{"id":"f25936e2-0392-4ad2-bbfc-db1e82503d9a","arxiv_id":"2506.02843","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Random registers in a vision transformer improve cross-domain few-shot transfer, and REAP strengthens this by replacing clustered image patches with random noise.","lead":"This paper finds that, for cross-domain few-shot learning, learnable prompt tokens hurt a vision transformer's transfer to new domains, while random noise tokens improve it. The authors propose REAP, which adds random registers to clustered image patches during source training, and report state-of-the-art results on four target datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 7's SAM equivalence drops SAM's worst-case maximization, so the central claim that random registers implement sharpness-aware minimization is unsupported.","rationale":"The paper's empirical phenomenon appears real: the ablations with error bars (Table 2) show a consistent improvement from random registers, and the REAP method is simple and reproducible in spirit. I do not object to the empirical SOTA claim as a conditional result; the reader's caveats about missing error bars in Table 1 and target-benchmark hyperparameter selection already make the verdict conditional. My concern is with the explanatory layer: the paper claims random registers are 'a kind of SAM' because they perturb attention maps. The derivation in Eqs. 5–7 conflates SAM's norm-constrained worst-case maximization with a fixed random draw, and the same derivation omits the value-token contribution of register tokens. This means the central mechanistic contribution—that random registers implement sharpness-aware minimization and thereby find a flattened minimum—is not supported by the presented equations. The paper demonstrates correlation and offers an analogy, not a derivation. Because this concern is exactly the reader's weakest assumption, my read does not change the conditional verdict. I recommend UNCHANGED, with the condition that the authors either provide a direct test of gradient alignment with the register-induced perturbation or reframe the contribution as noise-based regularization rather than an SAM equivalence.","tokens_in":21305,"tokens_out":10312,"duration_ms":130807,"concrete_test":"During source training, for about 100 fixed batches, compute the gradient ∂L/∂A of the classification loss with respect to the pre-softmax attention logits at a layer where random registers are active. Re-sample random register keys several times and measure the induced logit change ΔA. Compute the mean cosine similarity between ΔA and ∂L/∂A with a 95% confidence interval. SAM's perturbation is, by construction, the normalized gradient direction; if the mean cosine is near zero (e.g., |cos| < 0.1), random-register noise is not the SAM maximizer and Eq. 7's identification is unsupported. A consistently positive cosine would support a weak SAM-like effect. A complementary check is to zero out the value projections of register tokens; if the target-domain gain disappears, the mechanism is not purely attention perturbation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.3 (Eqs. 5–7) argues that random registers implement sharpness-aware minimization, but the argument is an analogy, not a derivation. In SAM, ε is the maximizer of L(ω+ε) over ||ε||₂≤ρ; in Eq. 6 the register term is a fixed random draw from random keys, never the computed, gradient-aligned worst case. Eq. 7 writes max_ε L(A+ε_R): if ε_R is fixed, the max operator has nothing to optimize, and if ε_R is meant to vary with ε, the notation is undefined. The text itself says 'can be regarded as' and 'viewed as,' yet the paper's contribution (2) states as fact that random registers 'novelty perturb attention maps for sharpness-aware minimization,' so the contribution rests on this identification. Moreover, Eq. 6 only shows the softmax denominator; appended register tokens also contribute through their value vectors to the output, so the perturbation is not confined to attention maps. If the SAM analogy fails, random registers may simply be stochastic input/feature noise, a regularization effect, and the flat-minimum explanation in Fig. 3b becomes correlational rather than causal. The empirical gains could still be real, but the central mechanistic claim and the motivation for REAP would be unproven.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies cross-domain few-shot learning (CDFSL) with ViT backbones and reports an empirical phenomenon: during source-domain training, learnable prompt tokens (registers) hurt target-domain transfer, while replacing them with random Gaussian registers consistently improves target accuracy, especially as the number of random registers grows. The authors interpret this via attention sharpness and CKA domain similarity, proposing that learnable registers absorb source-specific domain information and that random registers act as a form of sharpness-aware minimization (SAM) by perturbing attention maps. Based on this interpretation, they propose REAP, which, during source training, randomly clusters image patches and replaces selected clusters with random registers plus a small number of additional random registers, and, during target finetuning, switches to learnable registers. Experiments on four CDFSL benchmarks report state-of-the-art results under several settings, with ablations over the main components and additional results on multiple backbones.","tokens_in":21611,"tokens_out":5630,"duration_ms":67345,"significance":"If the empirical claims hold, the paper makes a useful and practical contribution: a cheap, architecture-agnostic way to improve ViT transfer under large domain gaps, with code and models released. The paper is also careful in several ways: it reports ablations of the main components, compares against many prior methods, groups comparisons by finetuning/transductive settings, and evaluates on multiple backbones. The sharpness and CKA analyses are valuable diagnostic evidence. However, the central mechanistic claim that random registers implement sharpness-aware minimization is not established by the formal argument in Section 2.3; as written, the SAM connection is an analogy rather than a derivation. Because this interpretation is listed as a contribution and motivates REAP, the paper needs either direct evidence for the SAM mechanism or a substantially softened framing. The empirical gains themselves are plausible and well supported, but the current contribution statement overstates the theoretical support.","major_comments":[{"comment":"The formal identification of random registers with SAM is not valid as stated. In Eq. (5), SAM maximizes L(ω+ε) over an adversarially chosen perturbation with ‖ε‖₂≤ρ, whereas in Eq. (7) the perturbation ε_R is a fixed random draw from the random registers. If ε_R is fixed, the max_ε operator is vacuous; if ε_R is intended to vary with ε, the notation is undefined. The text hedges with 'can be regarded as' and 'viewed as', but contribution (2) and Section 2.5 assert that random registers 'novelty perturb attention maps for sharpness-aware minimization,' so the central mechanistic claim rests on this identification. In addition, Eq. (6) shows only the softmax denominator; appended registers also contribute through their value vectors to the attention output, so the perturbation is not confined to attention maps. Please either reframe the SAM connection as a hypothesis and explicitly remove it from the claimed contributions, or provide direct evidence: for example, measure the alignment between register-induced noise and the SAM ascent direction, compare training trajectories against actual SAM, or show that random-register training reduces the sharpness measure in Eq. (4) more than an equivalent amount of input or feature noise. Without such evidence, the flat-minimum explanation remains correlational.","section":null},{"comment":"The sharpness evidence is post hoc and does not by itself establish that random registers perform sharpness-aware minimization. Eq. (4) measures the sensitivity of a trained model to attention perturbations, and Fig. 3b shows that the random-register model has lower sharpness than the learnable-register model. This is consistent with the flat-minimum interpretation, but it could also arise from stochastic input/feature regularization, which is a distinct mechanism. The causal claim that training with random registers is 'a kind of SAM' requires a training-time comparison: for example, compare the loss-landscape sharpness reached by random-register training with that reached by actual SAM, and show that random-register noise is not merely equivalent to adding Gaussian noise to image features or weights. Without this, Sections 2.5 and 3.1 should be rephrased from 'is' to 'is hypothesized to be'.","section":null},{"comment":"The starred REAP rows are placed in the transductive group, but the method description in Section 3 contains no transductive component: target finetuning is performed on the support set only, and no use of unlabeled query data is described. The meaning of REAP* must be stated explicitly. If it uses the query set or any additional unlabeled data, this changes the experimental setup and should be detailed; if it does not, the asterisk notation is misleading and should be removed or redefined. This matters because the SOTA claim in Table 1 relies on correctly grouping methods with and without transductive access.","section":null}],"minor_comments":[{"comment":"Please specify how the max is computed over Gaussian perturbations: the number of sampled perturbations, the variance schedule, and whether the max is taken over a finite sample or over the distribution. Without this, the sharpness values in Fig. 3b are not reproducible.","section":null},{"comment":"The displayed equation has unbalanced parentheses, making the argument of Lcls ambiguous. Please rewrite the equation with all parentheses matched and the variables clearly separated.","section":null},{"comment":"The notation in Eq. (8) introduces m, n−m, and ñ without first defining m in the surrounding text; the relationship between the anchor ratio a, the cluster size, and m should be stated before the equation. Also, when clusters are replaced by random tokens, the 'image perturbation' terms use the same notation as the original keys, which is confusing.","section":null},{"comment":"The column heading 'Mark' is unexplained; it appears to denote venue and year, but the caption should state this explicitly.","section":null},{"comment":"There are several typos and grammar issues, including 'as is explained as sharpness-aware minimization' (§2.4), 'By multiplied with randomized keys' (§3.1), and 'can be regarded asa novel way' (§2.3). A careful proofread is needed.","section":null}],"recommendation":"major_revision","confidential_remarks":"The paper makes heavy use of the authors' own prior works (Zou et al., a;b; FLoR) for baselines and sharpness tools; this is legitimate and not circular, but the reviewers should verify that the concurrent NeurIPS-24 citations were indeed available at submission. No evidence of citation suppression or data fabrication is apparent. The main empirical contribution is likely sound, but the SAM-equivalence claim needs either direct evidence or a clear downgrade to hypothesis status before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things you should know about arXiv:2506.02843. First, the empirical finding is real: for ViT-based CDFSL, learnable registers (prompt tokens) trained on the source domain hurt target-domain performance, while random Gaussian registers consistently help. Second, the proposed REAP method—cluster image patches and replace whole clusters with random registers during source training, then finetune with learnable registers on the target—beats prior SOTA on the four standard benchmarks. The gains are modest, though: 1-shot no-finetuning average 54.89 vs 54.12, and 5-shot finetuning 68.03 vs 67.61. The main comparison table has no error bars, so the SOTA claim rests on thin margins.\n\nWhat the paper does well: it documents the phenomenon with ablations, attention visualizations, CKA similarity, and a sensitivity study. It fairly credits Darcet et al.'s registers and the SAM/FLoR line. The cluster-based replacement is new and clearly ablated: random masking hurts, cluster masking alone is neutral, and adding the random-register replacement is what helps. Code and models are released.\n\nThe soft spot is Section 2.3. The claim that random registers are a form of sharpness-aware minimization is an analogy, not a derivation. In SAM, epsilon is the norm-constrained worst-case perturbation; here it is a fixed random draw from random keys. Equation 7 writes max_epsilon L(A + epsilon_R), but the max operator has nothing to optimize. The text hedges with 'can be regarded as' and 'viewed as,' but the contribution statement asserts the SAM identification as fact. If the analogy fails, the flat-minimum explanation is unsupported, and random registers could simply be stochastic feature noise with a regularizing effect. The empirical gains would still stand, but the paper should present the SAM link as an interpretation, not a justification. Minor point: Eq. 6 only shows the softmax denominator; the register value vectors also affect the output, so the perturbation is not confined to attention maps.\n\nThe citation pattern is honest. Self-citations provide baselines and sharpness tools rather than presupposing the result.\n\nThis paper deserves a serious referee. The phenomenon is new, the method is simple and reproducible, and the flaws are addressable. A revision that softens the SAM claim, adds error bars to the main comparison, and reports hyperparameter selection would be solid. I'd send it to review.","headline":"A genuinely interesting empirical finding with a simple effective method, but the SAM-equivalence claim is an analogy that should be reframed, and the SOTA margins are thin.","tokens_in":22134,"tokens_out":3042,"would_cite":false,"duration_ms":29342,"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":"Random registers — Gaussian-noise tokens appended to a vision transformer — improve cross-domain few-shot transfer while learned prompts hurt it, because the noise perturbs attention maps like sharpness-aware minimization.","keywords":["cross-domain few-shot learning","vision transformer","prompt tuning","random registers","sharpness-aware minimization","attention perturbation","domain similarity","CKA"],"falsifier":"Train the same ViT under the same protocol four ways — learned registers, random registers, a fixed constant token of the same dimension, and true SAM applied to attention or weights — then compare target-domain accuracy and a direct curvature measurement such as the sharpness metric of Eq. 4. If the fixed constant token matches the random register's gains, randomness itself is not the mechanism; if true SAM does not reproduce the same flattening and transfer gains under a matched perturbation budget, the sharpness-aware-minimization interpretation is unsupported.","tokens_in":21142,"feed_emoji":"🎲","tokens_out":11438,"duration_ms":90664,"temperature":0.7,"pith_summary":"This paper is trying to establish a counterintuitive result about how vision transformers should be tuned for cross-domain few-shot learning: learnable prompts, the standard way to adapt a ViT to a source dataset, actually hurt the model when it is evaluated on a different target domain, while replacing those prompts with tokens of random Gaussian noise consistently improves target-domain accuracy. The authors' explanation is that learned prompts absorb source-domain-specific information, causing the model to treat irrelevant visual patterns as recognition cues and making the loss landscape sharper, whereas random registers inject noise into the attention computation that acts like a sharpness-aware minimization and steers the model to flatter minima with more domain-agnostic features. On this basis the paper proposes REAP, which replaces clusters of similar image patches with random registers to make the perturbation stronger and more efficient, and then switches to learnable registers during target-domain finetuning. If the claim is right, random-register perturbation is a nearly free way to make ViT features transfer across large domain gaps, without the extra forward-backward cost of true sharpness-aware minimization.","feed_headline":"Random tokens beat learned prompts for cross-domain few-shot learning","feed_subtitle":"Replacing learned prompts with random noise flattens ViT's loss landscape and lifts target-domain accuracy.","key_machinery":"The load-bearing object is the random register: a token of Gaussian noise, $T_R \\sim \\mathcal{N}(0, \\tau^2)$ with a learnable scale $\\tau$, appended to the ViT input token sequence. Inside the attention softmax its key vectors generate the random denominator term of Eq. 6, which the paper treats as the perturbation $\\epsilon_R$ of a sharpness-aware minimization objective, the mechanism that supposedly flattens the loss landscape and pushes the model toward domain-agnostic features. The second component is cluster replacement: image patches are grouped by cosine similarity into clusters and a large portion of the clustered patches is replaced with random registers, so the perturbation strikes the semantic regions that dominate attention instead of being diluted across the whole image (Eq. 8). The third component is a two-stage use of the same tokens, random during source training and switched to learnable during target finetuning, so that the registers absorb target-domain information in the few-shot episode.","core_discovery":"The central claim is that random registers are not merely a harmless substitute for learned prompts but a transfer mechanism in their own right. Writing the attention softmax explicitly as in Eq. 6 of the paper, each query token multiplies random key vectors $\\tilde{K}$ produced by the registers, so the denominator term $\\sum_{k=1}^{\\tilde{n}} e^{Q_i \\tilde{K}_k^\\top}$ is a random noise $\\epsilon_R$ added to the attention map; the paper rewrites sharpness-aware minimization as $\\min_\\omega \\max_\\epsilon L(A + \\epsilon_R)$ and identifies random registers with this SAM-style perturbation. The supporting evidence is threefold: measured sharpness of the attention loss landscape rises with learnable registers and falls with random registers; CKA similarity between source and target features falls with learnable registers and rises with random registers; and attention maps on target domains drift off the object with learnable registers but stay on it with random registers. The paper further claims that because ViT attention is driven by continuous semantic regions, replacing clustered image patches with random registers amplifies the perturbation so that 16 registers suffice where hundreds were needed, and that switching the registers to the learnable state during target finetuning exploits their tendency to absorb domain-specific information. The claimed outcome is the top average accuracy on four target benchmarks (ChestX, ISIC2018, EuroSAT, CropDiseases) under 1-shot and 5-shot settings, with and without finetuning.","pith_inferences":["My inference: the same mechanism should transfer to domain generalization and unsupervised domain adaptation, where the failure mode — attention locked onto source-specific patterns — is identical; a cheap check is to run random-register perturbation on standard domain-generalization benchmarks.","My inference: because the SAM identification drops the worst-case maximization, part of the gain may be plain stochastic regularization; ablating the noise distribution (Gaussian versus Bernoulli drop versus sign-flip) would reveal which property of randomness carries the effect.","My inference: the cluster-replacement step is close in spirit to masked image modeling, so the method could double as a pretraining trick; a testable extension is whether the clustering constraint (semantic continuity) matters more than the masking ratio.","My inference: if the flat-minimum mechanism is real, random registers should also improve robustness to corruption-style shifts such as ImageNet-C, which would be a fast and independent way to test the sharpness claim."],"forward_implications":["Learned prompts should be dropped from the source-stage recipe for ViT-based cross-domain few-shot learning; random registers are a drop-in replacement that improves transfer without the extra optimization of true SAM.","The dual-role recipe — random registers at source training, learnable registers at target finetuning — becomes a usable default for few-shot adaptation of ViT backbones.","Cluster-based replacement of image patches makes the attention perturbation strong enough that only 16 appended registers are needed, against roughly a thousand for naive random registers.","The gains replicate across backbone initializations (DINO ViT-S, DINO ViT-B, iBOT, CLIP) and across shallow and deep register placements, so the effect is not tied to one architecture configuration.","With finetuning, REAP achieves the top average accuracy on all four target datasets in both 1-shot and 5-shot evaluations."],"supporting_citations":[{"why":"Supplies the SAM objective and generalization bound that the paper uses to interpret random-register noise as sharpness-aware minimization.","marker":"(Foret et al., 2021)"},{"why":"Introduces the term 'registers' for extra ViT input tokens, which the paper adopts and then randomizes.","marker":"(Darcet et al., 2024)"},{"why":"Defines visual prompt tuning, the learnable-prompt training paradigm whose harmful transfer effect the paper documents.","marker":"(Jia et al., 2022)"},{"why":"Provides the ViT-with-DINO-pretraining backbone and source-training setup that REAP builds on and compares against.","marker":"(Fu et al., 2023)"},{"why":"Supplies the CKA similarity used to measure domain similarity between source and target features.","marker":"(Kornblith et al., 2019)"},{"why":"Shows that ViT is robust to patch perturbation, the premise used to justify cluster-based image perturbation.","marker":"(Naseer et al., 2021)"},{"why":"Motivates the clustering design by showing that ViT attention relies on continuous image regions.","marker":"(Wei et al., 2024)"},{"why":"Defines the prototypical-network evaluation protocol used for target-domain few-shot testing.","marker":"(Snell et al., 2017)"}],"fun_headline_variants":["Random registers flatten ViT loss for cross-domain few-shot","Replacing prompts with noise boosts ViT transfer in few-shot","ViT few-shot: random registers beat learned prompts","Random registers: SAM-style perturbation for ViT few-shot","For ViT few-shot, random registers outperform tuned prompts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The identification of random registers with sharpness-aware minimization is an analogy: SAM maximizes the loss over a worst-case norm-constrained perturbation, while a random register is a single fixed draw of noise, so the paper's Eq. 7 drops the maximization step and the flat-minimum explanation rests on the assumption that the random draw behaves like the maximized perturbation.","fun_headline_variants_meta":{"raw":{"variants":["Random registers flatten ViT loss for cross-domain few-shot","Replacing prompts with noise boosts ViT transfer in few-shot","ViT few-shot: random registers beat learned prompts","Random registers: SAM-style perturbation for ViT few-shot","For ViT few-shot, random registers outperform tuned prompts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1702,"prompt_tokens":1114,"completion_tokens":588,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":730,"completion_tokens_details":{"reasoning_tokens":507}},"tokens_in":730,"tokens_out":588,"duration_ms":5908,"temperature":1.0,"reasoning_tokens":507,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:14:32.307606+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same ViT under the same protocol four ways — learned registers, random registers, a fixed constant token of the same dimension, and true SAM applied to attention or weights — then compare target-domain accuracy and a direct curvature measurement such as the sharpness metric of Eq. 4. If the fixed constant token matches the random register's gains, randomness itself is not the mechanism; if true SAM does not reproduce the same flattening and transfer gains under a matched perturbation budget, the sharpness-aware-minimization interpretation is unsupported.","supporting_citations":[{"cited_title":"Sharpness-aware minimization for efficiently improving generalization, 2021","cited_arxiv_id":null,"evidence_quote":"Supplies the SAM objective and generalization bound that the paper uses to interpret random-register noise as sharpness-aware minimization."},{"cited_title":"Visual prompt tuning","cited_arxiv_id":null,"evidence_quote":"Defines visual prompt tuning, the learnable-prompt training paradigm whose harmful transfer effect the paper documents."},{"cited_title":"Efficient vision-language pre-training by cluster masking","cited_arxiv_id":null,"evidence_quote":"Motivates the clustering design by showing that ViT attention relies on continuous image regions."},{"cited_title":"Prototypical networks for few-shot learning","cited_arxiv_id":null,"evidence_quote":"Defines the prototypical-network evaluation protocol used for target-domain few-shot testing."}],"review_version":1}