{"id":"abe7e9d0-dc6f-4d44-9c07-382e1e4f0933","arxiv_id":"2501.11827","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"PXGen is a post-hoc, training-free explanation framework that scores anchor samples with intrinsic and extrinsic criteria, groups them by thresholds, and selects representative examples via k-dispersion or k-center.","lead":"PXGen is a proposed post-hoc explanation framework for generative models: it scores user-chosen images against the model using latent-space and reconstruction measures, groups them, and picks representative examples. The paper demonstrates it on MNIST and CIFAR-10 and claims that the images it flags as most representative are more influential for the model than those selected by the VAE-TracIn baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The representativeness claim in §4.4 is validated only against PXGen's own KLD+MSE grouping, with no external ground truth, seeds, or error bars; Figure 8 does not establish that PXGen finds influential training samples.","rationale":"The reader's weakest assumption pins the collapse on the ad hoc threshold and the fidelity of KLD/MSE to true representativeness. My reading supports that, but locates the most load-bearing problem one step later: even if the threshold were principled, the validation in §4.4 does not establish that the HIHE group contains the most influential samples, because the success criterion is defined by the same framework's grouping and no independent ground truth or statistical error control is provided. The absence of seeds and error bars makes Figure 8's qualitative trend unable to support the comparative claim against VAE-TracIn. The paper's framework is understandable and the qualitative phenomena in Figures 3-6 are interesting, but the central quantitative claim is not evidenced. This does not change the reader's REJECT verdict, so I set verdict_should_be to UNCHANGED; the concrete oracle-based test would convert the concern into either support or refutation.","tokens_in":9434,"tokens_out":5789,"duration_ms":61931,"concrete_test":"Run the §4.4 experiment with an external influence oracle: for the same VAE and 5,923-sample training set, rank each training sample by the change in FID of a retrained model after removing that sample (or small random groups with 10+ seeds), producing a top-K ground-truth set. Then check whether PXGen's HIHE set significantly overlaps this oracle top-K, and whether M-HIHE beats random removal outside a 95% CI. In the same run, sweep the §4.2 threshold (median, 90th percentile, average-of-max) and verify the Figure 8 ordering is stable. If PXGen's selection does not outperform random or the ordering flips with threshold, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that PXGen identifies the most representative training samples. The only evidence is Section 4.4, where representativeness is defined by PXGen itself: training samples are split into HIHE/LILE/etc. using the §4.2 threshold (average of per-iteration maximum KLD/MSE), and the HIHE group is asserted to be most representative. The validation then checks whether a model retrained on HIHE-only data has lower FID to the original model than models after random or VAE-TracIn-based removal. This loop is not conclusive: (i) 'representative' is operationalized as low KLD+MSE, the same quantity PXGen optimizes, so the grouping and the success metric are not independent; (ii) no external ground truth such as exhaustive retraining-based influence scores is used; (iii) no random seeds, confidence intervals, or repeated runs are reported, so the Figure 8 gap may be noise; and (iv) the threshold is an arbitrary high/low cut that can reorder groups under monotone transformations of KLD/MSE. A model trained only on low-KLD/low-MSE samples may be similar to the original model because the original model is dominated by such easy samples, not because PXGen tracks causal influence. Thus the comparison with VAE-TracIn may conflate distributional similarity of easy subsets with representativeness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PXGen, a post-hoc explainable framework for generative models. Given a trained model, an anchor set, and user-specified intrinsic/extrinsic criteria (e.g., KLD and MSE), PXGen computes per-anchor feature values, partitions anchors into four groups (HIHE, HILE, LIHE, LILE) using thresholds taken as the average of per-iteration maxima over model-generated samples, and selects characteristic anchors via k-dispersion or k-center algorithms. Demonstrations on a VAE trained on MNIST \"0\" and a Soft-IntroVAE trained on CIFAR-10 Automobile identify qualitative phenomena called \"model delusion\" and \"aligned conception.\" The central validation claim (Section 4.4, Figures 7 and 8) is that PXGen can identify the most representative training samples, and that removing them changes the model's generated-output distribution more than random removal or VAE-TracIn-based removal.","tokens_in":9740,"tokens_out":6251,"duration_ms":62929,"significance":"If the central representativeness claim could be supported, PXGen would offer a practical, black-box, training-phase-free XAI method for encoder-decoder generative models, with the attractive properties of customizable multi-criteria input, no training access, and tractable selection algorithms. Strengths of the paper include the explicit post-hoc framing, the flexibility of the anchor/criteria setup, and the attempt to compare against VAE-TracIn. However, the current evaluation is not sufficient to establish the claim: the validation is circular with respect to PXGen's own grouping criteria, the thresholds are ad hoc, no uncertainty quantification is reported, and key experimental constructions are ambiguous. These issues are load-bearing for the paper's central contribution.","major_comments":[{"comment":"The validation of the representativeness claim is circular. The training data are partitioned into HIHE and other groups using exactly the same KLD+MSE criteria that PXGen uses to define \"representative,\" and the validation then removes or retains those same HIHE samples. Without an external ground truth for influence or representativeness (for example, leave-one-out retraining influence, a held-out likelihood drop, or a metric not used in the grouping), the FID similarity between M-HIHE and the original model only shows that models trained on low-KLD/low-MSE samples resemble a model trained on all samples, which may simply reflect that the original model is dominated by easy samples. The comparison to random removal and VAE-TracIn does not break this circularity because the baseline removals are not matched in terms of the sample-difficulty distribution.","section":"Section 4.4, Figures 7 and 8"},{"comment":"The threshold that defines all four anchor groups is ad hoc: it is \"the average of the maximum values obtained from multiple iterations\" of model-generated samples. The paper does not report the number of iterations, the variance across iterations, or any justification for why the average of maxima is a meaningful high/low cut. Because the grouping changes under monotone transformations of KLD or MSE, the division into HIHE/HILE/LIHE/LILE is not invariant, and every downstream claim about model delusion, aligned conception, and representative samples depends on this arbitrary choice. Either the threshold should be derived from a principled statistical model, or the authors should demonstrate that the reported phenomena are robust across a range of thresholds.","section":"Section 4.2"},{"comment":"No error bars, multiple seeds, or statistical tests are reported for any of the FID comparisons. The training set for MNIST \"0\" consists of only 5,923 samples, and the difference between M-HIHE, M-Others, M-Random, and M-TracIn in a single run may be well within retraining noise. The paper should report mean FID over several independent training runs with standard deviations or confidence intervals, and ideally a significance test, before claiming that PXGen 'finds representative training samples well compared with the state-of-the-art.'","section":"Figures 7 and 8, Section 4.4"},{"comment":"The description of M-Others is ambiguous and potentially inconsistent. The sentence \"we trained a model M-Others using training data, but removed the training data from the HIHE group in an amount equivalent to that of the other groups, prioritizing the removal of low anchor values\" does not clearly specify which samples are retained, which are removed, or how the \"amount equivalent to that of the other groups\" is computed. Without this detail, the reader cannot determine whether M-Others is a fair baseline with the same training-set size and budget as M-HIHE, and Figure 7's comparison is uninterpretable.","section":"Section 4.4, M-Others construction"},{"comment":"The 5% analysis is not defined precisely: the text says \"we focus solely on the subset of anchors with a KLD value within the 5% range among all anchors,\" but it is not clear whether this means the lowest 5%, the top 5%, or values within 5% of some reference. This selection is also post hoc: it is applied only after observing patterns in the reconstructed images, and it is used to support the model delusion and aligned conception phenomena. The selection rule should be stated explicitly and justified, and the paper should indicate whether the qualitative conclusions change if the 5% is varied.","section":"Section 4.2, 5% analysis"}],"minor_comments":[{"comment":"Typographical issues: \"three phrases\" should be \"three phases\" (Section 3), \"explaination\" should be \"explanation\" (Section 3.1), and \"challenginging\" should be \"challenging\" (Section 4.4).","section":"Throughout"},{"comment":"The phrase \"M-TracIn is trained by removing the same amount of low-help (harmful) training data\" is unclear. VAE-TracIn produces influence scores, but \"low-help\" is not defined; please specify whether low-scoring, negative-scoring, or least representative samples are removed, and how that relates to \"harmful.\"","section":"Section 4.4"},{"comment":"The caption says \"Left\" and \"Right\" but does not explicitly state that the left panel corresponds to class \"0\" and the right panel to class \"5\"; please make this explicit in the caption.","section":"Figure 7"},{"comment":"The k-dispersion algorithm is described as identifying \"the k most distinct individuals,\" but k-dispersion in the cited literature is an optimization problem with different variants (max-min, max-sum). Please specify which objective is used and how ties are broken.","section":"Section 4.3"},{"comment":"The phrase \"Another part involves obtaining information from the model and mapping it to the same multi-dimensional space as the anchors\" is vague. Please define what information is obtained (e.g., generated samples, encoder outputs) and how it is mapped.","section":"Section 3.2"},{"comment":"The complexity claim \"the calculation time is O(n^2)\" is not self-contained. It should specify n (anchor-set size) and clarify whether this covers the criteria computation, the grouping, and the selection algorithms together, or only one of them.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible high-level idea, but the central evaluation is currently circular and lacks the quantitative rigor needed for a comparative claim. I see a path to revision: add an external ground-truth influence measure, report variance over seeds, fix the M-Others description, and either justify or sensitivity-test the threshold. If the authors cannot provide non-circular validation, the representativeness claim should be substantially weakened or removed. I would not consider acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read PXGen over the weekend. Quick take: the framework is a sensible, clearly written recipe for post-hoc explanation of encoder-decoder generative models, but the headline claim—that it finds representative training samples better than VAE-TracIn—is not supported by the evidence as presented.\n\nWhat's actually new: the combination of an intrinsic criterion (KLD between encoded anchor and the model's prior) and an extrinsic criterion (reconstruction MSE/FID) to define four anchor groups, then k-dispersion/k-center to pick representatives. That specific recipe is new in a narrow sense, and the qualitative phenomena ('model delusion,' 'aligned conception') are plausible observations worth following up. The problem is real: post-hoc methods that don't need training-phase access are scarce for generative models.\n\nThe soft spots are mostly in the validation. Section 4.4 defines representativeness as low KLD plus low MSE, the same quantity PXGen optimizes. Showing that a model trained only on the HIHE group is close to the original could just mean the original model is dominated by easy samples, not that PXGen tracks causal influence. There is no external ground truth—no exhaustive retraining-based influence scores or an independent definition of representativeness. Figures 7 and 8 have no error bars or seeds, so the apparent advantage over VAE-TracIn might be noise. The threshold (average of per-iteration maxima) is ad hoc and not invariant to monotone transformations of KLD/MSE. Also, the construction of M-Others is unclear, and group sizes aren't reported, so we can't tell whether M-HIHE is just training on the majority of data. On top of that, the 'first post-hoc XAI for generative models' claim is too strong, and the '100x faster' claim is asserted without a benchmark.\n\nThe stress-test note is fair. I don't think it misreads the paper.\n\nWho is this for? Someone working on XAI for generative models might find the framework a useful starting point, but they'd need to rebuild the validation from scratch. It's not ready as a citable result on representativeness. I'd still send it to reviewers rather than desk-reject: the problem is relevant, the method is understandable, and a good reviewer can demand the missing controls. But I would not accept it in its current form.","headline":"PXGen is a clear, sensible recipe for post-hoc explanation of VAEs, but the representativeness claim is validated circularly and without error bars, so the headline comparison to VAE-TracIn does not hold as presented.","tokens_in":10253,"tokens_out":4981,"would_cite":false,"duration_ms":49613,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A post-hoc method identifies the most representative training samples of an encoder-decoder generative model without accessing the training phase, using only reconstruction error and latent-space divergence.","keywords":["explainable AI","generative models","post-hoc explanation","variational autoencoder","representative samples","influence analysis","model delusion","anchor set"],"falsifier":"Retrain a small VAE many times, each time deleting exactly one training sample, and measure the FID shift of the generated distribution relative to the full model; then rank samples by PXGen's HIHE anchor values and check whether the top-ranked samples are enriched in the largest per-sample FID shifts. If the enrichment is no better than random selection, the central claim that HIHE identifies the most representative training samples is undermined.","tokens_in":9202,"feed_emoji":"🎯","tokens_out":9755,"duration_ms":85280,"temperature":0.7,"pith_summary":"This paper proposes PXGen, a post-hoc explainable framework for generative models that works without access to training data or training dynamics. Given a model and a user-chosen anchor set, PXGen scores each anchor by two criteria: how far the model's encoded distribution for that sample is from the prior, and how well the model reconstructs the sample. A threshold derived from generated samples splits anchors into four groups, exposing two model behaviors: 'model delusion' (the model encodes confidently but reconstructs incorrectly) and 'aligned conception' (low latent affinity but accurate reconstruction). The paper's main empirical claim is that the high-affinity group contains the most representative training samples: models retrained only on that group generate images closer to the original model than random removal or VAE-TracIn-based removal, measured by FID. This gives a tractable, no-training-access route to auditing and curating generative models.","feed_headline":"Two scores reveal which samples a generative model relies on","feed_subtitle":"PXGen pinpoints the samples whose removal most changes what a generator outputs.","key_machinery":"The load-bearing mechanism is the anchor set plus an intrinsic and an extrinsic criterion. The intrinsic criterion is the KLD between the distribution the encoder assigns to an anchor and the model's prior $N(0,I)$; the extrinsic criterion is the reconstruction error between an anchor and its decoded version (MSE for MNIST, FID for CIFAR-10). Thresholds are computed as the average of the per-iteration maximum values over generated images, making the split depend on the model's typical outputs rather than on the anchor set. The high/low split yields the four groups HIHE, HILE, LIHE, and LILE, and $k$-dispersion or $k$-center then selects representative anchors from a group for visualization.","core_discovery":"On the paper's own terms, the discovery is that a black-box encoder-decoder generative model can be explained by measuring each candidate sample against two scalar criteria—KLD between the encoded latent distribution and the prior $N(0,I)$, and reconstruction error (MSE for simple images, FID for complex ones)—and then classifying samples as high or low on each axis. The four resulting groups carry distinct behavioral meanings: HIHE samples are what the model genuinely knows; HILE samples expose 'model delusion,' where the model believes it recognizes a concept but decodes it into something else; LIHE samples exhibit 'aligned conception,' where the model reconstructs them accurately despite low latent affinity; LILE samples are furthest from the model. The paper further claims that the HIHE group identifies the most representative training samples, validated by retraining experiments in which the FID distance to the original model is smallest when only HIHE data is kept, compared with random removal and with VAE-TracIn's influence scores.","pith_inferences":["If the two scalar criteria really capture representativeness, the same thresholding could be used before training as a data-quality filter, flagging samples that a model is likely to memorize poorly or reconstruct wrongly.","The framework's stated generality to GANs and diffusion models is a projection rather than a demonstrated result, since the intrinsic criterion used here is VAE-specific (KLD to a fixed prior); a testable extension would replace it with a latent-space distance defined for those architectures and repeat the removal experiment.","The model-delusion group suggests a direct test: an anchor the model encodes confidently but reconstructs wrongly should be the kind of sample that, if added to training data, most improves the model—this is a measurable prediction that the paper does not run.","The fuzziness of the 'average maximum' threshold implies that the groups are sensitive to how many generated images are sampled; one could make the method more robust by reporting group membership as a function of the sample size."],"forward_implications":["A deployed generative model whose training pipeline is proprietary can still be audited post hoc: PXGen points to concrete samples that most shape what the model generates, with no access to gradients, checkpoints, or training data.","The HIHE-versus-rest split gives a simple data-pruning recipe: keeping the high-affinity samples preserves a model's behavior, while removing them shifts behavior maximally.","The 'model delusion' and 'aligned conception' groupings turn a trained generator's latent behavior into inspectable categories, which could support debugging of learned biases and spurious concepts.","Because anchors need not be training data, users can probe a model with their own samples and see which concepts the model has internalized versus which it only reconstructs by accident.","PXGen's $O(n^2)$ cost per criterion makes it applicable at deployment time even for models whose training runs are too large to trace."],"supporting_citations":[{"why":"Supplies the variational autoencoder architecture that PXGen's intrinsic criterion (KLD to prior) is defined against and that the main demonstration uses.","marker":"[22]"},{"why":"Defines VAE-TracIn, the gradient-tracking baseline that PXGen must beat and claims to outperform in representative-sample selection.","marker":"[23]"},{"why":"Provides the Frechet Inception Distance metric used to measure similarity between original and retrained models, the basis of the validation.","marker":"[19]"},{"why":"Supplies Soft-IntroVAE, the second generative model used to demonstrate PXGen on CIFAR-10.","marker":"[6]"},{"why":"Supplies the k-dispersion algorithm used to pick the most distinct representative anchors from a group.","marker":"[18]"},{"why":"Supplies the k-center algorithm used to pick centrally representative anchors from a group.","marker":"[29]"},{"why":"Gives the Kullback-Leibler divergence used as the intrinsic criterion for measuring latent affinity.","marker":"[24]"},{"why":"Supplies the MNIST handwritten-digit benchmark used in the main demonstrations and in the representative-sample removal validation.","marker":"[7]"}],"fun_headline_variants":["Two scores reveal a generative model's hidden reliance on specific samples","PXGen explains generative models by categorizing samples on two axes","Post-hoc explainer for generative AI: Four sample types from two scores","Which samples does a generative model truly need? PXGen has a method","New post-hoc XAI categorizes generative-model samples into four roles"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a sample's importance to a model is faithfully summarized by two numbers—how far the model's encoded distribution for that sample sits from its prior, and how well the model reconstructs that sample—and that the boundary between 'high' and 'low' affinity is correctly placed at the average of the worst values achieved by the model's own generated samples; if that boundary is misplaced, the four groups and the representative-sample conclusion collapse.","fun_headline_variants_meta":{"raw":{"variants":["Two scores reveal a generative model's hidden reliance on specific samples","PXGen explains generative models by categorizing samples on two axes","Post-hoc explainer for generative AI: Four sample types from two scores","Which samples does a generative model truly need? PXGen has a method","New post-hoc XAI categorizes generative-model samples into four roles"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000952,"raw_usage":{"total_tokens":4090,"prompt_tokens":1007,"completion_tokens":3083,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":2991}},"tokens_in":623,"tokens_out":3083,"duration_ms":20036,"temperature":1.0,"reasoning_tokens":2991,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:48:45.558565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain a small VAE many times, each time deleting exactly one training sample, and measure the FID shift of the generated distribution relative to the full model; then rank samples by PXGen's HIHE anchor values and check whether the top-ranked samples are enriched in the largest per-sample FID shifts. If the enrichment is no better than random selection, the central claim that HIHE identifies the most representative training samples is undermined.","supporting_citations":[{"cited_title":"Advances in Neural Information Processing Systems 34, 2400–2412 (2021)","cited_arxiv_id":null,"evidence_quote":"Defines VAE-TracIn, the gradient-tracking baseline that PXGen must beat and claims to outperform in representative-sample selection."},{"cited_title":"Advances in neural information processing systems30 (2017) 14 Huang et al","cited_arxiv_id":null,"evidence_quote":"Provides the Frechet Inception Distance metric used to measure similarity between original and retrained models, the basis of the validation."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"Supplies Soft-IntroVAE, the second generative model used to demonstrate PXGen on CIFAR-10."},{"cited_title":"Operations research letters21(3), 133–137 (1997)","cited_arxiv_id":null,"evidence_quote":"Supplies the k-dispersion algorithm used to pick the most distinct representative anchors from a group."},{"cited_title":"Unpublished manuscript, March (2016)","cited_arxiv_id":null,"evidence_quote":"Supplies the k-center algorithm used to pick centrally representative anchors from a group."}],"review_version":1}