{"id":"44bc5d58-37cd-40d6-8792-bea13c1c663f","arxiv_id":"2507.05221","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CTA aligns a SimCLR-trained encoder to a frozen supervised encoder, then adapts only the self-supervised encoder at test time, improving corrupted-image classification over prior test-time training methods.","lead":"This paper introduces CTA, a test-time training method that aligns a self-supervised image encoder with a supervised one so that the self-supervised encoder can be adapted at test time without conflicting gradients. It reports state-of-the-art robustness on common image corruption benchmarks, with gains of about 4.5 points over the previous best method on CIFAR-10-C.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4)'s alignment loss cannot be instantiated as written: π(g(x)) is d-dimensional while w_i=h(f(x_i)) is C-dimensional, so the central alignment mechanism is undefined until this inconsistency is resolved.","rationale":"The reader identified persistence of alignment under test-time SimCLR updates as the weakest assumption. I agree that persistence is unproven, but there is a more fundamental blocker: the formal definition of the alignment loss is internally inconsistent. The central claim is that L_align transfers the supervised decision boundary into the SSL encoder; if L_align cannot be evaluated as written, the claim is not reproducible. The paper contains strong empirical support, including Tables 2, 4, and 5, Figure 2, and Table 6, and the CTA-C ablation shows the SSL encoder alone is already competitive, so I do not think the results are necessarily wrong. However, the missing tensor-shape specification is a correctness risk that should be resolved before the method is adopted. The persistence concern would then be tested by measuring classifier accuracy and alignment distance over the exact iteration count used in the final tables; the paper only reports 20-iteration DBI/distance, and the final-table iteration count is not stated. I therefore keep the verdict CONDITIONAL, contingent on the authors providing the corrected equation or code and the final-table adaptation protocol.","tokens_in":12118,"tokens_out":10311,"duration_ms":128635,"concrete_test":"Obtain the released code (or request it) and run a shape assertion for one batch on CIFAR-10: print the shape of π(g(x)) and the shape of the w_i tensor actually used in L_align. If w_i is f(x_i), re-run CIFAR-10-C Table 2 with the corrected Eq. (4) notation; if the reported 87.42 reproduces, the issue is a typo and the conditional can be lifted. If w_i is h(f(x_i)) with shape (B,10) while π(g(x)) is (B,2048), the paper's Eq. (4) does not describe the implemented method, and the central alignment claim must be re-derived from the actual loss. If code is unavailable, ask the authors for the exact tensor shapes and for whether C or d is the aligned dimension before accepting the mechanism.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.4 defines the alignment triplet with w_i = h(f(x_i)) and calls it \"the output of the supervised encoder.\" But Section 3.2 defines h(·) as the classifier, so h(f(x_i)) has dimension equal to the class count C (10/100/200), while Section 3.1 says π preserves the input dimension of g(·) to match the supervised encoder output. The cosine similarity in Eq. (4) therefore compares vectors of different dimensions; the loss cannot be computed. If the intended w_i is actually f(x_i) (the encoder feature), then Eq. (4) and the associated text contain a material misstatement, and the \"decision boundary distillation\" claim is really feature-space alignment rather than classifier-output alignment. If the intended w_i is truly h(f(x_i)), then the dimension-preservation statement is false and the architecture-agnostic claim needs revision. Either way, the central construction that the SSL encoder is aligned to the frozen classifier is not well defined by the paper, and every subsequent result, including the persistence evidence in Sec. 5.3, depends on an unstated implementation choice.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CTA (Cross-Task Alignment), a test-time training method that trains a supervised encoder and a self-supervised encoder on the source domain, then aligns the self-supervised encoder's representations to the supervised encoder's representations with a contrastive loss. At test time, the self-supervised encoder is used as the feature extractor for the frozen classifier and is updated with the SimCLR contrastive loss. The authors report consistent average accuracy gains over prior TTT methods on CIFAR-10-C, CIFAR-100-C, and TinyImageNet-C, and include an ablation (CTA-C) that isolates the alignment component. The central claimed mechanism is that alignment transfers the supervised decision boundary into the self-supervised encoder while preserving the intrinsic robustness of self-supervised learning.","tokens_in":12359,"tokens_out":6134,"duration_ms":69710,"significance":"If the results hold, CTA would be a practically useful TTT method because it does not require a specialized architecture and can be applied to a pretrained backbone by duplicating it and fine-tuning the copy. The CTA-C ablation is a strength: it isolates the effect of the alignment stage and suggests that the gains over ReC-TTT are not solely due to using a self-supervised objective. The evaluation covers three standard corruption benchmarks and compares with recent methods. However, the formulation of the central alignment loss is inconsistent as written, the experimental protocol is incompletely specified, and some hyperparameters appear to be selected using target-domain accuracy. These issues must be resolved before the reported state-of-the-art claims can be accepted.","major_comments":[{"comment":"The alignment loss as written cannot be instantiated. In Sec. 3.2, h(·) is the classifier, so w_i = h(f(x_i)) has dimension equal to the number of classes C (10, 100, or 200), whereas Sec. 3.1 states that π(g(x_i)) preserves the encoder dimension d. The cosine similarity in Eq. (4) therefore compares vectors of different dimensions and the loss is undefined. The surrounding text (Sec. 3.4 and Fig. 1) and the dimension-preservation statement in Sec. 3.1 indicate that the intended target is f(x_i), the supervised encoder's feature vector, not the classifier output. Please correct the definition of w_i and revise the 'decision boundary distillation' claims in the Abstract, Sec. 1, Sec. 3.5, and Sec. 6 accordingly: the construction as stated aligns feature spaces, not classifier outputs. Because every downstream result in the paper depends on this loss, the implementation choice must be stated explicitly.","section":"Sec. 3.4, Eq. (4)"},{"comment":"The final accuracy numbers in Tables 2, 4, and 5 do not state how many test-time adaptation iterations were used. Section 4.2 gives batch size and learning rate only; Section 5.4 discusses 20 iterations and Figure 3, but never states whether the reported tables correspond to a fixed iteration count, the best iteration, or some other protocol. In addition, Tables 4 and 5 contain no standard deviations for CTA or the comparison methods, and in Table 2 the CTA-C column has no variance. Please report the exact adaptation protocol and run-to-run or batch-level variability for all main results.","section":"Sec. 4.2 and Sec. 5.4"},{"comment":"The alignment temperature and test-time temperature appear to be selected using target-domain accuracy. Table 1 reports accuracy on the corrupted test sets as the alignment temperature varies, and Sec. 5.1 states that τ=0.5 was chosen for the final experiments and that smaller τ values were found to be better for test-time training. Since test-time training has no labels or validation set on the target distribution, this selection protocol can inflate the reported gains relative to methods whose hyperparameters are fixed on source data. Please clarify whether any target-domain information was used to choose τ_align and τ_con, and if so, provide a source-only or held-out validation procedure, or report sensitivity results that do not depend on target labels.","section":"Sec. 5.1, Table 1"}],"minor_comments":[{"comment":"The phrase 'extract layer hyperparameter' in the Introduction and 'update-layer hyperparameter' in Sec. 3.5 should be made consistent; the current wording is confusing.","section":"Introduction, Sec. 3.5"},{"comment":"The bibliography contains duplicate entries for the same works: [3] and [4] are both ReC-TTT, [15] and [16] are both TTT++, and [21] and [22] are both NC-TTT. These should be merged.","section":"References"},{"comment":"The sentence 'CTA-C demonstrates an average loss of 4.48% compared to CTA' should say '4.48 percentage points lower' to avoid confusion with the training loss.","section":"Sec. 5.5, Table 2"},{"comment":"The sentence about being 'on par with NC-TTT for the smallest amount of variability (i.e., ±0.22)' is unclear because Table 2 reports per-corruption standard deviations rather than a single average variability for the full benchmark; please clarify what quantity is being compared.","section":"Sec. 5.2, Table 2"},{"comment":"The caption refers to 'the second column' and 'the third column' but the figure layout appears to be a grid of rows and columns; please make the description unambiguous.","section":"Fig. 2 caption"}],"recommendation":"major_revision","confidential_remarks":"The dimension mismatch in Eq. (4) is the most serious issue and is likely a notation error rather than a fundamental flaw, but it must be corrected because the entire method is defined by that loss. I would also require the authors to state the exact test-time iteration count and to address the target-set hyperparameter selection concern before considering the paper for acceptance. The consistent gains across three benchmarks and the CTA-C ablation are encouraging; they suggest the method is worth publishing after the requested revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The stress-test note is right, and it's the thing to fix. Eq. (4) defines w_i = h(f(x_i)) and calls it the output of the supervised encoder, but h is the classifier, so w_i has C dimensions while π(g(x_i)) is d-dimensional. Cosine similarity is undefined across different dimensions. Either w_i should be f(x_i), in which case the 'decision boundary distillation' language is wrong and it's feature-space alignment, or the paper needs a projection head. Every downstream result depends on this choice, so the central mechanism is not actually specified.\n\nWhat's genuinely new: CTA's two-encoder setup with contrastive alignment, then using only the SSL encoder for test-time updates, is a real departure from shared-backbone TTT. The CTA-C ablation is a nice control: a frozen SSL encoder plus linear classifier already matches ReC-TTT, and the alignment step adds 4.5 points. That's a credible empirical story. The results on three corruption benchmarks are consistent and the gains over ReC-TTT are meaningful.\n\nSoft spots: the protocol is under-specified. No standard deviations on CIFAR-100-C or TinyImageNet-C, no iteration count for the main tables, and hyperparameters (τ_align) chosen based on target-benchmark performance. These are moderate and fixable. The bigger open question is whether source-only alignment persists under test-time SimCLR updates; the DBI/distance analysis in Sec. 5.3 is suggestive but not a proof. I'd also check whether the comparison methods' numbers are apples-to-apples (same backbone, same augmentation, same iterations).\n\nBottom line: if the w_i issue is a typo, this is a solid, novel contribution that deserves referee time. As written, it cannot be reproduced. I'd send it to review with a strong request to fix the alignment definition and report the missing experimental details. The central idea doesn't collapse; it needs careful repair.","headline":"A genuinely novel two-encoder TTT idea with strong results, but the main alignment loss contains a dimension mismatch that makes it undefined as written; this is fixable but must be addressed before the paper can be reproduced.","tokens_in":12901,"tokens_out":3522,"would_cite":false,"duration_ms":41191,"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":"A contrastive alignment step lets a self-supervised encoder serve as the feature extractor for a frozen classifier during test-time training, improving corrupted-image accuracy over prior TTT methods.","keywords":["test-time training","test-time adaptation","contrastive learning","representation alignment","distribution shift","SimCLR","CIFAR-10-C","self-supervised learning"],"falsifier":"A reader could settle the claim by measuring the distance between the feature distributions of $\\pi(g(\\cdot))$ and $f(\\cdot)$ on a target corruption after 20 adaptation iterations: CTA predicts this distance stays small or shrinks, while an unaligned shared-backbone model drifts. If the distance grows while accuracy still rises, the alignment story is wrong even if the benchmark numbers hold; alternatively, running the same two-stage alignment with a non-contrastive self-supervised task and seeing the gains vanish would show the effect is specific to SimCLR rather than alignment.","tokens_in":11921,"feed_emoji":"🔄","tokens_out":5640,"duration_ms":57061,"temperature":0.7,"pith_summary":"CTA's central claim is that the conflict between the supervised main task and the self-supervised auxiliary task in test-time training can be removed by separating them into two encoders and aligning them. The paper trains a SimCLR-style encoder and a supervised encoder on the same source data, then uses a contrastive cross-encoder loss to pull the self-supervised features toward the frozen supervised features. After this alignment, the self-supervised encoder can act as the feature extractor for the frozen classifier, and updating only the self-supervised encoder with the contrastive loss at test time improves accuracy on CIFAR-10-C, CIFAR-100-C, and TinyImageNet-C. A reader should care because the method needs no special architecture or chosen update layer, and it beat four recent test-time training baselines on those benchmarks.","feed_headline":"Aligned self-supervised encoder tops corrupted-image benchmarks","feed_subtitle":"A frozen classifier plus an adapted self-supervised copy beats multi-task test-time training by 3–5 points.","key_machinery":"The load-bearing mechanism is the cross-encoder contrastive alignment loss $\\mathcal{L}_{\\mathrm{align}}$ (Eq. 4), which treats each augmented view of an image as a positive pair with the frozen supervised encoder's feature vector for the same image, and all other batch features as negatives. This student-teacher setup transfers the supervised feature geometry into the self-supervised encoder while the SimCLR objective $\\mathcal{L}_{\\mathrm{con}}$ remains the encoder's native task. At test time only $\\mathcal{L}_{\\mathrm{con}}$ is used, so the updated features stay in the same contrastive space that was aligned to the classifier during source training.","core_discovery":"The discovery, on the paper's own terms, is that the decision boundary of the supervised classifier can be transferred into the self-supervised encoder through alignment, without ever training that encoder on labels. Concretely, after minimizing $\\mathcal{L}_{\\mathrm{align}}$ (Eq. 4) between the two encoders on source data, the test-time model $h(\\pi(g(\\cdot)))$ uses the self-supervised encoder as the feature extractor and the frozen classifier as the head; test-time adaptation with $\\mathcal{L}_{\\mathrm{con}}$ updates all parameters of $\\pi(g(\\cdot))$. The paper reports that this aligned self-supervised encoder beats the supervised encoder it was aligned to on two of three source datasets and is more robust on all three corrupted benchmarks, with the full CTA pipeline reaching 87.42 on CIFAR-10-C, 63.95 on CIFAR-100-C, and 41.76 on TinyImageNet-C. The claim is that alignment avoids gradient interference while preserving the intrinsic shift-robustness of self-supervised representations.","pith_inferences":["A natural extension the paper leaves implicit is testing whether the same alignment transfers when the self-supervised task is replaced with masked image modeling or other objectives; the method's architecture-agnostic claim predicts the gain should persist.","The inverse temperature behavior (larger $\\tau$ for alignment, smaller for test-time adaptation) suggests the teacher should be smoother than the student during adaptation; that asymmetry could be studied as a general principle for alignment-based adaptation.","If the representation-space alignment is the true cause of the gains, then CTA should work as a plug-in for any off-the-shelf frozen classifier, not only the ResNet50 fine-tuned here; a reader could test this by applying the two-stage alignment to a ViT or a detection backbone.","The paper's DBI and centroid-distance evidence implies that test-time contrastive updates push the self-supervised features closer to the frozen classifier's space, so a sharper test would be whether the distance stays bounded under much longer adaptation than the 20 iterations they report."],"forward_implications":["A pretrained supervised model can be turned into a TTT model by duplicating it, training the copy with a self-supervised loss, and aligning the copy to the original, with no change to the architecture.","The test-time adapter is the self-supervised encoder alone; the classifier stays frozen, so adaptation updates do not need to be filtered by layer choice or entropy signals.","Because the encoder never sees the cross-entropy loss, the method preserves the self-supervised representation's robustness under large corruptions, which the paper argues explains the gains over standard multi-task TTT.","On the benchmarks tested, the full CTA pipeline outperforms the baselines on average across all corruption types on CIFAR-10-C and on most corruption types on CIFAR-100-C and TinyImageNet-C, with gains of 4.51, 3.29, and 4.27 points over the previous best ReC-TTT."],"supporting_citations":[{"why":"Supplies the self-supervised task (SimCLR) used for both the pretraining and test-time adaptation losses.","marker":"[2]"},{"why":"Defines the TTT++ baseline whose shared-backbone conflicts CTA targets, and supplies the design that the self-supervised task and some hyperparameters are taken from.","marker":"[16]"},{"why":"Defines ReC-TTT, the most recent prior baseline that CTA compares against and outperforms on all three benchmarks.","marker":"[4]"},{"why":"Defines NC-TTT, a baseline and source of the test-time settings (batch size and learning rate) used by CTA.","marker":"[21]"},{"why":"Defines ClusT3, an architecture-based TTT baseline that CTA is compared against.","marker":"[8]"},{"why":"Supports the claim that combining features from differently trained encoders can beat either individual model, which motivates alignment as a learning objective.","marker":"[7]"},{"why":"Supports the claim that self-supervised representations are more robust to large distribution shifts than supervised ones.","marker":"[27]"},{"why":"Provides the multi-modal contrastive-learning inspiration for aligning two encoders in a shared latent space.","marker":"[24]"},{"why":"Documents gradient interference in multi-task learning, the failure mode CTA is designed to avoid.","marker":"[35]"}],"fun_headline_variants":["Cross-task encoder alignment boosts test-time robustness","Aligning encoders improves test-time training","Self-supervised encoder alignment beats multi-task TTT","Encoder alignment transfers boundary to self-supervised model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole method assumes that the decision boundary learned by the supervised encoder during source alignment continues to separate the classes after test-time SimCLR updates on corrupted images, so the frozen classifier still works on the adapted features.","fun_headline_variants_meta":{"raw":{"variants":["Cross-task encoder alignment boosts test-time robustness","Aligning encoders improves test-time training","Self-supervised encoder alignment beats multi-task TTT","Encoder alignment transfers boundary to self-supervised model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000175,"raw_usage":{"total_tokens":1274,"prompt_tokens":923,"completion_tokens":351,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":293}},"tokens_in":539,"tokens_out":351,"duration_ms":4290,"temperature":1.0,"reasoning_tokens":293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:30:12.950504+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the claim by measuring the distance between the feature distributions of $\\pi(g(\\cdot))$ and $f(\\cdot)$ on a target corruption after 20 adaptation iterations: CTA predicts this distance stays small or shrinks, while an unaligned shared-backbone model drifts. If the distance grows while accuracy still rises, the alignment story is wrong even if the benchmark numbers hold; alternatively, running the same two-stage alignment with a non-contrastive self-supervised task and seeing the gains vanish would show the effect is specific to SimCLR rather than alignment.","supporting_citations":[{"cited_title":"Ttt++: When does self-supervised test-time training fail or thrive? InThirty-Fifth Conference on Neural Information Processing Systems, 2021","cited_arxiv_id":null,"evidence_quote":"Defines the TTT++ baseline whose shared-backbone conflicts CTA targets, and supplies the design that the self-supervised task and some hyperparameters are taken from."},{"cited_title":"Rec-ttt: Contrastive feature reconstruction for test-time training, 2024","cited_arxiv_id":null,"evidence_quote":"Defines ReC-TTT, the most recent prior baseline that CTA compares against and outperforms on all three benchmarks."},{"cited_title":"Vargas Hakim, Mehrdad Noori, Milad Cheraghalikhani, Ali Bahri, Moslem Yazdan- panah, Ismail Ben Ayed, and Christian Desrosiers","cited_arxiv_id":null,"evidence_quote":"Defines NC-TTT, a baseline and source of the test-time settings (batch size and learning rate) used by CTA."},{"cited_title":"Clust3: Information invariant test- time training","cited_arxiv_id":null,"evidence_quote":"Defines ClusT3, an architecture-based TTT baseline that CTA is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the claim that combining features from differently trained encoders can beat either individual model, which motivates alignment as a learning objective."},{"cited_title":"V ogt, Philip H","cited_arxiv_id":null,"evidence_quote":"Supports the claim that self-supervised representations are more robust to large distribution shifts than supervised ones."},{"cited_title":"Gradient surgery for multi-task learning, 2020","cited_arxiv_id":null,"evidence_quote":"Documents gradient interference in multi-task learning, the failure mode CTA is designed to avoid."}],"review_version":1}