{"id":"e52271ec-4ae4-441b-a584-aeb0cdb79e97","arxiv_id":"2412.03058","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Simple transformed copies of in-distribution images, called peripheral-distribution samples, can replace real outliers when training an energy-based out-of-distribution detector.","lead":"This paper proposes tuning a trained image classifier with simple transformations of its own training images, such as blur, rotation, and cutout, to better flag unfamiliar inputs without needing curated outlier data. The authors report top out-of-distribution detection scores on CIFAR-10 and CIFAR-100 with only a 10-epoch fine-tune.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (12)'s energy-barrier loss is sign-inverted: minimizing log σ((E(xper)-E(xin))/β) drives PD energy below ID energy, contradicting Assumption 1 and the claimed mechanism; the central OEST* results cannot follow from the printed objective.","rationale":"The reader's weakest_assumption focuses on Assumption 1 and its admitted failure in Sec. V-C4, but the reader's rationale separately identifies the sign inversion in Eq. (12). I agree that Assumption 1 is a serious limitation: Theorem 1 only provides a guarantee when every OOD instance has a simple-transformation proxy x+ close in feature space with a large energy gap, and the authors explicitly document a main-setting failure on CIFAR-100/ResNet-18 (Table VII: OEST* is 3.83 AUROC worse than EBO). However, the sign error is the single most load-bearing issue for the paper's central claim. The objective in Eq. (12) is not a harmless typo because the whole contribution of OEST* is the energy-barrier loss; as printed, minimizing it drives E(xper) below E(xin), the reverse of the barrier required by Assumption 1 and illustrated in Fig. 1b. The reported numbers in Tables II-III are claimed for OEST*, and the paper provides no derivation that would make the printed objective equivalent to the intended barrier, e.g. via a negative sign or a swapped numerator. Therefore the current manuscript cannot be judged correct until either the text or the code is fixed and the experiments are re-run. This warrants rejection of the submission as written, not merely conditional acceptance, because the discrepancy affects the central contribution and the authors' own released code is the only way to resolve it. Should the code check show a one-line sign typo and reproduce the results with the corrected sign, the appropriate path would be a corrected resubmission; my verdict refers to the manuscript in its current form. I also note the proof of Theorem 1 is essentially a Lipschitz bound and appears valid apart from minor presentation; the weakness is in the assumption, not the algebra, which is why I do not escalate that point beyond the reader's existing objection.","tokens_in":27349,"tokens_out":12029,"duration_ms":113885,"concrete_test":"Clone the public repository (github.com/victor-yifanwu/Outlier-Exposure-by-Simple-Transformations), locate the Lenergy* computation, and check the sign of the logit passed to log-sigmoid against Eq. (12). Then retrain CIFAR-10/ResNet-18 under the Sec. V-A3 setup (10 epochs, α=0.2, β=10) in two variants: (a) exactly as printed, and (b) with the argument flipped so minimization pushes E(xper)-E(xin) upward. If variant (b) reproduces the average AUROC 96.30 while variant (a) does not, the printed equation is wrong and the paper's central mechanism is mis-specified; if variant (a) reproduces the table, the text and code disagree and the manuscript must be corrected and re-evaluated either way.","verdict_should_be":"REJECT","load_bearing_attack":"The printed OEST* objective cannot implement the claimed energy barrier. Eq. (12) defines Lenergy*(xin,x'in) = log σ((E(xper)-E(xin))/β) and Algorithm 1 minimizes L = LCE + α·Lenergy*. Since log σ(z) is monotone increasing with derivative 1-σ(z) > 0 and unbounded below as z → -∞, minimizing the loss drives (E(xper)-E(xin))/β as negative as possible, i.e. E(xper) < E(xin). This is the exact opposite of the barrier required by Eq. (9)/Assumption 1, E(x+)-E(x) > γα ≥ 0, and of the mechanism described in the abstract and Fig. 1b. The reported SOTA AUROC values in Tables II-III are therefore not consequences of the stated method; either the released code contains a sign flip absent from the text, or the empirical improvement has a different cause. This is more immediately disabling than the admitted failure of Assumption 1 for CIFAR-100/ResNet-18 (Sec. V-C4), because it puts the theoretical and algorithmic core in direct contradiction with the experiments as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two OOD-detection tuning schemes, OEST and OEST*, which fine-tune a pretrained classifier using cross-entropy plus an energy-based loss on \"peripheral-distribution\" (PD) data obtained by simple transformations of in-distribution samples. The central theoretical claim is that if an energy barrier exists between PD and ID samples (Assumption 1), then OOD samples receive higher energy than ID samples with high probability (Theorem 1). The authors report strong AUROC and FPR95 numbers on CIFAR-10 and CIFAR-100 benchmarks, as well as on MNIST/SVHN, and provide a public code repository. The paper is clearly written and the experimental comparison is broad, but the theoretical and algorithmic core contains a load-bearing sign error in the OEST* objective, and the theoretical result is conditional on essentially the same energy-barrier behavior that the loss is intended to create.","tokens_in":27674,"tokens_out":4397,"duration_ms":48602,"significance":"If the stated mechanism were correct, the idea of replacing manually curated outlier data with a diverse set of simple transformations and using an energy-difference loss would be a useful and practical contribution to OOD detection. The paper also ships a public implementation, provides extensive benchmarking against many baselines, and includes ablations over transformations, backbones, and hyperparameters, which are strengths. However, the significance of these results is currently overshadowed by the sign error in Eq. (12), which makes the printed OEST* objective implement the opposite of the claimed energy barrier, and by the fact that Theorem 1 is a direct restatement of Assumption 1 rather than a guarantee about what the training loss achieves. These issues must be resolved before the paper's central claims can be accepted.","major_comments":[{"comment":"Eq. (12) defines Lenergy*(xin,x'in) = log σ((E(xper)-E(xin))/β), and Algorithm 1 minimizes L = LCE + α·Lenergy*. Since log σ(z) is monotonically increasing with positive derivative, minimizing this loss drives E(xper)-E(xin) toward minus infinity, i.e., it pushes peripheral-distribution energy below in-distribution energy. This is the exact opposite of the energy barrier required by Assumption 1, which postulates E(x+;f)-E(x;f) > γα ≥ 0, and it is also opposite to the mechanism described in the abstract and Figure 1b. As written, the reported OEST* results in Tables II, III, and VIII cannot be consequences of the stated objective. The text provides no alternative sign convention or maximization step that would rescue Eq. (12); either the released code contains a sign flip absent from the manuscript, or the empirical improvement has a different cause. This is the load-bearing flaw of the paper and must be fixed and verified before any further evaluation.","section":"Eq. (12) and Algorithm 1"},{"comment":"The proof of Theorem 1 merely combines Assumption 1 with the Cauchy-Schwarz inequality: Assumption 1 already asserts E(x+;f)-E(x;f) > B‖x'-x+‖ + γα, and the proof bounds E(x';f)-E(x+;f) ≥ -B‖x'-x+‖, so the conclusion E(x';f)-E(x;f) > γα is a direct consequence of the assumption. The theorem therefore does not establish that the proposed training objective creates or maintains the energy barrier; it only says that if such a barrier exists, then a separation guarantee follows. This is a conditional restatement rather than a theoretical justification of the method. The paper's own Remark after Theorem 1 acknowledges this, and Section V-C4 further states that Assumption 1 fails on CIFAR-100 with ResNet-18, which is precisely the setting of the headline improvement in Table III. Consequently, the theory does not cover the main empirical claim in the configuration where OEST* is reported to be state of the art.","section":"Theorem 1 and Assumption 1 (Section IV-B, Appendix A)"},{"comment":"Section V-C4 concedes that the energy-barrier assumption is violated for the CIFAR-100/ResNet-18 setup, yet Tables III and VII report the largest absolute improvements in that setup (average AUROC 88.03% and a 12.29% AUROC gain over EBO with ResNet-50). This creates an internal inconsistency: the paper presents OEST* as a theoretically grounded method whose success is explained by Theorem 1, but the very benchmark used to demonstrate the method's advantage is one where the authors say the assumption does not hold. At minimum, the paper needs to separate the empirical claim from the conditional theoretical claim and provide an alternative explanation for the CIFAR-100 results, or the theoretical framing should be substantially weakened.","section":"Section V-C4 and Table III"}],"minor_comments":[{"comment":"The margin hyperparameter is written as m_per in Eq. (11) but as \"m_pre\" in Section V-A3; please unify the notation.","section":"Eq. (11) and Section V-A3"},{"comment":"The text says six simple transformations are applied for SVHN but then lists five (noise, blur, perm, rotation, and sobel); either a transformation such as cutout is missing from the list, or the count is wrong.","section":"Appendix C-B"},{"comment":"The caption contains the typo \"orqange\" for \"orange\" in the description of rotated CIFAR-10 samples.","section":"Figure 1 caption"},{"comment":"The table header \"Dtrain in pre-train + fine-tune\" is unclear; the first column appears to mix a data label with a training-scheme label and should be split or reworded for readability.","section":"Appendix C-C3, Table IX"},{"comment":"The brackets around the expression in Eq. (12) are visually confusing; the outer square brackets add no mathematical meaning and should be removed to make the loss definition clearer.","section":"Eq. (12)"}],"recommendation":"reject","confidential_remarks":"The sign discrepancy between Eq. (12) and the claimed energy-barrier mechanism is the kind of issue that warrants checking the released code before any revision is considered. If the code uses a sign that actually raises peripheral energy relative to in-distribution energy, the manuscript must be rewritten to describe that objective explicitly; if the code uses the printed sign, the benchmark results are unexplained by the stated method. I would also strongly encourage the editor to verify the code and the reported numbers before any resubmission, since the paper's central claim currently contradicts its own objective function."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The central objective in Eq. (12) is spelled backwards. Lenergy* = log σ((E(xper)-E(xin))/β), and Algorithm 1 minimizes it. Since log σ is monotone increasing, minimizing drives E(xper)-E(xin) toward -∞, i.e., PD energies end up lower than ID energies. That is the opposite of the energy barrier Assumption 1 requires and the opposite of what the abstract and Fig. 1b claim. Unless the released code flips the sign, the reported SOTA numbers cannot follow from the printed method. This is the first thing to sort out.\n\nWhat is actually new and good: the peripheral-distribution framing is a reasonable repackaging of the idea that augmented views can serve as cheap outliers (CSI, RotPred), and the paper is upfront about that lineage. The energy-barrier assumption and Theorem 1 give a clean, if conditional, statement of why augmented data should help. The empirical story is also worth taking seriously: a 10-epoch fine-tune with no curated outliers improves both near and far OOD detection on CIFAR-10/100, and the authors include ablations over transformations, backbones, and hyperparameters. They also admit in Sec. V-C4 that the assumption fails on CIFAR-100 with ResNet-18, which is honest.\n\nSoft spots: the sign error is load-bearing, not cosmetic. Theorem 1 rests on Assumption 1, which is strong and only provable for a linear classifier in the appendix; the paper itself shows the assumption does not hold in a main configuration. The reported numbers lack error bars, so I can't tell whether the gains over RotPred are significant. And the relation to CSI/RotPred deserves a more direct comparison in the main text.\n\nWho is this for? Researchers working on OOD detection who want a cheap tuning alternative to curated outlier exposure. The paper deserves a serious referee, but not in its current form. I'd recommend major revision: fix the sign in Eq. (12), verify the code implements the corrected loss, provide error bars, and either strengthen Assumption 1 or reframe it as a heuristic.","headline":"Eq. (12)'s energy-barrier loss is sign-inverted, so the paper's central objective as written contradicts its own theory; the empirical results may still hold, but the text needs a major fix before it can be trusted.","tokens_in":28193,"tokens_out":4726,"would_cite":false,"duration_ms":44395,"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":"Tuning a pretrained classifier for ten epochs on its own simply augmented images makes it reliably flag out-of-distribution inputs, and the paper's energy-barrier loss achieves top results on CIFAR-10 and CIFAR-100.","keywords":["out-of-distribution detection","energy-based models","outlier exposure","data augmentation","peripheral-distribution samples","energy barrier","CIFAR-10","CIFAR-100"],"falsifier":"A concrete test: on the CIFAR-100-as-ID, CIFAR-10-as-OOD setting with ResNet-18, measure the actual energy gaps $E(x')-E(x)$ over many OOD images and compare with the paper's reported AUROC of 75.22, which is below the untuned energy baseline; a sharper check is to compute, for each OOD image, the nearest peripheral-distribution sample in feature space and test whether Assumption 1's inequality holds, because if the inequality fails on most OOD images yet the energy ordering still holds, the theorem is not the mechanism, and if the inequality holds but the ordering fails, the assumption is insufficient.","tokens_in":27138,"feed_emoji":"🔍","tokens_out":7809,"duration_ms":70159,"temperature":0.7,"pith_summary":"The paper claims that a classifier can be turned into a reliable out-of-distribution detector without any external outlier data, by tuning it for ten epochs on 'peripheral-distribution' samples: its own training images run through simple transformations such as rotation, cutout, blur, noise, permutation, and Sobel filtering. The central idea is an energy barrier: the transformed samples are pushed to higher energy than the original images, and the paper proves (Theorem 1, under Assumption 1) that this forces any genuinely out-of-distribution input to receive higher energy than an in-distribution input with probability $1-\\alpha$. If correct, this would make OOD detection a cheap post-hoc add-on to existing classifiers, avoiding the fragile reliance on curated outlier datasets. Empirically the improved version OEST* reports average AUROC 96.30% on CIFAR-10 and 88.03% on CIFAR-100, with the lowest FPR95 among compared methods in both settings.","feed_headline":"Simple image tweaks teach models to flag the unfamiliar","feed_subtitle":"A 10-epoch tune on transformed input copies lifts mean AUROC to 96.3% (CIFAR-10) and 88.0% (CIFAR-100).","key_machinery":"Peripheral-distribution (PD) data and the energy-barrier inequality carry the argument. PD data are augmented samples $\\mathcal{B}^+ = \\bigcup_{S\\in\\mathcal{S}}\\{S(x_i)\\}$ generated from ID images by a fixed set of simple transformations, placed in feature space between ID and OOD samples. The load-bearing identity is Theorem 1, proved from Assumption 1 by inserting $x^+$ into $E(x')-E(x)$ and bounding $E(x')-E(x^+) \\geq -B\\|x'-x^+\\|$ using the bounded class vectors of the linear classifier. The energy-barrier loss of Eq. (12) is the training mechanism that establishes the barrier: it minimizes $\\log\\sigma\\big((E(x_{\\text{per}})-E(x_{\\text{in}}))/\\beta\\big)$, a function of the energy difference only, so the intractable $\\log Z$ cancels, which is the theoretical fix over the earlier energy-bounded loss of Eq. (11).","core_discovery":"The paper's central claim is that OOD detection does not require seeing real outliers during training; it is enough to establish an energy barrier between in-distribution images and their augmented 'peripheral' versions. The authors define this barrier in Assumption 1: for a random ID sample $x$ and any OOD sample $x'$, some augmented sample $x^+$ satisfies $E(x^+;f)-E(x;f) > B\\|x'-x^+\\| + \\gamma_\\alpha$ with probability $1-\\alpha$. Theorem 1 then shows $E(x';f)-E(x;f) > \\gamma_\\alpha$, so OOD inputs sit above ID inputs in energy. To create this barrier, the paper replaces the classical energy-bounded loss, which implicitly ignores the changing partition function during training, with an energy-barrier loss $\\mathcal{L}_{\\text{energy}^*} = \\log\\sigma\\big((E(x_{\\text{per}})-E(x_{\\text{in}}))/\\beta\\big)$ that depends only on energy differences and is therefore invariant to the partition function. The result is an extension of energy-based OOD scoring that uses simple transformations as surrogate outliers and a theoretically motivated loss.","pith_inferences":["If the barrier mechanism is what matters, the transformation set should be chosen so that augmented samples sit just outside the ID cluster but inside the region where OOD data live; a dataset whose natural augmentations stay fully ID would need a different set, and the energy gap on held-out validation could select it.","The partition-function cancellation is a general design principle: any auxiliary loss stated purely in energy differences is immune to the changing normalizer, so the same barrier loss could be applied to other logit-based scores or representation-level energies, not just the softmax free energy.","The method suggests a testable equivalence between OOD detection and augmentation-graph geometry: OOD samples are those reachable from ID by a path of transformation steps, so the energy barrier should transfer along that graph; one could verify whether OOD inputs with larger graph distance to ID get larger energy gaps.","Because only ten epochs of tuning are needed, the same recipe could be applied on top of large pretrained feature extractors without retraining the backbone, potentially giving OOD detection for foundation models at negligible cost, a direction the paper does not test."],"forward_implications":["Tuning a pretrained classifier for roughly ten epochs on its own augmented images can rival or beat methods trained from scratch with curated outliers; on CIFAR-10 the ID accuracy drops only about 0.09% while average AUROC rises to 96.30%.","The energy-barrier loss makes the maximum-likelihood spirit of energy-based tuning rigorous, because it removes the dependence on the partition function, which the paper shows fluctuates substantially during tuning.","Under Assumption 1, OOD detection reduces to thresholding the Helmholtz free energy: with probability $1-\\alpha$ every OOD input has higher energy than a random ID input.","The assumption holds only when the feature extractor clusters ID tightly: on CIFAR-100 with ResNet-18 the method underperforms, while with ResNet-34 and ResNet-50 the AUROC gains over the energy baseline become +10.10 and +12.29 points.","Combining several simple transformations beats any single one on average, so the method does not need the per-dataset transformation search that single-augmentation contrastive approaches require."],"supporting_citations":[{"why":"Supplies the energy-based OOD score and the energy-bounded loss that the paper replaces with its energy-barrier loss.","marker":"[23]"},{"why":"Establishes the precedent of treating augmented samples as negatives; the paper generalizes this to a broad set of transformations.","marker":"[37]"},{"why":"Provides the pretrained ResNet classifiers and benchmark protocol used in all main CIFAR experiments.","marker":"[20]"},{"why":"Defines the unified OOD evaluation benchmark and the near/far OOD dataset split used in the main results tables.","marker":"[19]"},{"why":"Gives the energy-based model formalism, Gibbs distribution, and Helmholtz free energy on which the score function rests.","marker":"[36]"},{"why":"Supplies the augmentation-overlap style assumption that Assumption 1 adapts from contrastive learning theory.","marker":"[79]"},{"why":"Together with [79], supports the geometric assumption that an augmented sample can sit close to an OOD sample in feature space.","marker":"[80]"},{"why":"The preliminary conference version that introduced simple transformations for OOD detection and is extended here with the energy-barrier analysis.","marker":"[38]"}],"fun_headline_variants":["No real outliers needed: simple transforms train OOD detectors","Energy barrier loss lifts OOD detection without crafted outliers","Simple tweaks set energy barrier for out-of-distribution detection","OEST*: surrogate outliers via simple transformations improve OOD","Energy-based OOD detection gets a partition-free loss boost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof only works if every out-of-distribution input has some simply transformed in-distribution image that is simultaneously close to it in feature space and already carries a large energy gap over the original image; when that proxy image does not exist, as the paper finds for CIFAR-100 with a ResNet-18, Theorem 1 provides no separation guarantee.","fun_headline_variants_meta":{"raw":{"variants":["No real outliers needed: simple transforms train OOD detectors","Energy barrier loss lifts OOD detection without crafted outliers","Simple tweaks set energy barrier for out-of-distribution detection","OEST*: surrogate outliers via simple transformations improve OOD","Energy-based OOD detection gets a partition-free loss boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000191,"raw_usage":{"total_tokens":1397,"prompt_tokens":1055,"completion_tokens":342,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":260}},"tokens_in":671,"tokens_out":342,"duration_ms":3850,"temperature":1.0,"reasoning_tokens":260,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:49:32.397256+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: on the CIFAR-100-as-ID, CIFAR-10-as-OOD setting with ResNet-18, measure the actual energy gaps $E(x')-E(x)$ over many OOD images and compare with the paper's reported AUROC of 75.22, which is below the untuned energy baseline; a sharper check is to compute, for each OOD image, the nearest peripheral-distribution sample in feature space and test whether Assumption 1's inequality holds, because if the inequality fails on most OOD images yet the energy ordering still holds, the theorem is not the mechanism, and if the inequality holds but the ordering fails, the assumption is insufficient.","supporting_citations":[{"cited_title":"CSI: Novelty Detection via Contrastive Learning on Distributionally Shifted Instances,","cited_arxiv_id":null,"evidence_quote":"Establishes the precedent of treating augmented samples as negatives; the paper generalizes this to a broad set of transformations."},{"cited_title":"A Tu- torial on Energy-Based Learning,","cited_arxiv_id":null,"evidence_quote":"Gives the energy-based model formalism, Gibbs distribution, and Helmholtz free energy on which the score function rests."},{"cited_title":"Chaos is a Ladder: A New Theoretical Understanding of Contrastive Learning via Augmentation Overlap,","cited_arxiv_id":null,"evidence_quote":"Supplies the augmentation-overlap style assumption that Assumption 1 adapts from contrastive learning theory."},{"cited_title":"Connect, Not Collapse: Explaining Contrastive Learning for Unsupervised Domain Adaptation,","cited_arxiv_id":null,"evidence_quote":"Together with [79], supports the geometric assumption that an augmented sample can sit close to an OOD sample in feature space."},{"cited_title":"OEST: Outlier Exposure by Simple Transformations for Out-of-Distribution Detection,","cited_arxiv_id":null,"evidence_quote":"The preliminary conference version that introduced simple transformations for OOD detection and is extended here with the energy-barrier analysis."}],"review_version":1}