance threshold to their class centroid, while fine-grained prototypes are target features whose average prediction confidence is above the batch threshold and whose augmented-view confidence has low standard deviation; each unlabeled sample is pseudo-labeled by the positive class with the smallest average cosine distance to its prototypes. The topology is the complete set of pairwise cosine distances between the $N$ source classifier weights $\\mu$ and the $N$ target classifier weights $f$. The prototype topology distillation loss $L_{\\mathrm{ptd}} = L_{\\mathrm{com}} + L_{\\mathrm{sep}}$ weights each point-to-point cosine distance by a soft assignment, with target class proportions $p(f_i)$, so that every target prototype is drawn toward the source prototype it matches while source prototypes are not collapsed onto a single target prototype. The self-organization objective $L_{\\mathrm{ptfs}} = L_{\\mathrm{ce}} + L_{\\mathrm{con}}$ uses these pseudo-labels and supervised contrastive learning to make same-class target features cluster and different-class features separate.","core_discovery":"On the paper's own terms, the central discovery is that the two difficulties of CI-SFUDA can be attacked at the prototype level rather than at the image level. Positive classes, meaning the source classes present in a given target session, are mined by combining two accumulation statistics: the per-class average softmax-normalized similarity between target features and stored source centroids, and the per-class cumulative prediction probability of the frozen source model; a class is positive only if both exceed the mean over all source classes. Reliable pseudo-labels are then generated from multi-granularity prototypes, coarse-grained ones being source classifier weights plus confident target features near class centroids, and fine-grained ones being augmented-consistent confident features, and a cross-entropy plus supervised-contrastive loss organizes the target features. To prevent forgetting, the source and target classifier-weight vectors of the positive classes are treated as point sets, and two distillation losses, a compactness loss and a separability loss, pull each target prototype toward its corresponding source prototype while keeping the target prototypes separable from one another; a small exemplar memory provides replay. With a ViT-B backbone, this yields average final accuracies of 93.0 on Office-31-CI, 82.5 on Office-Home-CI, and 88.8 on ImageNet-Caltech-CI, which the paper reports as state-of-the-art.","pith_inferences":["Beyond the paper, the above-mean mining rule is the natural failure point to stress-test: under a heavily imbalanced target session or a poorly calibrated source model, a rare present class could fall below the mean and be dropped, while an absent high-confidence class could be admitted.","Beyond the paper, the same two-signal mining and prototype topology distillation could be extended to open-set or novel-class arrivals, where the target label space is not a subset of the source label space, by replacing the mean threshold with a learned rejection boundary.","Beyond the paper, because the topology distillation acts only on the $N \\times N$ prototype distances, the mechanism could be attached to large foundation-model backbones with lightweight classifier heads without storing source images.","Beyond the paper, one could test cutting the contrastive term in later sessions; the ablations show it helps the full pipeline, but removing it after the first session might reduce the reported training overhead."],"forward_implications":["A model can be adapted to new target classes in sequence without labeled source data and without replaying old target images, keeping only a small exemplar memory of ten examples per class.","Old-class accuracy remains relatively stable across sessions because the classifier is anchored to the source topology instead of drifting toward the latest batch.","The method outperforms both class-incremental domain adaptation methods that need source data and source-free universal methods that need all target data available at once.","The results across Office-31-CI, Office-Home-CI, ImageNet-Caltech-CI, and the supplementary DomainNet-126-CI suggest the approach scales to larger label spaces.","Ablations show that removing the positive-class mining, the self-organization, the topology distillation, or the contrastive loss each reduces final accuracy, indicating every module contributes to the reported gains."],"supporting_citations":[{"why":"Defines the class-incremental unsupervised domain adaptation setting and dataset protocol that GROTO follows, and supplies the exemplar-replay baseline.","marker":"[16]"},{"why":"Establishes the class-incremental domain adaptation baseline with proxy-source regularization that GROTO outperforms under the source-free constraint.","marker":"[13]"},{"why":"Is the source-free universal domain adaptation method whose feature-decomposition approach GROTO compares against.","marker":"[28]"},{"why":"Is the frozen-foundation-model source-free universal baseline used in comparisons and ablations.","marker":"[35]"},{"why":"Is the unified source-free domain adaptation baseline handling label-space mismatch, used as a comparison.","marker":"[34]"},{"why":"Provides the supervised contrastive loss used in the target feature self-organization loss.","marker":"[12]"},{"why":"Supports the premise that target prototypes move toward same-label source prototypes, motivating the compactness loss.","marker":"[36]"},{"why":"Supplies the exemplar-replay formulation used in the memory bank and replay loss.","marker":"[41]"},{"why":"Provides the ViT-B backbone used by GROTO and the compared baselines.","marker":"[4]"}],"fun_headline_variants":["GROTO: Prototype topology distillation tames incremental adaptation","No source data? GROTO still adapts incrementally via prototype topology","Mining positive classes and distilling topology eases class-incremental adaptation","Prototype-level attack on class-incremental source-free domain adaptation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that, in every target session, the classes actually present can be identified simply as those whose average similarity to source centroids and cumulative target prediction probability both exceed the mean across all source classes; if that fails for a class, its data receives no reliable pseudo-labels and the whole adaptation chain suffers.","fun_headline_variants_meta":{"raw":{"variants":["GROTO: Prototype topology distillation tames incremental adaptation","No source data? GROTO still adapts incrementally via prototype topology","Mining positive classes and distilling topology eases class-incremental adaptation","Prototype-level attack on class-incremental source-free domain adaptation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000724,"raw_usage":{"total_tokens":3303,"prompt_tokens":1056,"completion_tokens":2247,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":672,"completion_tokens_details":{"reasoning_tokens":2171}},"tokens_in":672,"tokens_out":2247,"duration_ms":15611,"temperature":1.0,"reasoning_tokens":2171,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:34:42.826720+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build a synthetic target session with ten classes where one class has only a few samples, such as five percent of the batch, while another absent source class receives high confidence from the frozen source model; run the positive-class mining and measure its precision and recall against the ground-truth present classes, and if the rare class is missed or the absent class is flagged as positive, the above-mean accumulation rule has failed and the downstream pseudo-labels, prototypes, and topology distillation inherit the error.","supporting_citations":[{"cited_title":"Prototype- guided continual adaptation for class-incremental unsuper- vised domain adaptation","cited_arxiv_id":null,"evidence_quote":"Defines the class-incremental unsupervised domain adaptation setting and dataset protocol that GROTO follows, and supplies the exemplar-replay baseline."},{"cited_title":"Class- incremental domain adaptation","cited_arxiv_id":null,"evidence_quote":"Establishes the class-incremental domain adaptation baseline with proxy-source regularization that GROTO outperforms under the source-free constraint."},{"cited_title":"Lead: Learn- ing decomposition for source-free universal domain adapta- tion","cited_arxiv_id":null,"evidence_quote":"Is the source-free universal domain adaptation method whose feature-decomposition approach GROTO compares against."},{"cited_title":"Source- free domain adaptation with frozen multimodal foundation model, 2024","cited_arxiv_id":null,"evidence_quote":"Is the frozen-foundation-model source-free universal baseline used in comparisons and ablations."},{"cited_title":"Supervised contrastive learning","cited_arxiv_id":null,"evidence_quote":"Provides the supervised contrastive loss used in the target feature self-organization loss."},{"cited_title":"A prototype-oriented framework for unsupervised domain adaptation","cited_arxiv_id":null,"evidence_quote":"Supports the premise that target prototypes move toward same-label source prototypes, motivating the compactness loss."},{"cited_title":"Large scale incre- mental learning","cited_arxiv_id":null,"evidence_quote":"Supplies the exemplar-replay formulation used in the memory bank and replay loss."}],"review_version":1}