{"id":"52fe1896-2d16-46ff-a0fc-4c924b98629f","arxiv_id":"2505.18514","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"BiTTA guides test-time model adaptation with a few binary correct/incorrect feedback labels, combining feedback-guided updates on uncertain samples with agreement-based self-adaptation on confident ones.","lead":"A new test-time adaptation method, BiTTA, uses simple correct/incorrect human feedback on a few sample predictions to guide model adaptation under domain shift. It outperforms prior adaptation methods on corruption and domain-shift benchmarks, and beats a full-class-label active TTA method under the paper's equal labeling-cost comparison.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ABA's agreement criterion (Eq. 8) may select confidently wrong self-training targets on severe shifts; since ABA alone accounts for 82.64/87.20 on CIFAR10-C, the central accuracy claim hinges on an unvalidated assumption.","rationale":"The reader identified the same underlying weakness: ABA's agreement criterion is the method's most load-bearing component because it provides the majority of the accuracy gain (82.64 vs 87.20 on CIFAR10-C) without any oracle signal. The paper's only evidence for the criterion's reliability is Figure 4b, which is shown for CIFAR10-C alone; on shifts where the model is less accurate, dropout consistency is not a guarantee of correctness. I considered alternative concerns: the 13.3%p headline inconsistency, the alpha=1 vs alpha=2 discrepancy, and the PACS comparison where full-class SimATTA (75.80) slightly exceeds BiTTA (75.57) under equal sample count. These are real but secondary; they affect reporting and scope, whereas the ABA assumption affects the method's validity on the target use case (unseen shifts). A concrete test, measuring the accuracy of S_ABA on the harder corruption benchmarks and comparing against an oracle-selected variant, would settle whether the assumption holds. If it fails, the paper should either add a confidence threshold to ABA or restrict the claim to shifts where agreement correlates with accuracy. Thus the reader's CONDITIONAL verdict is appropriate, with the condition being this additional validation.","tokens_in":31275,"tokens_out":14308,"duration_ms":113618,"concrete_test":"On CIFAR100-C and Tiny-ImageNet-C, replicate Figure 4b by logging the ground-truth accuracy of S_ABA (Eq. 8) per batch across the online stream. If the mean accuracy of agreement samples is not substantially above the overall accuracy (e.g., within 10 p.p.), the self-training targets in Eq. 12 are too noisy. Additionally, run BiTTA with S_ABA replaced by a set of samples whose pseudo-labels are verified correct (oracle-ABA); if real-ABA accuracy is more than about 3 p.p. below oracle-ABA, the agreement criterion is not a sufficient confidence signal on that shift and the central claim requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that ABA's self-training on S_ABA (Eq. 8) is beneficial. This holds only if agreement between the deterministic prediction and the MC-dropout argmax implies the pseudo-label is correct. The paper validates this only via Figure 4b on CIFAR10-C, where the source model already attains roughly 57% accuracy. On more severe shifts (e.g., Tiny-ImageNet-C, SrcValid around 21%), a model can be consistently wrong across dropout masks; agreement then selects confidently wrong targets, and the reward +1 in Eq. 9 becomes a positive reward for reinforcing errors. The ablation shows ABA alone reaches 82.64% on CIFAR10-C, contributing the bulk of BiTTA's 87.20%; a failure of this assumption on other shifts would reduce BiTTA to BFA-only (about 58.9%), below several baselines. No per-dataset accuracy of S_ABA is reported for CIFAR100-C or Tiny-ImageNet-C, so the assumption is unverified exactly where the method's margin over baselines is thinnest. If ABA degrades, the claim of state-of-the-art accuracy is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces a test-time adaptation setting in which an oracle provides binary correct/incorrect feedback on a few model predictions per batch. The authors propose BiTTA, a dual-path method combining Binary Feedback-guided Adaptation (BFA) on uncertain samples with Agreement-Based self-Adaptation (ABA) on confident samples, both expressed through a REINFORCE-style policy-gradient objective with Monte Carlo dropout. Experiments on corruption benchmarks (CIFAR10-C, CIFAR100-C, Tiny-ImageNet-C) and domain shifts (PACS) claim that BiTTA outperforms adapted TTA baselines and full-class active TTA (SimATTA) under equal labeling budgets, with the abstract reporting a 13.3 percentage-point average improvement.","tokens_in":31523,"tokens_out":12945,"duration_ms":93288,"significance":"If the central claims hold, the paper makes a useful contribution by proposing a cheaper supervision interface (binary correctness feedback) for test-time adaptation and showing that it can outperform methods that use full-class labels. The manuscript is strong in its breadth of experiments—multiple corruption severities, domain generalization, batch-size-1 and non-iid streams, additional architectures, and a public code release—and the ablation isolating BFA and ABA clarifies the contribution of each path. However, the significance is tempered by (i) the undefined and internally inconsistent headline improvement, (ii) a mismatch between the stated policy-gradient derivation and the implemented loss, and (iii) an unvalidated assumption in the ABA selection rule that is load-bearing for the accuracy claims on the most severe shifts.","major_comments":[{"comment":"The claimed \"13.3%p accuracy improvement over state-of-the-art baselines\" is not defined and is inconsistent with the detailed tables: in Table 1, BiTTA's margin over the strongest binary-feedback baseline is +3.42 pp on CIFAR10-C, +1.79 pp on CIFAR100-C, and +1.03 pp on Tiny-ImageNet-C, while in Table 2 the margin is +12.94 pp on PACS domain-wise and +2.97 pp on PACS mixed; no averaging protocol in the text produces 13.3 pp. Please specify the baseline, the set of benchmarks, and the aggregation rule, and update the abstract accordingly.","section":"Abstract / Section 4"},{"comment":"The REINFORCE estimator in Eq. (3) is derived for actions sampled from π_θ, but Algorithm 1 computes the gradient for the deterministic prediction y* = argmax f_θ(x); this is a biased estimator of ∇J and the connection between Eq. (3) and the implemented cross-entropy loss in Eq. (12) is not established. Please either sample actions from the MC-dropout policy during training or reframe the objective as a weighted self-training loss and remove the policy-gradient claim.","section":"Section 3, Eq. (3), Algorithm 1, Eq. (12)"},{"comment":"The ABA selection rule assumes that agreement between the deterministic prediction and the MC-dropout argmax implies the pseudo-label is correct; the validation for this is only Figure 4b on CIFAR10-C, where source accuracy is already 57%. On CIFAR100-C and Tiny-ImageNet-C, source accuracy is 33% and 21%, where a model can be consistently wrong across dropout masks, so agreement may select confidently wrong targets; since ABA-only accounts for 82.64% of BiTTA's 87.20% on CIFAR10-C, a failure of this assumption would substantially erode the headline accuracy. Please report the accuracy of S_ABA versus its complement per corruption type for CIFAR100-C and Tiny-ImageNet-C, and show the online accuracy trajectory of ABA-selected samples to demonstrate the assumption holds under severe shifts.","section":"Section 3.2, Eq. (8)-(9), Section 4 ablation"},{"comment":"The main balancing hyperparameters are stated inconsistently: Section 3.3 says \"We set α=β=1 in all experiments\", Section 4 says \"α=2 and β=1\", and the Figure 10 caption says α is set to 1. This ambiguity directly affects the reported objective and the sensitivity analysis; please correct the statements and ensure the figure matches the main configuration.","section":"Section 3.3 and Section 4"},{"comment":"The \"equal labeling cost\" comparison that supports the claim \"up to 32% higher accuracy\" is not fully specified: Figure 5 displays \"+9%, +30%, +32%\" without defining whether these are relative or percentage-point improvements, and the number of feedback samples used by each method under the Shannon-cost rule is described only in the appendix; clarify the exact budgets (e.g., 3 binary vs 1 full-class sample for CIFAR-10) and state the error bars and significance of the differences.","section":"Section 4, Figure 5, Appendix D.1"}],"minor_comments":[{"comment":"Figures 1, 5, and several Appendix B figures contain garbled unicode characters in their captions (e.g., \"/uni00000028/uni00000057/...\" in Figure 5), which appear to be a rendering artifact; the figures and captions should be regenerated.","section":"Figures 1, 5, and Appendix B"},{"comment":"Figure 2's x-axis labels are concatenated as \"TENT*EATA*CoTTA*RoTTA*SoTTA*BiTTA\"; please use separate labels and a legend.","section":"Figure 2"},{"comment":"In Eq. (10), the ABA term β E_{x∈S_ABA}[R_ABA(x,y)] is constant (=β) since R_ABA=1 for all x∈S_ABA by definition; clarify that the actual update uses the gradient E[∇ log π] rather than the literal objective.","section":"Eq. (10)"},{"comment":"Table 1 and several other tables report only averages over three seeds without standard deviations; given that Table 2 reports standard deviations and some baselines (e.g., SimATTA* on PACS) have very high variance, error bars should be reported consistently.","section":"Tables 1 and others"},{"comment":"The description of the binary-feedback baseline modification says \"an equal number of random binary-feedback data\" but does not state the selection rule (random vs uncertainty) or the exact loss weighting for each adapted baseline; please specify the full protocol for reproducibility.","section":"Section 4, baseline adaptations"},{"comment":"The statement that BiTTA uses \"three binary feedback samples within each 64-sample test batch, accounting for less than 5%\" is correct, but the later claim of \"log(num_class) times more feedback samples\" under equal labeling cost should be made more explicit with the actual numbers per dataset.","section":"Appendix D.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely within scope for a test-time adaptation or active learning venue, but the unresolved quantitative claim and the ABA validity concern require careful revision. If the authors provide the requested per-dataset analysis of agreement-selected samples and clarify the RL derivation, the result would be of interest. I do not see a circularity problem; the reuse of SoTTA/AETTA weights and designs is acknowledged through citations."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The setting is the real contribution here: test-time adaptation with a few correct/incorrect feedback labels is a practical middle ground between full-class active TTA and fully unsupervised TTA, and I don't think anyone has framed it quite this way before. BiTTA's dual-path structure (BFA on uncertain samples, ABA on confident ones) is a reasonable design, and the experimental coverage is broad: three corruption benchmarks, PACS, extra datasets, ablations, sensitivity analyses, and released code. The result that binary feedback can beat full-class SimATTA under equal labeling cost is notable if it holds up.\n\nThat said, the paper as written has several soft spots. The abstract's \"13.3%p accuracy improvement\" is never defined and doesn't match the tables: against the best binary-feedback baseline the margins are roughly 3.4 pp on CIFAR10-C, 1.8 pp on CIFAR100-C, 1.0 pp on Tiny-ImageNet-C, and 11.2 pp on PACS, averaging around 4–5 pp, not 13.3. The authors also state alpha=beta=1 in Section 3.3, alpha=2/beta=1 in Section 4, and alpha=1 in Figure 10's caption; one of these is wrong. The REINFORCE framing is loose: Eq. 3 assumes actions sampled from pi, but the algorithm uses the deterministic argmax and then optimizes cross-entropy on MC-dropout probabilities. The actual loss is pseudo-labeling with a confidence gate, which is fine, but the RL story oversells it.\n\nThe stress-test concern about ABA is legitimate. Agreement between the deterministic and MC-dropout predictions does not guarantee correctness on severe shifts; Figure 4b only validates this on CIFAR10-C, where source accuracy is already decent. On Tiny-ImageNet-C the margin over the best baseline is only about one point, and if ABA reinforces confidently wrong targets there, the overall claim of consistent SOTA would not survive. The paper does not report agreement-sample accuracy on the harder benchmarks, so this is an unaddressed risk rather than a demonstrated failure. Missing error bars in Table 1 are a minor issue since Table 12 reports them.\n\nOverall, the central finding—that a few binary labels can improve TTA over no labels and over full-class active TTA—is plausible and worth taking seriously. The paper needs revision to fix the numerical inconsistencies, tighten the RL framing, and add per-dataset evidence on ABA's reliability under severe shift. I would send it to peer review; it deserves referee time, but as a conditional accept, not a clean one.","headline":"A genuinely new TTA-with-binary-feedback setting and a plausible method, but the headline accuracy number is undefined, there are internal inconsistencies, and the agreement-based self-training path is under-validated on severe shifts.","tokens_in":32109,"tokens_out":2253,"would_cite":true,"duration_ms":20744,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Binary correct/incorrect feedback on a few uncertain samples, paired with self-adaptation on confident ones, can guide test-time adaptation past severe distribution shift and even beat full-class active labeling at equal labeling cost.","keywords":["test-time adaptation","binary feedback","reinforcement learning","MC-dropout","active learning","domain shift","self-adaptation"],"falsifier":"Run BiTTA with BFA disabled (ABA-only) on a corruption type or severity not in the paper's suite and measure whether agreement-sample accuracy stays high. If on some shift the agreement samples' accuracy falls near chance while ABA-only accuracy drops below BFA-only accuracy, the self-adaptation path is amplifying noise rather than knowledge. Concretely, report the accuracy of the agreement set $S_{\\text{ABA}}$ on ImageNet-C or a new severity-5 corruption and the delta between BiTTA and its BFA-only ablation.","tokens_in":31060,"feed_emoji":"👍","tokens_out":4347,"duration_ms":35009,"temperature":0.7,"pith_summary":"Test-time adaptation usually tunes a pretrained model on unlabeled test data, which can fail under severe shifts. This paper proposes a cheaper alternative: ask a human only whether the model's prediction is correct or incorrect, a single bit, on a handful of uncertain samples per batch. The paper's method, BiTTA, combines this binary feedback with self-training on confidently predicted unlabeled samples, both optimized through a reinforcement-learning objective. On corruption and domain-shift benchmarks, BiTTA reports an average 13.3 percentage-point accuracy gain over state-of-the-art baselines, and it outperforms an active-test-time-adaptation method that receives full class labels under equal labeling cost. If it holds, the result means a thumb up or down on a few predictions can substitute for expensive full labels during deployment.","feed_headline":"Three binary labels per batch beat full-class active TTA","feed_subtitle":"BiTTA needs only correct/incorrect feedback yet tops SimATTA's full labels on corruption benchmarks.","key_machinery":"The load-bearing object is the MC-dropout policy estimate $\\pi_\\theta(y|x)$, the mean softmax over $N$ stochastic forward passes, combined with REINFORCE policy gradients. MC-dropout does double duty: it selects the $k$ least-confident samples for binary feedback (BFA) and identifies agreement samples for self-adaptation (ABA). The combined loss is $L_{\\text{BiTTA}} = \\alpha L_{\\text{BFA}} + \\beta L_{\\text{ABA}}$ with $L_{\\text{BFA}}$ minimizing cross-entropy on correct-memory samples and maximizing it on incorrect-memory samples, and $L_{\\text{ABA}}$ minimizing cross-entropy only on agreement samples. The mechanism is what lets one bit of feedback generalize: the oracle bit is sparse but direct, and the agreement path supplies the dense, unlabeled self-supervision.","core_discovery":"The central claim is that binary correct/incorrect feedback, used sparingly, is enough to guide test-time adaptation past severe distribution shift, and that it can beat full-class active labeling once labeling cost is accounted for. BiTTA formulates adaptation as policy-gradient reinforcement learning: the model's MC-dropout softmax is the policy, the predicted label is the action, and rewards come from two complementary paths. Uncertain samples (lowest MC-dropout confidence in the predicted class) receive ±1 rewards from the oracle and are used to push the model toward correct predictions and away from incorrect ones; confident samples, identified by agreement between deterministic and MC-dropout predictions, receive a +1 self-reward for maintaining their current prediction. On CIFAR10-C the method reaches 87.20% accuracy with 3 labels per 64-sample batch, versus 83.78% for the best binary-feedback-augmented TTA baseline and 83.13% for full-class SimATTA under equal sample count, and it also outperforms SimATTA under equal labeling cost.","pith_inferences":["The binary-feedback setting could be extended to non-classification tasks such as segmentation or detection, where correct/incorrect is defined per prediction; the RL formulation only requires a scalar reward.","The equal-labeling-cost comparison suggests the binding constraint in active TTA may be the number of human interactions rather than the information content per label, so methods that use cheap labels to query more samples could dominate regardless of label richness.","A testable extension is to replace the oracle with a learned reward model calibrated on a handful of binary labels and measure how BiTTA's margin shrinks as reward noise increases."],"forward_implications":["If the central claim holds, a deployment pipeline can request only yes/no feedback from operators or users, cutting annotation cost by a factor of roughly $\\log(\\text{num\\_classes})$ per label compared with full-class labeling.","BiTTA's dual-path objective means the method does not have to choose between active labeling and unlabeled self-training; both share one policy-gradient update.","The agreement-based selection criterion is dynamic and threshold-free, so it can track confidence drift across a continual stream without re-tuning a confidence threshold.","Because MC-dropout provides the uncertainty estimate with no architectural change, the approach transfers to standard pretrained backbones."],"supporting_citations":[{"why":"Supplies the MC-dropout uncertainty estimates that drive both BFA sample selection and ABA agreement detection.","marker":"(Gal & Ghahramani, 2016)"},{"why":"Provides the REINFORCE gradient estimator that turns non-differentiable binary rewards into parameter updates.","marker":"(Williams, 1992)"},{"why":"Defines the active TTA baseline SimATTA that BiTTA must beat, and supplies the full-label comparison setup.","marker":"(Gui et al., 2024)"},{"why":"Supplies the CIFAR-10-C, CIFAR-100-C, and Tiny-ImageNet-C corruption benchmarks used for evaluation.","marker":"(Hendrycks & Dietterich, 2019)"},{"why":"Provides empirical evidence that binary feedback is far cheaper and more reliable than full-class labeling, motivating the setting.","marker":"(Joshi et al., 2010)"},{"why":"Introduces the test-time adaptation paradigm that BiTTA extends with binary feedback.","marker":"(Wang et al., 2021)"}],"fun_headline_variants":["Binary yes/no feedback outperforms full-class active TTA","With only correct/incorrect labels, BiTTA beats full-labeling TTA","BiTTA's three binary labels per batch top full-class active adaptation","Minimal labeling: correct/wrong feedback wins over full-class TTA"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a sample whose deterministic prediction matches its MC-dropout prediction is confidently correct enough to serve as a self-training target; if that agreement no longer tracks accuracy on an unseen shift, the ABA path could reinforce the model's errors.","fun_headline_variants_meta":{"raw":{"variants":["Binary yes/no feedback outperforms full-class active TTA","With only correct/incorrect labels, BiTTA beats full-labeling TTA","BiTTA's three binary labels per batch top full-class active adaptation","Minimal labeling: correct/wrong feedback wins over full-class TTA"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000315,"raw_usage":{"total_tokens":1782,"prompt_tokens":936,"completion_tokens":846,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":767}},"tokens_in":552,"tokens_out":846,"duration_ms":7465,"temperature":1.0,"reasoning_tokens":767,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:30:21.218760+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BiTTA with BFA disabled (ABA-only) on a corruption type or severity not in the paper's suite and measure whether agreement-sample accuracy stays high. If on some shift the agreement samples' accuracy falls near chance while ABA-only accuracy drops below BFA-only accuracy, the self-adaptation path is amplifying noise rather than knowledge. Concretely, report the accuracy of the agreement set $S_{\\text{ABA}}$ on ImageNet-C or a new severity-5 corruption and the delta between BiTTA and its BFA-only ablation.","supporting_citations":[],"review_version":1}