{"id":"69382d9a-dd18-4f44-ab73-58ee830e0b42","arxiv_id":"2501.09361","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A feature augmentation plus proxy-label contrastive learning framework reports SOTA FSCIL accuracy, but its evaluation uses extra proxy prototypes that bias comparisons.","lead":"This paper proposes a training method for few-shot class incremental learning that mixes augmented image features and adds proxy labels to improve base-class separation. The authors report state-of-the-art accuracy on three benchmarks, but the evaluation method gives their model extra class prototypes, which may inflate the gains.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on an evaluation that gives FACL two votes per base class while baselines get one; if recomputed without proxy aggregation, the reported gains likely collapse.","rationale":"The reader's weakest assumption is exactly the load-bearing issue. It is not a minor missing baseline: Section 3.3.4 explicitly changes the evaluation rule for FACL by aggregating original and proxy predictions, while all compared methods use one prototype per class. This makes the reported numbers not comparable under the standard FSCIL protocol. The base-session gap in Table 5 (86.20 vs 74.77 for CE) is the clearest symptom, and since the full method is the only one evaluated with the integrated rule, the gains cannot be attributed to representation quality. Additional weaknesses, such as the absence of error bars and the underspecified test-time procedure, are secondary to this evaluation bias. A fair re-evaluation could potentially rescue the method, and I would not rule out that FACL retains some value as a regularizer, but as submitted the central SOTA claim is unsupported. Therefore the REJECT verdict stands.","tokens_in":16128,"tokens_out":3018,"duration_ms":30524,"concrete_test":"Using the released code, rerun CIFAR100 with the Section 3.3.4 evaluation modified to ignore proxy-label predictions, keeping only original-class prototypes in Eq. 11, and rerun the Table 5 ablation with that same evaluation for every row. If FACL's session-0 accuracy drops from 86.20 by more than a few points, or if its average advantage over MICS/SAVC falls below the reported gaps (4.77% on CIFAR100, 0.6% on miniImageNet, 1.59% on CUB200), then the headline SOTA result is an artifact of the multi-prototype evaluation.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim ('outperforms current state-of-the-art methods across all datasets') is supported only by tables computed with an evaluation rule that is not applied to any baseline. Section 3.3.4 states: 'For evaluation in our approach, aggregate the original label predictions and proxy label predictions to compute the accuracy for the original labels.' Since proxy labels are generated for each base class (Section 3.3.2, P=2), this gives each base class two chances to be correct at test time: the original prototype and the proxy prototype. Baselines such as CEC, FACT, TEEN, ALICE, MICS, and SAVC are evaluated by standard nearest-mean classification with one prototype per class. The comparison is therefore not apples-to-apples. The likely result is visible in the ablation: the full model's session-0 accuracy is 86.20 vs 74.77 for plain CE, and the largest jump occurs exactly when proxy labels and feature augmentation are added. If the proxy branch is removed, session-0 accuracy may fall close to the CE baseline, and the 4.77% average CIFAR100 gain could shrink to noise. The paper also does not report error bars or standard deviations, so it is impossible to tell whether residual gains exceed run-to-run variance. The method may still be a useful training regularizer; the SOTA claim, as stated, is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FACL (Feature Augmentation driven Contrastive Learning) for few-shot class incremental learning (FSCIL). The method augments feature vectors, assigns proxy labels to the augmented features, and adds a contrastive loss to train the base-session representation, aiming to leave room for future incremental classes. Experiments on CIFAR100, miniImageNet, and CUB200 report average accuracy improvements of 4.77%, 0.6%, and 1.59% over prior state-of-the-art methods, with base-session gains as large as 86.20% on CIFAR100 (vs. 74.77% for plain cross-entropy).","tokens_in":16328,"tokens_out":5526,"duration_ms":50286,"significance":"If the reported gains were reliable, FACL would be a useful contribution to FSCIL by structuring the base-session feature space with feature augmentation and proxy classes. However, the central SOTA claim is undermined by an asymmetric evaluation protocol that credits proxy-label predictions to the original classes only for FACL, an underspecified inference procedure, and the absence of error bars. These issues mean the paper's headline results are not yet supported.","major_comments":[{"comment":"The evaluation aggregates original-label and proxy-label predictions, so a correct proxy prediction is counted as correct for the original class. Since each base class has P=2 labels under Eq. (5), FACL effectively has two prototypes per base class at test time, while baselines such as CEC, FACT, TEEN, ALICE, MICS, and SAVC are evaluated with one prototype per class. Tables 2–4 therefore do not provide an apples-to-apples comparison. This asymmetry likely explains the large session-0 gain (86.20 vs. 74.77 in Table 5). The authors must re-run the evaluation with a single prototype per class, or apply the same proxy-aggregation rule to all baselines, and report those results.","section":"3.3.4, Eq. (11)"},{"comment":"The inference procedure is underspecified. Training uses F_comb = concat(Z_ori, F_aug) (Eq. 4) and the classifier loss in Eq. (8) operates on augmented versions F_a of F_comb. However, evaluation Eq. (11) computes cosine similarity between f(x) (the original feature only) and prototypes c_s_y (Eq. 10). It is unclear how the classifier trained on concatenated features is used at test time, how proxy prototypes are constructed for incremental sessions, and how the 'integrated' evaluation combines classifier logits with NCM distances. Without a precise inference algorithm, the reported numbers cannot be reproduced.","section":"3.3.1–3.3.4"},{"comment":"No error bars, standard deviations, or number of seeds are reported. The claimed miniImageNet improvement is 0.6% average accuracy, which is within typical run-to-run variance for FSCIL benchmarks. The authors should report mean ± std over at least 3 seeds for their method and the closest baselines (e.g., MICS, SAVC) to support the claim of significant improvement.","section":"Tables 2–4"},{"comment":"The ablation is confounded with the evaluation protocol. The CE baseline uses one prototype per class, while the full model (CE+SSCL+PC+FA) benefits from proxy-prediction aggregation. The +12.33% final-session gain over CE therefore conflates the effect of the method with the effect of the evaluation rule. The ablation should be performed under a single evaluation protocol, either with proxy aggregation for all rows or without it for the full model.","section":"4.4, Table 5"}],"minor_comments":[{"comment":"The citation 'MICS [24]' is inconsistent with the reference list, where MICS is [16] (the same table also has a typo in 'SA VC'; it should be 'SAVC' or 'SA-VC').","section":"Table 4"},{"comment":"The paper calls the contrastive loss 'self-supervised' but cites [15] (Supervised Contrastive Learning) and uses MoCo [6]; please clarify which loss is actually used and provide correct references.","section":"2.3, 3.3.3"},{"comment":"The notation in Eq. (7) is under-specified; please define S(j), N(j), za, zk+, and explain how the feature and label queue is used.","section":"Eq. (7)"},{"comment":"The t-SNE and confusion-matrix visualizations are qualitative; consider adding quantitative metrics, such as inter-class/intra-class distance or average cosine similarity, to support the claim of improved feature separation.","section":"4.3"},{"comment":"The sentence 'our approach maintains the same number of proxy classes as the original ones' is confusing because Eq. (5) with P=2 doubles the label space; please rephrase.","section":"Discussion"},{"comment":"The statement 'We follow the CEC protocol' should specify whether the same data splits, evaluation code, and hyperparameter selection procedure are used for all baselines.","section":"4.1"}],"recommendation":"major_revision","confidential_remarks":"The evaluation asymmetry in Section 3.3.4 is the most serious issue; if the authors cannot demonstrate gains under a fair single-prototype protocol, the SOTA claim should be withdrawn. The paper currently does not support its central claim, but the flaws are experimentally fixable, so a major revision is appropriate if the authors can re-run the evaluation and provide a precise inference specification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's core idea is sensible: mix augmented image features to create proxy classes, add a supervised contrastive loss, and train the base session with this expanded label space. The combination is new, though each piece is known from ALICE, MICS, and SAVC. The ablations are reasonably thorough—they test different mixture sources, the delta coefficient, and compare against Mixup/CutMix/Manifold-Mixup. That part is fine.\n\nThe problem is the evaluation. Section 3.3.4 says they aggregate original and proxy label predictions for accuracy, which effectively gives each base class two prototypes at test time. Baselines use nearest-mean with one prototype per class. That is not apples-to-apples. It likely explains the big base-session jump on CIFAR100 (86.20 vs 74.77) and the 4.77% average gain. The ablations are confounded in the same way: rows with proxy classes get the integrated evaluation, the CE-only row does not. So the component contributions are not measurable from these tables.\n\nThere is also a train-test mismatch that is never clarified. Training uses F_comb, a concatenation of original and augmented features, but Eq. 11 evaluates on f(x). How the integrated classifier handles the dimension difference is unspecified. And no error bars or multiple seeds are reported, so even the residual gains after fixing the evaluation could be noise.\n\nThe stress-test note holds up on reading. This is not a minor oversight; it is load-bearing for the central claim. The method might still be a useful training regularizer, but the paper does not demonstrate that with the current protocol.\n\nRecommendation: send it to a serious referee. The method is concrete and the flaw is fixable in revision. A referee should ask for re-evaluation against baselines with a single prototype per class, error bars, and a clear description of the inference procedure. But as is, I would not cite the numbers.","headline":"The SOTA claim is undermined by an evaluation that gives FACL two votes per base class while baselines get one; the idea is a plausible regularizer but the reported margins are not trustworthy.","tokens_in":16912,"tokens_out":2257,"would_cite":false,"duration_ms":50665,"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":"This paper shows that mixing two augmented feature vectors to create proxy classes, together with a self-supervised contrastive loss, expands the base-session feature space so that new classes can be added in few-shot increments with less…","keywords":["few-shot class incremental learning","feature augmentation","proxy classes","self-supervised contrastive learning","base session representation","prototype evaluation","CIFAR100","miniImageNet"],"falsifier":"Run FACL and the strongest baselines under identical evaluation rules: evaluate FACL using only its original-class prototypes (dropping proxy predictions), and evaluate baselines with the same proxy-label aggregation (adding a second prototype per class). If FACL's average-accuracy lead over the best baseline shrinks below the reported margins or reverses, the central claim is not about the representation but about the evaluation protocol.","tokens_in":15856,"feed_emoji":"🧩","tokens_out":11559,"duration_ms":89135,"temperature":0.7,"pith_summary":"The paper asks how to train the base session in few-shot class incremental learning so that later classes, which arrive with only a few labeled examples, can be added without collapsing into old classes. It proposes Feature Augmentation driven Contrastive Learning (FACL), which blends feature vectors from transformed images, assigns proxy labels to those mixtures, and trains with a contrastive loss to push base-class representations apart. The result, the authors report, is a base feature space with more room for future classes: average accuracy improves by 4.77% on CIFAR100, 0.6% on miniImageNet, and 1.59% on CUB200 over prior state-of-the-art methods, with the largest gains in the base session. The claim is that a deliberately expanded base representation is the key to accommodating few-shot novel classes without retraining the backbone.","feed_headline":"Feature mixing lifts few-shot class-incremental accuracy up to 4.8%","feed_subtitle":"Proxy classes and contrastive loss expand the base feature space so new classes fit without overwriting old ones.","key_machinery":"The load-bearing mechanism is the proxy-class feature mixture: $F^i_{\\text{aug}} = \\delta Z^i_{\\text{aug}} + (1-\\delta) Z^j_{\\text{aug}}$ with $\\delta = 0.5$, where $Z_{\\text{aug}}$ are features of images subjected to a 180° rotation plus RGB permutation. Each mixture receives a proxy label $y_p = y\\times 2 + p$, effectively doubling the label space during base training and creating placeholders for future classes. A MoCo-style self-supervised contrastive loss is applied to two augmented views of $F_{\\text{comb}}$, and inference uses cosine-similarity nearest-mean classification over both original and proxy prototypes. These components together are what the paper credits for separating base classes and leaving room for incremental classes.","core_discovery":"FACL's central claim is that the base session, not the incremental sessions, should do the work of preparing for future classes. The authors generate an augmented feature vector $F^i_{\\text{aug}} = \\delta Z^i_{\\text{aug}} + (1-\\delta) Z^j_{\\text{aug}}$ by mixing features of two differently transformed images, assign these mixtures proxy labels $y_p = y \\times P + p$ (with $P=2$), and train the classifier on both original and proxy labels. A MoCo-style self-supervised contrastive loss on the combined features $F_{\\text{comb}} = \\mathrm{concat}(Z_{\\text{ori}}, F_{\\text{aug}})$ then separates base-class representations. At evaluation, the model aggregates original-label and proxy-label predictions under the nearest-mean classifier. The paper reports that this recipe outperforms current state-of-the-art FSCIL methods on all three benchmarks, with base-session accuracy rising to 86.20 on CIFAR100 relative to 79.13 for the closest prior method.","pith_inferences":["One testable extension would be to apply the same proxy-label evaluation protocol to existing baselines, giving each baseline a second prototype per class, to isolate how much of FACL's gain comes from the representation versus the aggregation rule.","The proxy-class idea could transfer to other continual-learning settings, such as task-incremental or online continual learning, where reserving placeholder prototypes might reduce inter-task confusion without rehearsal.","The choice of $\\delta=0.5$ and only one augmentation type (rotation plus RGB permutation) suggests a cheap optimization surface: sweeping $\\delta$ and the augmentation family per dataset could yield further gains, or reveal where the mixture stops being semantically valid.","Because the base-session accuracy jumps roughly 11 points when feature augmentation is added, the paper implicitly claims that overfitting to base classes is a major bottleneck; an independent test would be to measure FACL's base-session calibration or feature-space coverage versus baselines."],"forward_implications":["If FACL is right, base-session training should be judged not only by its own accuracy but by how much unused feature space it leaves for later classes.","Frozen-backbone FSCIL methods can improve without touching the incremental session by changing only how the base representation is shaped.","Proxy-class placeholders make the number of base training labels a tunable resource: doubling the label space via $P=2$ is a simple lever that other FSCIL pipelines could adopt.","The contrastive loss and feature mixing appear to be complementary: the ablation shows the largest gains only when both are present, suggesting the expanded space and the separation pressure reinforce each other.","The reported gains are largest on the base session and on CIFAR100, so the method's benefit may scale with the number of base classes and the difficulty of separating them."],"supporting_citations":[{"why":"CEC supplies the evaluation protocol and a primary frozen-backbone baseline that FACL builds on and compares against.","marker":"[44]"},{"why":"FACT establishes the idea of reserving feature space for future classes via virtual prototypes, which FACL extends with proxy classes.","marker":"[48]"},{"why":"MoCo provides the momentum contrastive architecture used for the self-supervised contrastive loss.","marker":"[6]"},{"why":"The supervised contrastive loss formulation is cited for the SSCL separation loss in Eq. 7.","marker":"[15]"},{"why":"TOPIC defines the FSCIL benchmark setup and provides an early baseline and the fine-tune reference.","marker":"[32]"},{"why":"F2M is a flat-minima frozen-backbone method whose base-session representation goal motivates FACL.","marker":"[28]"},{"why":"TEEN is a recent training-free prototype-calibration baseline that FACL compares against.","marker":"[37]"},{"why":"SAVC is a semantic-aware virtual contrastive constraint baseline included in the comparison tables.","marker":"[31]"}],"fun_headline_variants":["Feature mixing lifts few-shot incremental accuracy by 4.8%","Proxy labels and contrastive loss expand base space for FSCIL","Contrastive feature augmentation boosts FSCIL base separation","Prepping base features with mixing improves few-shot class-incremental"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on comparing FACL's aggregated original-plus-proxy predictions against baselines that use a single prototype per class, so the reported gains could be an artifact of the evaluation protocol rather than the learned representation.","fun_headline_variants_meta":{"raw":{"variants":["Feature mixing lifts few-shot incremental accuracy by 4.8%","Proxy labels and contrastive loss expand base space for FSCIL","Contrastive feature augmentation boosts FSCIL base separation","Prepping base features with mixing improves few-shot class-incremental"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001582,"raw_usage":{"total_tokens":6319,"prompt_tokens":962,"completion_tokens":5357,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":5284}},"tokens_in":578,"tokens_out":5357,"duration_ms":38837,"temperature":1.0,"reasoning_tokens":5284,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:06:02.768076+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FACL and the strongest baselines under identical evaluation rules: evaluate FACL using only its original-class prototypes (dropping proxy predictions), and evaluate baselines with the same proxy-label aggregation (adding a second prototype per class). If FACL's average-accuracy lead over the best baseline shrinks below the reported margins or reverses, the central claim is not about the representation but about the evaluation protocol.","supporting_citations":[{"cited_title":"Few-shot incremental learning with contin- ually evolved classifiers","cited_arxiv_id":null,"evidence_quote":"CEC supplies the evaluation protocol and a primary frozen-backbone baseline that FACL builds on and compares against."},{"cited_title":"Forward compatible few-shot class-incremental learning","cited_arxiv_id":null,"evidence_quote":"FACT establishes the idea of reserving feature space for future classes via virtual prototypes, which FACL extends with proxy classes."},{"cited_title":"Few-shot class- incremental learning","cited_arxiv_id":null,"evidence_quote":"TOPIC defines the FSCIL benchmark setup and provides an early baseline and the fine-tune reference."},{"cited_title":"Overcoming catastrophic for- getting in incremental few-shot learning by finding flat min- ima","cited_arxiv_id":null,"evidence_quote":"F2M is a flat-minima frozen-backbone method whose base-session representation goal motivates FACL."},{"cited_title":"Few-shot class-incremental learning via training-free prototype calibration","cited_arxiv_id":null,"evidence_quote":"TEEN is a recent training-free prototype-calibration baseline that FACL compares against."},{"cited_title":"Learning with fantasy: Semantic-aware virtual contrastive constraint for few-shot class-incremental learning","cited_arxiv_id":null,"evidence_quote":"SAVC is a semantic-aware virtual contrastive constraint baseline included in the comparison tables."}],"review_version":1}