{"id":"6a406f6c-a822-4fc0-a776-d40c2a17babf","arxiv_id":"2505.12952","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A threshold-free OOD detection method that labels wild data as an extra class and clusters per-sample training losses to separate in-distribution from out-of-distribution data.","lead":"This paper proposes marking all unlabeled 'wild' data as an extra class, then using the difference in training losses to separate in-distribution from out-of-distribution samples without setting a threshold. The method reports very low false-positive OOD rates on standard and harder benchmarks, suggesting a simple route to OOD detection in deployed models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 1's proof does not model LoD's K+1-class loss, where ID and OOD wild samples share the K+1 label, so the theoretical loss-ordering guarantee is unproven.","rationale":"The reader identified the transfer from a two-class Gaussian model to the K+1-class deep network and the non-positivity of the bound as the weakest assumption, and also flagged the Eq. (6) omission. My analysis goes further: the proof's loss expressions are derived for samples with opposite observed labels, whereas in LoD both wild ID and wild OOD share the K+1 label. This is a model mismatch, not merely a loose bound. The Proposition 1 result is therefore not a valid theoretical guarantee for the method's central filtering mechanism. This strengthens the reader's conditional verdict: the theoretical claims should be revised or explicitly weakened, while the empirical evaluation remains convincing enough that the method should not be rejected outright. I partially agree with the reader because they focused on transfer and non-positivity, but the more fundamental issue is the mismatch between the proof's label-noise model and the actual training loss. The verdict remains CONDITIONAL as the reader stated; no adjustment is needed.","tokens_in":19854,"tokens_out":5966,"duration_ms":60776,"concrete_test":"Re-derive Proposition 1 from the actual LoD objective (Eq. 2) using a K+1-class linear model with ID samples drawn from the K ID Gaussians and OOD samples drawn from a separate Gaussian, both labeled as class K+1. Compute the CE loss -log p_{K+1}(x) for both groups and check whether an analogous lower bound l_in - l_out ≥ c > 0 holds under the early-learning conditions of Lemma 1. If the derived bound is not positive, or if positivity requires assumptions not stated in the paper (e.g., OOD logits exceeding ID logits for class K+1), then Proposition 1 does not support the threshold-free filtering claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 1 (Sec. 4.2 and Appendix A) is derived from a two-class Gaussian model with symmetric label noise (Eq. 7), where a sample's observed label is flipped with probability Δ. In that model, label-clean samples have loss log(1+e^{-θ^T x}) and label-noise samples have loss log(1+e^{+θ^T x}), i.e., opposite signs in the logit. In LoD's actual objective (Eq. 2), however, both ID and OOD samples in D_wild are assigned the same observed label y_{K+1}; neither group receives a flipped opposite label. Therefore the central inequalities l_clean ≤ e^{-θ^T(v+z)} and l_noise ≥ 1 - e^{-θ^T(v+z)} do not apply to the CE loss of the K+1-th class. The proof never introduces the K+1 logit or the K ID classes; it proves a different statement. Consequently the claimed lower bound l_in - l_out ≥ 1 - 2e^{-θ^T v + (1/2)||θ||^2 δ^2} - O(R/√n) is not tied to the LoD training loss. Even within the binary model, positivity requires 1 - 2e^{-θ^T v + ...} > 0, which is not shown for the learned θ. Separately, the gradient update in Eq. (6) omits ∇L_{N3}, the OOD term from the decomposition in Eq. (5), so the 'joint dominance' argument in Sec. 4.1 is also unsupported by the displayed equation. The empirical results are strong, but the theoretical foundation as written does not establish the central loss-ordering claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes LoD, a two-stage OOD detection framework that uses unlabeled wild data. In the filtering stage, all wild samples are intentionally labeled as a single K+1-th class and trained together with labeled ID data in a fully supervised K+1-class setting, so that OOD wild samples become effectively clean labels while ID wild samples become conflicting labels. The mean cross-entropy losses of wild samples over training are then clustered with K-means into two groups, ID (high loss) and OOD (low loss), removing the need for a manually tuned threshold. In the detection stage, the filtered OOD samples and labeled ID data train a binary OOD detector. The paper argues that this reformulation addresses model bias and threshold selection, provides a theoretical analysis of a claimed loss gap, and reports extensive experiments on standard and hard benchmarks.","tokens_in":20259,"tokens_out":3618,"duration_ms":37158,"significance":"If the loss-ordering claim holds, LoD is a simple, data-centric, threshold-free approach to wild-data OOD detection, and the paper has several strengths: the problem reformulation is clear and elegant, the experimental evaluation is extensive (standard and hard benchmarks, unseen OOD datasets, multiple network architectures, and ratio/epoch ablations), and the reported gains over OE, Energy, WOODS, and SAL are consistent. I also credit the paper for curating harder same-source benchmarks. However, the theoretical foundation as written does not establish the central loss-gap guarantee: the proof models a different binary label-noise setting and the displayed gradient update omits one of the three terms. The empirical contribution is strong, but the advertised theory needs substantial repair.","major_comments":[{"comment":"The gradient update in Eq. (6) drops the OOD term. Eq. (5) decomposes the loss into three parts with gradients ∇L_N1, ∇L_N2, and ∇L_N3, but Eq. (6) writes θ_{t+1} = θ_t − η(∇L_N1 + ∇L_N2), omitting ∇L_N3. Consequently the displayed update contradicts the decomposition and cannot support the claim that labeled ID data and OOD wild data jointly dominate learning. Even if the term is reinserted, the sample-size argument |B_train_in| > |B_wild| ≥ N2 does not by itself imply that ∇L_N3 is non-negligible, since gradient magnitudes depend on the model state, not only on batch counts.","section":"Sec. 4.1, Eq. (5)–(6)"},{"comment":"The proposition is proved in a two-class Gaussian model with symmetric label noise, where a clean sample has log-likelihood log(1+e^{−θ^T x}) and a noisy sample has log(1+e^{+θ^T x}). In LoD's actual objective, Eq. (2), both ID and OOD wild samples are assigned the same observed label y_{K+1}; there is no binary sign flip and the ID wild samples are not labeled with an opposite binary class. The inequalities l_clean ≤ e^{−θ^T(v+z)} and l_noise ≥ 1−e^{−θ^T(v+z)} therefore do not apply to the K+1-class cross-entropy loss. The proof establishes a statement about a different training objective, not about the LoD training loss.","section":"Sec. 4.2 and Appendix A, Proposition 1"},{"comment":"Even within the binary model, the claimed lower bound is not shown to be positive. The right-hand side 1 − 2e^{−θ^T v + (1/2)||θ||^2 δ^2} − O(R/√n) can be negative unless θ^T v is sufficiently large relative to ||θ||^2 δ^2; Lemma 1 only guarantees the normalized gradient direction, not a lower bound on θ^T v itself. In addition, the proposition and its proof use δ^2 while Eq. (9) derives the Gaussian expectation with σ^2, so the notation conflates the label-noise parameter with the Gaussian variance. Without a positivity argument, the proposition cannot justify the K-means separation used in Eq. (3).","section":"Sec. 4.2, Proposition 1 and Eq. (9)"},{"comment":"The main empirical tables report exact 0.00 FPR95 and 100.00 AUROC values for many settings, but no error bars, standard deviations, or per-seed results are provided in the main text; Appendix B mentions averaging over five splits for hard benchmarks, yet the main tables do not report the spread. Given the saturation at 0/100, error bars are necessary to assess whether the differences among methods are meaningful and whether the 0.00 values are stable across runs.","section":"Tables 1 and 2"}],"minor_comments":[{"comment":"The derivation uses E[e^{−θ^T z}] = e^{(1/2)||θ||^2 σ^2}, but Proposition 1 and the text use δ^2; please make the notation consistent.","section":"Appendix A, Eq. (9)"},{"comment":"The caption contains a typo: 'different rations' should be 'different ratios'.","section":"Figure 4 caption"},{"comment":"The loss-recording step V ← V ∪ {l_i | i ∈ (1, |B_wild|)} is underspecified: it should state how losses are associated with individual samples in D_wild when batches are sampled with replacement, and whether each sample's loss is recorded once per epoch.","section":"Algorithm 1, line 6"},{"comment":"The captions of Table 4 and Table 5 refer to 'Table 2 and Table 3' in a way that collides with the main-text numbering; please renumber the appendix tables and references.","section":"Appendix C and D captions"},{"comment":"The phrase 'with high probability' is used without a quantitative probability in either Lemma 1 or Proposition 1; please state the failure probability or clarify the o(1) convention.","section":"Sec. 4.2, Lemma 1/Proposition 1"},{"comment":"Calling OOD wild samples 'label-clean' under the synthetic K+1-th class is conceptually nonstandard, since every wild sample receives the same artificially assigned label; a brief clarification of what 'clean' means in this synthetic-label setting would help readers.","section":"Sec. 3.2, Remark"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead LoD. The one thing to know: the core move is better than I expected from the title. Label every wild sample as class K+1, train a K+1 classifier, then K-means the per-sample mean losses to separate ID from OOD without thresholds. That genuinely reframes unlabeled-wild OOD filtering as label-noise learning, and the paper shows it works across standard and hard benchmarks with consistent gains over SAL, WOODS, OE, and Energy. This is the paper's real value and it should be preserved.\n\nWhat is new: the deliberate K+1 relabeling is not in the cited baselines, and using loss-curve clustering rather than a threshold is a clean way around the threshold dilemma. The authors also curated hard benchmarks (same dataset, different classes) and included experiments across networks, ratios, and unseen OOD sets. That is a solid empirical package. Code is linked. I would believe the method works in the tested settings, though I would want error bars before trusting numbers like 0.00 FPR95: several are exact zeros, and no standard deviations or per-split intervals are reported.\n\nWhere it softens: the theoretical section does not establish what the paper claims. Eq. (5) decomposes the loss into three terms; Eq. (6) writes the gradient update as θ_{t+1} = θ_t − η(∇L_N1 + ∇L_N2), dropping the OOD term ∇L_N3. That is a concrete error, and the \"joint dominance\" conclusion does not follow from the displayed equation. Proposition 1 is worse. It is derived from the two-class Gaussian model with symmetric label flipping (Eq. 7), where clean and noisy samples have opposite labels and opposite logit signs. LoD's actual objective, Eq. (2), assigns both ID and OOD wild samples the same K+1 label; neither group receives an opposite label. So the inequalities l_clean ≤ e^{−θ^T x} and l_noise ≥ 1 − e^{−θ^T x} do not apply to the K+1 cross-entropy loss. The stress-test note is right: the proof proves a statement about a different labeling scheme, and even inside that scheme the lower bound's positivity is not shown. This is not a missing step; the object being analyzed is not LoD's loss.\n\nMinor issues: the ratio ablation supports the gradient intuition but does not rescue it, and the \"long early-learning\" explanation for 100–500 epochs is plausible but hand-wavy. The citations to the label-noise literature are appropriate.\n\nBottom line: this paper deserves serious refereeing because the empirical method is novel, simple, and probably useful to the OOD detection and label-noise communities. But the authors should be asked to either repair the proof or remove the formal claim and present Proposition 1 as intuition. I would accept a revised version with error bars and a corrected theory section.\n\nSend it out.","headline":"The empirical recipe is a genuine new angle and the results are strong; the theory is not load-bearing as written, so treat the paper as an empirical contribution until the proof is repaired.","tokens_in":20779,"tokens_out":3083,"would_cite":true,"duration_ms":33015,"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":"The paper claims that intentionally labeling all unlabeled wild data as a single fake extra class makes OOD samples separable from ID samples by K-means on training losses, with no threshold required.","keywords":["out-of-distribution detection","unlabeled wild data","label-noise learning","early learning","K-means filtering","threshold-free detection","Huber contamination model","loss gap"],"falsifier":"Train LoD on a wild set whose OOD classes are chosen to be nearest neighbours of the ID classes, then plot the per-sample mean-loss histograms for true ID and true OOD samples. If the two histograms overlap so strongly that two-cluster K-means purity on the wild set falls to chance level, or if the measured mean-loss gap is not positive, the central claim is falsified. A simpler observational check: on the paper's own TinyImageNet hard benchmark, the reported FPR95 around 45% already indicates imperfect cluster separation, so measuring whether cluster accuracy tracks the Proposition 1 bound would settle the issue.","tokens_in":19658,"feed_emoji":"🎯","tokens_out":8144,"duration_ms":76949,"temperature":0.7,"pith_summary":"This paper tries to show that a known difficulty in out-of-distribution (OOD) detection — how to use unlabeled wild data that mixes in-distribution (ID) and OOD images — can be handled by deliberately labeling all wild data as a single extra class, the K+1-th class. The labeling converts ID samples in the wild into label-noise samples and OOD samples into label-clean samples, so the network's own training losses separate the two: noisy ID samples keep higher, more fluctuating losses while clean OOD samples settle to lower losses. That separation is strong enough that plain K-means clustering on mean training losses picks out the OOD samples, eliminating the need for a threshold. The paper backs this with a theoretical loss-gap bound and reports false-positive rates near 0% on standard benchmarks while keeping competitive ID accuracy. A sympathetic reader would care because threshold-free filtering and joint ID/OOD dominance remove two design obstacles in wild-data OOD detection.","feed_headline":"Fake extra-class labels separate OOD from ID with no threshold","feed_subtitle":"Relabel wild data as one extra class, clean OOD samples fit fast, noisy ID samples don't — K-means does the rest.","key_machinery":"The load-bearing mechanism is intentional label-noisification: assigning the single K+1-th class to all unlabeled wild data, which converts the OOD-filtering problem into a label-noise learning problem with a controllable noise ratio. The work it does is to produce the early-learning loss asymmetry — clean OOD samples fit quickly, noisy ID samples do not — and the resulting mean-loss gap that two-cluster K-means exploits. The theoretical engine is Proposition 1, which transfers an early-learning guarantee from a two-class Gaussian model to a lower bound on the difference between mean cross-entropy losses of ID and OOD wild samples; the clustering step replaces threshold selection with the distance rule $d_1 < d_2$. A useful named model is the Huber contamination model $P_{\\mathrm{wild}} = (1-\\pi) P_{\\mathrm{in}} + \\pi P_{\\mathrm{out}}$, which formalizes the wild mixture.","core_discovery":"The central claim is that OOD filtering in unlabeled wild data can be reformulated as a controllable label-noise learning problem. Label every sample in the wild set as class K+1 and train a K+1-way classifier together with clean K-class labeled ID data. During the early-learning phase, the K+1-labeled ID samples behave like mislabeled data: their cross-entropy losses stay higher and fluctuate more, whereas K+1-labeled OOD samples behave like clean data and converge to low losses. Proposition 1 formalizes this as a gap in mean losses bounded below by $1 - 2 e^{-\\theta^T v + \\frac{1}{2}\\|\\theta\\|^2 \\delta^2} - O(R/\\sqrt{n})$, so with enough samples the two populations separate; K-means on per-sample mean losses with two clusters then identifies OOD without any threshold. The framework then trains a binary OOD detector on the filtered OOD candidates plus labeled ID data. The empirical claim is that this yields near-zero FPR95 on standard benchmarks and large margins over prior wild-data methods on harder same-dataset benchmarks.","pith_inferences":["Editorial inference: the same loss-gap mechanism should transfer to other training objectives that exhibit early-learning behavior, such as symmetric losses or robust cross-entropy variants, provided the asymmetry (ID noisy, OOD clean) is preserved; testing this would clarify whether the K+1 relabeling itself is the essential ingredient.","Editorial inference: the threshold-free property depends on the mean-loss distributions being bimodal and separable; on the hardest reported benchmark, TinyImageNet, FPR95 remains around 45%, so the universal no-threshold promise is bounded by how distinguishable OOD is from ID in feature space.","Editorial inference: because the noise ratio is controlled by batch composition, an online stopping rule based on cluster separation (e.g., a bimodality index or silhouette score) could make the early-learning window self-terminating instead of fixed at a chosen epoch count."],"forward_implications":["If the loss-gap premise holds, OOD filtering from wild data needs no threshold: a fixed two-cluster K-means on mean training losses is sufficient.","Because wild OOD samples are correctly labeled as the extra class and jointly dominate learning with labeled ID data, the model-bias problem that makes networks fit OOD as ID is reduced.","The label-noise ratio is controllable through the batch ratio $|B_{\\mathrm{train}}^{\\mathrm{in}}| : |B_{\\mathrm{wild}}|$, and the paper reports that larger ratios improve filtering performance.","The framework is data-centric and architecture-agnostic, working across WideResNet-40-2, ResNet18, and ResNet34 without modifying network structure.","On hard benchmarks where ID and OOD come from the same dataset, the paper reports substantially lower FPR95 than prior wild-data methods, with the largest gains on CIFAR10-based splits."],"supporting_citations":[{"why":"Defines the Huber contamination model for unlabeled wild data and the goal of training OOD detectors in their natural habitats, which LoD adopts as its problem setup.","marker":"[Katz-Samuels et al., 2022]"},{"why":"Supplies the two-step filtering-then-detector strategy, the detector training objective, and the benchmark protocol that LoD extends and compares against.","marker":"[Du et al., 2024]"},{"why":"Provides Lemma 1 (early-learning succeeds), the theoretical result on loss separation under label noise that Proposition 1 is built on.","marker":"[Liu et al., 2020a]"},{"why":"Contributes the clustering-training-losses idea and the sigmoid-activation bounds used in the proof of Proposition 1.","marker":"[Yue and Jha, 2024]"},{"why":"Establishes the empirical memorization and early-learning phenomenon in deep networks that motivates the label-noise reformulation.","marker":"[Arpit et al., 2017]"},{"why":"Introduces outlier exposure, the auxiliary-OOD-data baseline that wild-data methods such as LoD are contrasted with.","marker":"[Hendrycks et al., 2018]"}],"fun_headline_variants":["Turn wild data into label noise to catch OOD with no threshold","Loss gaps from fake labels let K-means spot OOD without thresholds","No-threshold OOD detection via loss-difference on noisy-labeled wild data","Label wild data as extra class; K-means separates OOD without tuning","Fake extra labels create a loss gap; K-means finds OOD for free"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands or falls on the assumption that real out-of-distribution images, once labeled as the fake extra class, are consistently easier for the deep network to fit than real in-distribution images that were deliberately mislabeled as that same class; the simplified two-Gaussian proof does not by itself guarantee that this ordering holds for a deep K+1 classifier on arbitrary wild mixtures.","fun_headline_variants_meta":{"raw":{"variants":["Turn wild data into label noise to catch OOD with no threshold","Loss gaps from fake labels let K-means spot OOD without thresholds","No-threshold OOD detection via loss-difference on noisy-labeled wild data","Label wild data as extra class; K-means separates OOD without tuning","Fake extra labels create a loss gap; K-means finds OOD for free"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000773,"raw_usage":{"total_tokens":3461,"prompt_tokens":1025,"completion_tokens":2436,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":641,"completion_tokens_details":{"reasoning_tokens":2336}},"tokens_in":641,"tokens_out":2436,"duration_ms":14757,"temperature":1.0,"reasoning_tokens":2336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:22:52.883350+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LoD on a wild set whose OOD classes are chosen to be nearest neighbours of the ID classes, then plot the per-sample mean-loss histograms for true ID and true OOD samples. If the two histograms overlap so strongly that two-cluster K-means purity on the wild set falls to chance level, or if the measured mean-loss gap is not positive, the central claim is falsified. A simpler observational check: on the paper's own TinyImageNet hard benchmark, the reported FPR95 around 45% already indicates imperfect cluster separation, so measuring whether cluster accuracy tracks the Proposition 1 bound would settle the issue.","supporting_citations":[{"cited_title":"Training ood de- tectors in their natural habitats","cited_arxiv_id":null,"evidence_quote":"Defines the Huber contamination model for unlabeled wild data and the goal of training OOD detectors in their natural habitats, which LoD adopts as its problem setup."},{"cited_title":"Ctrl: Clus- tering training losses for label error detection.IEEE Trans- actions on Artificial Intelligence,","cited_arxiv_id":null,"evidence_quote":"Contributes the clustering-training-losses idea and the sigmoid-activation bounds used in the proof of Proposition 1."},{"cited_title":"A closer look at memo- rization in deep networks","cited_arxiv_id":null,"evidence_quote":"Establishes the empirical memorization and early-learning phenomenon in deep networks that motivates the label-noise reformulation."}],"review_version":1}