{"id":"e3dbebcf-d49c-40a4-9d44-b6ab57379461","arxiv_id":"2508.19581","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A discriminator trained to distinguish clean from corrupt image-label pairs can be used during sampling to correct the score of a noisy-label conditional diffusion model, improving class-wise fidelity without retraining.","lead":"This paper proposes Score-based Discriminator Correction (SBDC), an inference-time guidance method that uses a small discriminator to steer pre-trained conditional diffusion models away from errors caused by noisy training labels. It requires no retraining of the diffusion model and adds only minor inference-time overhead.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's bound is vacuous and the discriminator gradient does not equal the required score correction; the central derivation is unsupported.","rationale":"The reader's weakest assumption (noise detector bias) is a practical limitation that affects performance but not the core derivation. The concern raised here is more load-bearing: even with a perfect noise detector, the theoretical link between the discriminator gradient and the clean-distribution score correction is not demonstrated. Theorem 1's bound is vacuous because it is satisfied by a zero gradient, and a direct derivation shows the discriminator log-odds differs from the required log-likelihood ratio. This undermines the central claim that SBDC 'recovers the clean distribution,' although the method may still work as an effective heuristic. The verdict remains CONDITIONAL rather than REJECT because the empirical results are consistent and could stand if the paper reframes the contribution as a heuristic with a weaker, empirically validated property, or if a corrected theory is supplied. The concrete test would settle whether the heuristic's gradient actually points toward the clean conditional, which is the minimal support the central claim needs.","tokens_in":22744,"tokens_out":11460,"duration_ms":100341,"concrete_test":"On a two-class Gaussian mixture with known clean conditionals p(x|y) and injected symmetric label noise at rate η, train an EDM on the noisy labels and train the discriminator as in Algorithm 1 with a perfect noise detector. For a grid of timesteps t, compute the exact correction ∇ log p(xt|y)/p_noisy(xt|y) in closed form (Gaussian convolution) and the discriminator logit gradient ∇ log D/(1−D). Measure their cosine similarity and MSE across t and η. If the similarity is not close to 1 or the MSE does not approach zero as discriminator capacity grows, Theorem 1's claim fails and the guidance is not implementing Eq. (8). Report these numbers for η ∈ {0.2, 0.5, 0.8}.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central theoretical claim is that the discriminator gradient approximates the correction term in Eq. (8), i.e., ∇ log p(xt|y)/pθ(xt|y~). Theorem 1 in Appendix A does not establish this. In the proof, the score-matching reformulation cancels both terms using ∇ log p(xt|x0, y) = ∇ log p(xt|x0, y~) = ∇ log p(xt|x0), leaving LHS = E||∇gθ*||² + C1 and then bounding E||∇gθ*||² ≤ L. The final inequality LHS ≤ L + C1 is satisfied even by a useless discriminator with ∇gθ* = 0, because C1 is the squared norm of the very correction term being approximated. Thus the theorem is a tautology, not a proof of approximation. Independently of the proof, the discriminator trained per Eq. (9) estimates the log-odds p(r=1|xt,y~)/(1−p(r=1|xt,y~)). Under symmetric noise this equals log p(xt|y~, clean) − log Σ_{y'≠y~} p(xt|y')/(K−1) + const, whereas Eq. (8) requires log p(xt|y) − log p_noisy(xt|y~), which subtracts the full noisy conditional. These differ unless the noise rate is zero or the clean class dominates; the guidance is therefore not the theoretically derived clean-distribution score. Empirical gains may reflect a useful heuristic, but the paper's stated derivation is invalid.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Score-based Discriminator Correction (SBDC), an inference-time guidance method for conditional diffusion models trained on noisy labels. The method trains a small discriminator to distinguish clean from corrupted (image, label) pairs, using labels from a noise detector to pseudo-label training data. During sampling, the gradient of the discriminator's logit is added to the score network, gated to a limited time interval, with the aim of steering the noisy conditional model toward the clean distribution. The paper claims a theoretical guarantee (Theorem 1) that the discriminator gradient approximates the required score correction, and reports experiments on CIFAR-10/100, Tiny-ImageNet, ImageNet-128, and real-world datasets (Food101, Clothing1M), showing improvements over the TDSM baseline on class-wise metrics.","tokens_in":23085,"tokens_out":10580,"duration_ms":88046,"significance":"If the theoretical claim were sound, the method would be a valuable practical contribution: it is a plug-and-play module that requires no retraining of the diffusion model, adds negligible inference overhead, and improves conditional alignment under label noise. The empirical study is extensive, covering multiple noise types and levels, and includes useful ablations that identify the limited guidance interval as a key ingredient. The proposed augmentations (Pseudo-clean Shuffle and SiMix) are interesting and appear to provide benefits beyond the main guidance. However, the central theoretical justification is not correct: the proof of Theorem 1 in Appendix A contains unjustified cancellations and yields a vacuous bound, and the discriminator trained with Eq. (9) does not estimate the log-ratio required in Eq. (8). Since the paper explicitly states that the method is 'derived' from the theory, the unsupported theory is a load-bearing weakness.","major_comments":[{"comment":"The proof is invalid. The transition from Eq. (18) to Eq. (19) drops the cross terms between the gradient of the discriminator logit and the target log-ratio without justification; the two score-matching terms that are cancelled involve different label variables (y and y~) and only coincide if the marginal distribution of x0 is identical under both, which is not established. More importantly, the resulting bound in Eq. (10) is vacuous: the constant C1 is exactly the squared norm of the target correction term, so a discriminator with zero gradient satisfies the bound trivially. Thus Theorem 1 does not establish that the discriminator gradient approximates the correction term, and the claim in Sec. 3.2 that the guidance is 'derived' from the theory is unsupported.","section":"Appendix A, Theorem 1"},{"comment":"The discriminator trained with Eq. (9) estimates the log-odds of the pseudo-clean versus pseudo-corrupt status given the observed label, not the log-ratio of clean and noisy conditional densities. For symmetric noise, the discriminator logit gradient equals the gradient of log p(xt|y~) minus log sum_{y' != y~} p(xt|y')/(K-1) (up to additive constants), whereas Eq. (8) requires the gradient of log p(xt|y) minus log p_noisy(xt|y~), which subtracts the full noisy conditional. These quantities differ unless the noise rate is zero or the clean class dominates; the paper provides no argument closing this gap. Consequently, the guidance signal implemented in Eq. (11) is not the theoretically derived clean-distribution score.","section":"Sec. 3.2, Eqs. (8)-(11)"},{"comment":"On original (clean) CIFAR-10, applying SBDC degrades FID from 1.88 (oracle) to 3.19, yet the paper neither reports nor analyzes this failure mode. This is important because it shows the method is not universally beneficial and can harm performance when the training labels are already clean. The limited-interval analysis in Sec. 4.5 does not address clean-data behavior, and the conclusion that SBDC 'consistently outperforms' baselines is not accurate for this setting.","section":"Table 3"},{"comment":"The theoretical claim assumes an optimal discriminator, but in practice the discriminator is trained on pseudo-labels from a noise detector. Table 9 shows that for asymmetric noise, CORES recall is only 58.5%, meaning many true corrupt samples are treated as clean. The paper acknowledges the importance of recall but does not analyze how imperfect pseudo-labels affect the correction term or provide any robustness bound. This disconnect between the idealized theory and the practical implementation is not addressed.","section":"Sec. 3.4 and Table 5"}],"minor_comments":[{"comment":"The phrase 'limiting the usage of our guidance to the early phase' contradicts Sec. 4.5, where the guidance interval is the middle phase (steps 8-16 of 18). Please correct the abstract.","section":"Abstract"},{"comment":"The sentence 'we set Sclip min = 1.5 and Sclip min = 50.0' should read 'Sclip max = 50.0'.","section":"Sec. 4.5"},{"comment":"The notation 'log p^phi_theta(x_t)' and 'log D^t_phi(x_t)/(1-D^t_phi(x_t))' is typeset inconsistently; the superscripts are misplaced. Please clarify.","section":"Eq. (5)"},{"comment":"Reported Density values are around 100, whereas the standard Density metric of Naeem et al. is typically O(1). Please clarify the scaling or the metric definition used.","section":"Tables 1 and 5"},{"comment":"In lines 5-7, the comparison 'y_t^prev != y_t^prev' is self-referential; the intended comparison is between consecutive steps. Please fix the indexing.","section":"Algorithm 4"},{"comment":"The paper claims the method is 'derived' from theory, but the proof of Theorem 1 is in the supplementary material (Appendix A). Please indicate in the main text that the proof is provided in the appendix.","section":"Introduction"}],"recommendation":"reject","confidential_remarks":"The empirical results are extensive and the method may work as a heuristic, but the central theoretical claim is unsupported in a way that cannot be fixed by a local proof repair: the discriminator is estimating a different quantity than the required score correction. The authors would need to either change the method or fully reframe the paper as an empirical study, which is beyond a normal revision. I would not recommend acceptance without a substantially corrected theory."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper has a genuinely useful practical idea—train a small discriminator on noise-detector-filtered samples and use its gradient to steer a frozen conditional diffusion model during the early/middle sampling phase—and the class-wise results across CIFAR-10/100, Tiny-ImageNet, FOOD101, and Clothing-1M are consistent enough to take seriously. But the theoretical proof in Appendix A is not just sloppy; it's wrong, and the central claim that the discriminator gradient recovers the clean score does not follow.\n\nWhat I like: the method is cheap, requires no retraining of the diffusion model, and the limited-interval gate is well motivated by their confidence/instability analysis. The ablations show SiMix and Pseudo-clean Shuffle help, and the gains on class-wise metrics (CW-FID, CW-Density, CW-Coverage) show up in almost every setting they tried. That is a real empirical signal.\n\nThe problems: Theorem 1 doesn't prove what the paper says. In the Appendix A derivation, the term with the + sign is treated like the usual denoising score matching target, which it isn't; the cancellation that gets to Eq. 19 doesn't hold. The final bound is satisfied by a zero discriminator, so it's vacuous. Independently, the discriminator trained per Eq. (9) estimates the log-odds of the pseudo-clean label given (xt, y~), not the ratio log p(xt|y)/p_theta(xt|y~). Those coincide only in degenerate cases. So Eq. (8) is best read as a heuristic correction, not a derivation of the clean distribution. The empirical results are also mixed on global FID—on CIFAR-10 SBDC makes FID worse than the unmodified EDM in every noise setting they report, even while class-wise metrics improve. The abstract's 'superiority over previous state-of-the-art' is generous given that. Several tables (CIFAR-100, real-world, ImageNet) have no error bars, and no code is released.\n\nWho this is for: researchers working on noisy-label conditional generation, and anyone thinking about inference-time corrections to frozen diffusion models. It is a useful proof of concept and a nice example of a paper where the empirical story is stronger than the theory.\n\nRecommendation: send it to review, but with a clear message that the theoretical section needs to be removed or completely redone, and the empirical claims need to be calibrated to what the metrics actually show. If the authors resubmit with a heuristic framing and released code, this could be a solid contribution.","headline":"A promising inference-time fix for noisy-label diffusion with a broken theoretical appendix; worth engaging with as a heuristic, not as a proven method.","tokens_in":23599,"tokens_out":7688,"would_cite":false,"duration_ms":64250,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A discriminator trained on clean versus corrupted image-label pairs can steer a noisy conditional diffusion model back to the intended class-conditional distribution at sampling time, without retraining the model.","keywords":["diffusion models","noisy labels","score-based guidance","discriminator correction","conditional generation","inference-time correction","noise detection","class-conditional quality"],"falsifier":"Take a diffusion model trained on clean CIFAR-10, corrupt its labels with a known transition matrix, and compare the true correction $\\nabla_{x_t}\\log(p(x_t|y)/p_\\theta(x_t|\\tilde{y}))$ against the discriminator gradient on a fixed batch of intermediate samples; if their cosine similarity is near zero, the approximation in Theorem 1 is not what the method is actually computing. A simpler behavioral test is to measure class-conditional accuracy (CAS) and FID with the gate off: if the correction is working, CAS should rise while FID stays roughly constant.","tokens_in":22559,"feed_emoji":"🎨","tokens_out":9441,"duration_ms":78880,"temperature":0.7,"pith_summary":"The paper claims that when a conditional diffusion model is trained on labels that are partly wrong, its score estimate—the gradient of the log-density that drives the reverse process—can be corrected during sampling by adding the gradient of a discriminator that separates clean image-label pairs from corrupted ones. That correction is derived as the log-likelihood ratio between the clean and noisy conditional distributions, so the discriminator acts as a learned stand-in for the missing piece of the score. The authors further claim that applying this correction only in the middle phase of the sampling trajectory, where class decisions are most unstable, improves class-conditional fidelity and diversity and keeps the added inference cost small. If true, the result is a practical way to make large pre-trained conditional generators robust to label noise without retraining them.","feed_headline":"One discriminator fixes noisy-label diffusion outputs at sampling","feed_subtitle":"Trained on clean versus corrupted pairs, it corrects class-conditional scores during sampling at near-zero extra cost.","key_machinery":"The load-bearing identity is Eq. (8), the score-correction decomposition $\\nabla_{x_t}\\log p(x_t|y)=\\nabla_{x_t}\\log p_\\theta(x_t|\\tilde{y})+\\nabla_{x_t}\\log\\frac{p(x_t|y)}{p_\\theta(x_t|\\tilde{y})}$, which isolates the mismatch caused by noisy labels. The mechanism that realizes it is a small time-dependent discriminator $D^t_\\theta(x_t,\\tilde{y})$ trained with binary cross-entropy on clean and corrupted samples, whose logit gradient approximates the likelihood-ratio term; a $\\gamma$-gate restricts this guidance to the middle 'conditional phase' where class assignments are unstable, and two augmentations—Pseudo-clean Shuffle, which flips labels of clean samples to balance the corrupt set, and SiMix, a similarity-based input mixing scheme—stabilize the discriminator's training and gradients.","core_discovery":"The central claim is that the clean conditional score can be recovered from a noisy-label score model by adding a correction term: $\\nabla_{x_t}\\log p(x_t|y)=\\nabla_{x_t}\\log p_\\theta(x_t|\\tilde{y})+\\nabla_{x_t}\\log\\frac{p(x_t|y)}{p_\\theta(x_t|\\tilde{y})}$. The paper trains a time-dependent discriminator on pseudo-clean and pseudo-corrupt pairs, filtered from the noisy training set by an off-the-shelf noise detector, and uses its logit gradient $\\nabla_{x_t}\\log\\frac{D}{1-D}$ as the learned approximation to that intractable ratio. Theorem 1 states that for an optimally trained Lipschitz discriminator, the expected squared error between the discriminator gradient and the true correction is bounded by a constant plus the norm of the correction itself; the empirical sections then show that gating this correction to an early-to-middle interval of the reverse process improves class-wise FID, density, and coverage across symmetric, asymmetric, instance-dependent, and real-world label noise, with almost no increase in inference time.","pith_inferences":["The same likelihood-ratio correction could be applied to text-conditioned diffusion models by training the discriminator on matched versus mismatched text-image pairs rather than clean versus corrupted labels, which the paper itself names as future work.","The Theorem 1 bound depends on the magnitude of the true correction, so at very high noise rates the approximation can degrade; the more informative bottleneck may be the noise detector's recall, which the paper reports dropping to 58.5 percent on asymmetric CIFAR-10.","The gate interval could be predicted from the paper's confidence and instability curves instead of tuned per dataset, which would make the method easier to deploy on new domains.","A direct test of the mechanism would be to compute the true clean-to-noisy score ratio on held-out samples with a known noise transition matrix and compare its cosine similarity with the discriminator gradient."],"forward_implications":["Pre-trained conditional diffusion models can be made robust to label noise at inference time, with only a small discriminator and no retraining of the generative model.","Applying the discriminator correction only during the early-to-middle sampling phase improves class-conditional metrics while adding only marginal wall-clock time to generation.","The correction improves class-wise fidelity and diversity even at high noise rates, where the authors report the baseline and prior method produce many class-incorrect images.","Because the method is plug-and-play, it can be stacked on top of other noisy-label diffusion training schemes; the authors show it also improves models trained with the transition-matrix method they compare against.","The practical ceiling of the correction depends on the noise detector's recall, so better detection of corrupted labels should translate directly into stronger generative correction."],"supporting_citations":[{"why":"Supplies the discriminator-guidance technique of using a discriminator's logit gradient to refine a score model, which the correction step extends.","marker":"[24]"},{"why":"The transition-matrix-based noisy-label diffusion baseline that SBDC must beat and the main experimental comparison point.","marker":"[34]"},{"why":"Provides the EDM score-model architecture used as the pre-trained base model in the experiments.","marker":"[23]"},{"why":"Provides a noise detector used to split the training set into pseudo-clean and pseudo-corrupt subsets.","marker":"[11]"},{"why":"Provides the alternative noise detector used in the ablation study.","marker":"[36]"},{"why":"Shows that applying guidance in a limited interval improves sample quality, motivating the gamma-gate schedule.","marker":"[29]"},{"why":"Supplies the classifier-guidance form used in the gated correction equation.","marker":"[14]"}],"fun_headline_variants":["Discriminator fixes noisy-label diffusion at sampling","Score-based discriminator corrects noisy label diffusion","Sampling-time fix for noisy-label diffusion models","Discriminator steers diffusion away from label noise","Early discriminator boost for noisy-label diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the noise detector's clean/corrupt split reflects the true labels; if that split is biased, as the paper's own asymmetric-noise recall of 58.5 percent suggests, the discriminator correction steers samples toward the detector's notion of clean rather than the true clean distribution.","fun_headline_variants_meta":{"raw":{"variants":["Discriminator fixes noisy-label diffusion at sampling","Score-based discriminator corrects noisy label diffusion","Sampling-time fix for noisy-label diffusion models","Discriminator steers diffusion away from label noise","Early discriminator boost for noisy-label diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000522,"raw_usage":{"total_tokens":2518,"prompt_tokens":928,"completion_tokens":1590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":1519}},"tokens_in":544,"tokens_out":1590,"duration_ms":11196,"temperature":1.0,"reasoning_tokens":1519,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:50:21.529626+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a diffusion model trained on clean CIFAR-10, corrupt its labels with a known transition matrix, and compare the true correction $\\nabla_{x_t}\\log(p(x_t|y)/p_\\theta(x_t|\\tilde{y}))$ against the discriminator gradient on a fixed batch of intermediate samples; if their cosine similarity is near zero, the approximation in Theorem 1 is not what the method is actually computing. A simpler behavioral test is to measure class-conditional accuracy (CAS) and FID with the gate off: if the correction is working, CAS should rise while FID stays roughly constant.","supporting_citations":[{"cited_title":"Refining generative process with discriminator guidance in score-based diffusion mod- els","cited_arxiv_id":null,"evidence_quote":"Supplies the discriminator-guidance technique of using a discriminator's logit gradient to refine a score model, which the correction step extends."},{"cited_title":"Label- noise robust diffusion models","cited_arxiv_id":null,"evidence_quote":"The transition-matrix-based noisy-label diffusion baseline that SBDC must beat and the main experimental comparison point."},{"cited_title":"Learning with instance-dependent label noise: A sample sieve approach, 2021","cited_arxiv_id":null,"evidence_quote":"Provides a noise detector used to split the training set into pseudo-clean and pseudo-corrupt subsets."},{"cited_title":"Confident learning: Estimating uncertainty in dataset labels","cited_arxiv_id":null,"evidence_quote":"Provides the alternative noise detector used in the ablation study."}],"review_version":2}