{"id":"8da2a1e5-66c2-4335-8b34-a135db4a131e","arxiv_id":"2507.23237","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"Under a new unlabeled-set composition that mixes base and novel classes, the ALDC method reports average accuracy of 69.77% on mini-ImageNet, 10.9 to 11.5 points above prior Semi-FSCIL baselines.","lead":"This paper proposes a new semi-supervised few-shot class-incremental learning setting where unlabeled data contains both old and new classes, and a method, ALDC, that uses ambiguity-based pseudo-labeling and distribution calibration. The authors report state-of-the-art results on three benchmarks, but the comparison may be unfair because their base-session accuracy is much higher than all baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ALDC's reported SOTA gains are almost entirely the session-0 gap: it starts at 82.45% accuracy while all baselines start near 72.35%, and no matched base-session protocol is described.","rationale":"The reader's weakest_assumption correctly identifies the most load-bearing flaw: the claimed SOTA improvement is confounded by an unmatched base-session model. The numbers in Table I make this concrete. ALDC's session-0 accuracy is 82.45 versus 72.35 for the strongest prior baselines (UaD-CE and Us-KD). Since GSemi-FSCIL incremental sessions only provide novel-class few-shot samples and unlabeled data, session-0 accuracy is a pure measure of base-session training; ALT and B2N cannot influence it. The average accuracy gain over UaD-CE* (11.51 points) is almost exactly the session-0 gap (10.10 points), so the proposed modules account for at most ~1.4 points of average improvement, and even that could be due to the different base model's feature space. The paper provides no implementation details for base training (no backbone, epochs, optimizer, learning rate, augmentation) and no matched-protocol comparison. Therefore the central SOTA claim is not supported by the evidence as presented. This is an external-validity problem rather than an internal contradiction, but it is decisive for the paper's main assertion. I agree with the reader's verdict; the additional concerns about Eq. (4) and test-set hyperparameter selection are secondary, but they reinforce the rejection.","tokens_in":10401,"tokens_out":5217,"duration_ms":53719,"concrete_test":"Re-run ALDC's incremental sessions using the exact base-session protocol of UaD-CE (e.g., ResNet-18, 90 epochs, SGD with cosine decay, standard augmentation) and report session-0 accuracy. If session-0 accuracy drops to ~72.35%, then the 11.51% average gain over UaD-CE* in Table I is inherited from base-session initialization, not from ALT/B2N. As a complementary check, run UaD-CE* with ALDC's base-session recipe; if it also reaches ~82% at session 0, the proposed modules are not the source of the improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: ALDC sets new SOTA in GSemi-FSCIL. In Table I, ALDC reports 82.45% session-0 accuracy, while every baseline (FSCIL, Semi-FSCIL, and GSemi-FSCIL) begins between 62.98% and 72.35%. The session-0 gap to the strongest Semi-FSCIL baseline, UaD-CE, is 10.10 points; the reported average improvement over UaD-CE* is 11.51 points. Thus almost all of the claimed advantage is present before any incremental learning happens. The paper never specifies the base-session training protocol for ALDC (backbone choice, epochs, optimizer, learning rate, augmentation, batch size); the only mention is 'Train backbone, classifier by D1' in Algorithm 1. Without a matched base-session recipe, the SOTA comparison is confounded: the gains could come entirely from a stronger base model, and the ALT and B2N modules could be contributing little or nothing. This is not an internal inconsistency, but it invalidates the central comparative claim as presented. The ablation tables cannot rescue the comparison, because the ablations are all built on the same 82.45% base model; they show only the marginal effect of ALT/B2N on top of that base model, not that ALDC outperforms prior methods under equal conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a new setting, Generalized Semi-FSCIL (GSemi-FSCIL), in which the unlabeled pool in each incremental session contains samples from base and previously seen classes in addition to current novel classes. To handle this, the authors propose ALDC, consisting of an Ambiguity-guided Learnable Threshold (ALT) that pseudo-labels high-confidence unlabeled samples, and a Base-to-Novel (B2N) module that calibrates novel-class feature distributions using base-class statistics. The paper claims state-of-the-art results, with average accuracy gains of 10.93% over Semi-FSCIL and 11.51% over GSemi-FSCIL on mini-ImageNet, and presents curves for CIFAR-100 and CUB-200.","tokens_in":10751,"tokens_out":6500,"duration_ms":69037,"significance":"The proposed GSemi-FSCIL setting is a reasonable step toward more realistic semi-supervised FSCIL, and the observation that existing Semi-FSCIL methods degrade when the unlabeled set is generalized is potentially useful. If the reported gains survive a matched experimental protocol, ALDC would be a strong baseline for this setting. However, the paper's central SOTA claim is currently not established: the session-0 accuracy gap and the lack of a shared base-session training recipe dominate the reported average improvements, and the core B2N equation has formal problems. The paper does not include reproducible code or a parameter-free derivation, so the empirical comparison is the sole basis for the claims.","major_comments":[{"comment":"The central SOTA claim is confounded by unequal session-0 conditions. ALDC reports 82.45% accuracy at session 0, while every baseline starts at 72.35% or lower; the reported average improvements of 10.93% and 11.51% are nearly the entire session-0 gap (10.10 points over UaD-CE). No base-session training protocol is given: Algorithm 1 only says 'Train backbone, classifier by D1', with no backbone, epochs, optimizer, learning rate, or augmentation details. Because the ablations in Tables II and III all build on the same 82.45% base model, they show only the marginal effect of ALT/B2N on that base, not that ALDC outperforms prior methods under matched initial conditions.","section":"Table I / Section V-A"},{"comment":"The B2N calibration equation is dimensionally inconsistent as written. The numerator sums over i in D0 of class means (and covariances), but the denominator is N^t_c + 1, the per-class sample count plus one; if D0 consists of many base classes, the output is not a per-class mean or covariance estimate. The text says that only base classes estimated to be 'more similar' to the novel class should be used, but Eq. (4) pools all of D0. No derivation or normalization is supplied. Since B2N is one of the two core modules, this needs correction and validation.","section":"Eq. (4) / Section IV-B"},{"comment":"The sensitivity analysis selects m and alpha by reporting the last-session accuracy, which appears to be test-set accuracy, and the manuscript never states which values of m and alpha are used in the main comparisons of Table I. This is a test-set model-selection issue that can bias the reported gains. The main-table numbers should be tied to a pre-specified validation choice, or the full sensitivity surface should be reported for all sessions rather than only session 8.","section":"Fig. 3(d) / Section V-D"},{"comment":"The three-dataset claim is not fully verifiable: CIFAR-100 and CUB-200 results are presented only as curves, with no numerical tables, no dataset-specific protocols, and no GSemi-FSCIL baseline numbers for those datasets. To support the claim of state-of-the-art results on three benchmark datasets, the paper should provide session-by-session accuracy tables and the same matched-protocol description for all datasets.","section":"Fig. 3(a-b) / Section V-A"}],"minor_comments":[{"comment":"The definition D_t = {(x_i,y_i)}_{i=0}^{N times K} uses undefined notation, and the condition 't > 1' should be 't >= 1' for the first incremental session.","section":"Section III"},{"comment":"The symbol t is used both as the session index and as the threshold, which is confusing; the initial value of the threshold is never specified.","section":"Algorithm 1"},{"comment":"The smoothing coefficient m is not defined, and although a sensitivity grid is reported in Fig. 3(d), the manuscript does not state which m value was used for the main results.","section":"Eq. (2)"},{"comment":"The covariance formula sums over x_j in D_t, but it should sum over samples of class c; the meaning of N^t_c is not stated.","section":"Eq. (3)"},{"comment":"The 'Baseline' average accuracy is 66.16 in Table II but 66.27 in Table III for identical per-session numbers; this inconsistency should be corrected.","section":"Tables II and III"}],"recommendation":"major_revision","confidential_remarks":"The main barrier is not the novelty of the setting but the absent matched training protocol; without a rerun of all baselines under the same base-session initialization, the SOTA claim cannot stand. The Eq. (4) issue is likely fixable but must be corrected and re-evaluated. The GSemi-FSCIL setting alone is a modest contribution; the value of the paper rests on the corrected empirical study."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's main contribution is the new GSemi-FSCIL setting, which puts base and all previously seen novel classes into the unlabeled pool. That is a legitimate and realistic extension of Semi-FSCIL, and the paper shows that existing methods do degrade under it. The ALT and B2N modules are sensible heuristics, and the ablations give some evidence that each component helps. Credit where due: the problem framing is useful, and the empirical point about existing methods struggling with base-class unlabeled data is well taken.\n\nThe soft spots are serious, though. The central SOTA comparison is not properly controlled. ALDC starts at 82.45% session-0 accuracy while every baseline starts near 72.35%, and the paper never specifies the base-session training protocol — backbone, epochs, optimizer, augmentation, none of it. So the 10-point gap could come from a stronger base model, not from the incremental modules. The stress-test note claims this means 'almost all' of the gain is inherited from initialization. That arithmetic is actually wrong: averaging the per-session differences, the incremental sessions show a larger average gap (about 11.7 points) than session 0 (10.1 points). So the modules are not doing nothing. But the unmatched base-session recipe still means the comparison is confounded, and the claimed SOTA advantage is not demonstrated.\n\nEq. (4) is another real problem. Summing all base-class means and covariances and dividing by the number of novel samples is not a sensible calibration operation as written; the notation is ambiguous, and the formula looks dimensionally off unless it is a typo for averaging over base classes. The authors need to derive or clearly state what this operation is doing. Hyperparameters m and α are selected on the test benchmark, which is a model-selection issue. The choice of 50 unlabeled samples is arbitrary. There are also internal inconsistencies: the average accuracies in Table I do not match the per-session numbers, which suggests careless reporting.\n\nWho gets value from this? Researchers working on few-shot class-incremental learning under semi-supervised setups, especially those interested in realistic unlabeled data composition. The new benchmark could be useful to the community, and the paper deserves a serious referee. But as it stands, the empirical case is not convincing, and the method section needs clarification. I would recommend major revision: specify the base-session protocol, match it with baselines (or at least show results with a shared backbone), fix Eq. (4), and correct the table arithmetic. If those are addressed, the paper could be a solid contribution.","headline":"The new GSemi-FSCIL setting is a real contribution, but the SOTA claim is undermined by an unmatched base-session protocol and an unclear calibration formula.","tokens_in":11215,"tokens_out":4639,"would_cite":false,"duration_ms":47205,"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":"The paper redefines Semi-FSCIL so unlabeled data includes old and new classes, and proposes ALDC, a learnable threshold plus base-to-novel distribution calibration, which sets new state-of-the-art results on three benchmarks.","keywords":["few-shot class-incremental learning","semi-supervised learning","generalized semi-FSCIL","distribution calibration","pseudo-labeling","ambiguity-guided threshold","class-incremental learning"],"falsifier":"Retrain ALDC with the base-session protocol used by the baselines so that session-0 accuracy matches theirs, then recompute average accuracy over all sessions; if the 10.93% advantage shrinks or vanishes, the claimed improvement comes from initialization rather than from ALT and B2N.","tokens_in":1539,"feed_emoji":"🎯","tokens_out":1726,"duration_ms":73413,"temperature":0.7,"pith_summary":"This paper argues that existing Semi-FSCIL settings are unrealistic because the unlabeled pool contains only current novel classes. It proposes Generalized Semi-FSCIL (GSemi-FSCIL), where unlabeled data also includes base and previously seen novel classes, and claims prior methods degrade in this setting. To address the resulting base-versus-novel confusion, it introduces ALDC: an ambiguity-guided learnable threshold selects which unlabeled samples to trust, and a base-to-novel calibration corrects biased few-shot class distributions using abundant base-class statistics. On mini-ImageNet, CIFAR-100, and CUB200, the paper reports state-of-the-art average accuracy, outperforming the prior Semi-FSCIL state of the art by 10.93% on mini-ImageNet.","feed_headline":"New method beats prior few-shot incremental learning by ~11%","feed_subtitle":"A learnable threshold plus base-to-novel calibration exploits ambiguous unlabeled samples to set state-of-the-art on three benchmarks","key_machinery":"The machinery is the dual-branch Ambiguity-guided Learnable Threshold (ALT) combined with Base-to-Novel distribution calibration (B2N). ALT computes base and novel similarity scores $S_i^b$ and $S_i^n$ as maximum cosine similarities to the respective class weights, then sets a threshold $\\tau = \\frac{1}{N_u}\\sum_i |S_i^b - S_i^n| + m$ to decide whether an unlabeled sample gets a pseudo-label or is treated as high-ambiguity. B2N estimates per-class mean $\\mu_c$ and covariance $\\Sigma_c$ for novel classes, averages them with the statistics of the most similar base classes via $\\mu'_c = \\frac{\\sum_{i\\in D_0}\\mu_i + \\mu_c}{N_t^c+1}$ and $\\Sigma'_c = \\frac{\\sum_{i\\in D_0}\\Sigma_i + \\Sigma_c}{N_t^c+1} + \\alpha$, then samples synthetic features $x \\sim \\mathcal{N}(\\mu'_c, \\Sigma'_c)$ to augment the few-shot training set. This lets the classifier be retrained with an augmented set that the paper reports as converting an $N$-way $K$-shot task into an $N$-way $(K+10)$-shot task.","core_discovery":"The central claim is that the main obstacle in generalized semi-supervised few-shot incremental learning is the presence of high-ambiguity unlabeled samples that resemble both base and novel classes, and that these samples can be exploited rather than discarded. The method computes per-sample similarity to base and novel class weights, uses a learnable threshold to separate confident samples from ambiguous ones, and for ambiguous samples calibrates the novel-class Gaussian feature distribution with base-class mean and covariance statistics, generating synthetic training features. The paper reports that this dual mechanism both protects base-class performance and improves novel-class accuracy, yielding new state-of-the-art results on three benchmarks.","pith_inferences":["The GSemi-FSCIL protocol is a more faithful model of real deployments, where unlabeled streams mix old and new classes; if adopted as a standard evaluation, it could shift how semi-supervised incremental methods are compared.","Because B2N transfers covariance and mean from data-rich base classes to data-poor novel classes, the same mechanism is a natural fit for other label-scarce incremental settings where a related source class exists, such as class-incremental learning with pretrained features.","A direct stress test is to hold the base-session training recipe fixed across all methods, isolating whether the reported gain comes from ALT and B2N rather than from a stronger initialization."],"forward_implications":["Prior Semi-FSCIL methods lose accuracy when the unlabeled pool includes base and previously seen novel classes, confirming that the generalized setting is genuinely harder.","Using a dynamic, learnable threshold beats both discarding ambiguous samples and using a fixed threshold, according to the paper's ablations.","Calibrating novel-class distributions with base-class statistics raises novel-class accuracy across sessions.","On mini-ImageNet, ALDC reaches 69.77% average accuracy, 10.93% above the best Semi-FSCIL method and 11.51% above the best GSemi-FSCIL baseline."],"supporting_citations":[{"why":"Defines the FSCIL benchmark and supplies the TOPIC baseline used in comparisons.","marker":"[8]"},{"why":"Introduces continually evolved classifiers (CEC) that the paper adapts into the Semi-CEC baseline.","marker":"[9]"},{"why":"Presents the forward-compatible FACT method used as a state-of-the-art FSCIL baseline.","marker":"[10]"},{"why":"Supplies the uncertainty-aware distillation method UaD-CE, a main Semi-FSCIL baseline.","marker":"[15]"},{"why":"Provides the Us-KD baseline and the SS-iCarl and SS-NCM-CNN comparison points.","marker":"[17]"},{"why":"The distribution calibration technique that B2N builds on for generating synthetic few-shot features.","marker":"[30]"},{"why":"Offers the strong baseline for semi-supervised incremental few-shot learning that motivates the single-stage comparison.","marker":"[14]"}],"fun_headline_variants":["Ambiguous unlabeled samples boost few-shot incremental learning","Calibrating with confusing data sets new state of the art in FSCIL","Learnable threshold turns ambiguity into advantage for class-incremental learning","Exploiting ambiguous samples improves semi-supervised few-shot in","Use base classes to fix novel-feature drift for better incremental learning"],"cache_read_input_tokens":13440,"weakest_assumption_plain":"The headline gains assume the starting foundation is the same: ALDC enters at about 82% base-session accuracy while the methods it beats start near 72%, so if the base-session recipe is stronger, the improvement may come from initialization rather than from the new threshold and calibration modules.","fun_headline_variants_meta":{"raw":{"variants":["Ambiguous unlabeled samples boost few-shot incremental learning","Calibrating with confusing data sets new state of the art in FSCIL","Learnable threshold turns ambiguity into advantage for class-incremental learning","Exploiting ambiguous samples improves semi-supervised few-shot in","Use base classes to fix novel-feature drift for better incremental learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1324,"prompt_tokens":898,"completion_tokens":426,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":514,"completion_tokens_details":{"reasoning_tokens":336}},"tokens_in":514,"tokens_out":426,"duration_ms":5325,"temperature":1.0,"reasoning_tokens":336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:55:02.138191+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain ALDC with the base-session protocol used by the baselines so that session-0 accuracy matches theirs, then recompute average accuracy over all sessions; if the 10.93% advantage shrinks or vanishes, the claimed improvement comes from initialization rather than from ALT and B2N.","supporting_citations":[{"cited_title":"Few-shot class-incremental learning,","cited_arxiv_id":null,"evidence_quote":"Defines the FSCIL benchmark and supplies the TOPIC baseline used in comparisons."},{"cited_title":"Few-shot incre- mental learning with continually evolved classifiers,","cited_arxiv_id":null,"evidence_quote":"Introduces continually evolved classifiers (CEC) that the paper adapts into the Semi-CEC baseline."},{"cited_title":"For- ward compatible few-shot class-incremental learning,","cited_arxiv_id":null,"evidence_quote":"Presents the forward-compatible FACT method used as a state-of-the-art FSCIL baseline."},{"cited_title":"Uncertainty-aware distillation for semi-supervised few-shot class-incremental learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the uncertainty-aware distillation method UaD-CE, a main Semi-FSCIL baseline."},{"cited_title":"Uncertainty-guided semi-supervised few-shot class-incremental learn- ing with knowledge distillation,","cited_arxiv_id":null,"evidence_quote":"Provides the Us-KD baseline and the SS-iCarl and SS-NCM-CNN comparison points."},{"cited_title":"Free lunch for few-shot learning: Distribution calibration,","cited_arxiv_id":null,"evidence_quote":"The distribution calibration technique that B2N builds on for generating synthetic few-shot features."},{"cited_title":"A strong baseline for semi-supervised incremental few-shot learning,","cited_arxiv_id":null,"evidence_quote":"Offers the strong baseline for semi-supervised incremental few-shot learning that motivates the single-stage comparison."}],"review_version":1}