{"id":"26c0c69c-b4c8-49b0-818d-1ea6ff6fb5cf","arxiv_id":"2504.20902","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"C2B detects biases in visual classifiers by using an LLM to propose bias categories, retrieving unlabeled images for those categories, and measuring accuracy differences in the classifier.","lead":"This paper introduces a pipeline that guesses what a visual classifier might be biased about by asking a large language model, then pulling web images for those guesses and testing the classifier on them. It aims to let anyone audit a downloaded model for bias without needing labeled data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Retrieved-image validity is the load-bearing premise: the paper's own VQA analysis shows under 50% of retrieved images contain both the intended target and bias class, so bias scores computed on these pseudo-labels may reflect retrieval artifacts rather than classifier bias.","rationale":"Reviewing the central claim in good faith, I agree with the reader that the load-bearing step is the validity of the retrieval-based pseudo-labels entering Eq. (5) and being used to compute bias scores in Eq. (1). The paper is commendably transparent: Appendix A and Appendix F explicitly state that less than 50% of retrieved images contain both target and bias classes per VQA, and Appendix B shows that controlling retrieval (IMAGENET-GT) changes the results. This is not an external disagreement with consensus; it is an internal inconsistency between the strong claim that C2B 'automatically identifies its biases' and the reported quality of the data feeding the metric. On the other hand, the paper does provide a working pipeline, evaluates on two datasets, compares with a supervised baseline, and tests threshold sensitivity in Appendix C. The novelty is real and the honesty of the reporting is a credit. The concern does not justify rejecting the paper; it justifies the CONDITIONAL verdict already given by the reader. My proposed test would settle whether the top bias scores are stable when only correctly-retrieved images are used. If the concern lands, the required revision is to either filter retrieval by VQA or ground-truth confirmation before scoring, or report bias scores with uncertainty intervals; if it does not land, the current claims are empirically supported.","tokens_in":32447,"tokens_out":4095,"duration_ms":40385,"concrete_test":"On the ImageNet-X validation set, use ground-truth factor annotations to directly measure the accuracy of a tested model (e.g., ResNet50V2) on images that actually contain the target class and the bias attribute (e.g., 'birdhouse' plus 'rectangular'), and compute bias scores from this cleanly labeled subset. Then compare the resulting bias ranking and sign with C2B's ranking from unfiltered CLIP top-20 retrieval over CC12M for the same captions. If the top-10 detected bias classes do not substantially overlap, or any headline bias direction (e.g., birdhouse shape) flips, the reported bias detection is dominated by retrieval mislabels rather than classifier behavior. As a second arm, bootstrap the full-retrieval bias scores and report whether the top-scoring biases are significantly different from zero after multiple-comparison correction.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (5) assumes the top-k retrieved images are valid pseudo-labeled samples for the target-bias pair. Appendix F (Tables 17-18) reports VQA-confirmed 'both' accuracy of 46.5% (Bing, CelebA), 42.4% (CC12M, CelebA), 42.8% (Bing, ImageNet-X), and 38.1% (CC12M, ImageNet-X). Thus the majority of images used to compute bias scores in Eq. (1) do not contain both the intended target class and bias attribute. Accuracies computed over these images can differ across bias classes because of retrieval difficulty, image quality, or off-target content, not because of model bias. For example, 'spherical birdhouse' images may be rarer and harder to retrieve, producing a systematically different image distribution than 'rectangular birdhouse' images; a lower accuracy on the former is then a retrieval artifact, not a detected bias. Appendix B's controlled experiments weaken the counterargument that the method still works: even with target labels fixed (IMAGENET-GT), C2B's hit rate drops while precision rises, which is consistent with retrieval noise dominating the external-setting scores. No error bars or significance tests are provided, and with k=20 images per cell, the accuracy differences underlying phi are high-variance. The central claim therefore rests on an unvalidated premise that cannot be inferred from the reported GT-matching or VQA-agreement numbers alone.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Classifier-to-Bias (C2B), a pipeline for unsupervised bias detection in pretrained visual classifiers. Given only a textual task description, C2B uses an LLM to propose bias attributes and bias classes for each target class, generates captions for each target-bias pair, retrieves top-k images from CC12M or Bing via CLIP-style retrieval, and then computes per-class accuracy differences on this pseudo-labeled image set to score each proposed bias. The authors claim this is the first framework that detects biases without task-specific labeled data and without a predefined bias list. They evaluate on CelebA and ImageNet-X, comparing against the supervised B2T baseline using two evaluation protocols: matching detected biases to ground-truth annotations via SBERT similarity, and measuring agreement with VQA-generated bias pseudo-labels. The paper also includes ablations on retrieval accuracy, domain shift, threshold sensitivity, LLM choice, and retrieved-image diversity.","tokens_in":32739,"tokens_out":4477,"duration_ms":50616,"significance":"If the central claim held, C2B would fill a real gap: a user downloading a model could audit it with no labeled validation set and no predefined bias ontology. The paper has clear strengths in presentation and reproducibility: the code repository is public, the exact prompts are reported in Appendix I, and the supplementary material contains unusually extensive ablations, including controlled retrieval from labeled subsets (Appendix B), VQA-based retrieval checks (Appendix F), and threshold sensitivity analysis (Appendix C). However, the significance is currently limited by a load-bearing weakness: the pseudo-labeled retrieved images are only weakly validated, and the paper's own VQA analysis shows that fewer than half of retrieved images contain both the intended target and bias classes. Because the bias scores in Eq. (1) are computed over these images, the reported 'detected biases' may reflect retrieval artifacts rather than classifier behavior. The core idea is promising and modular, but the evidence as presented does not yet establish that C2B measures what it claims to measure.","major_comments":[{"comment":"The load-bearing premise of the method is that the top-k retrieved images in Eq. (5) form valid pseudo-labeled datasets D(b,y), so that the accuracy values in Eq. (1) measure classifier performance on images of class y exhibiting bias class b. The paper's own VQA evaluation contradicts this premise: the 'both' accuracy is 46.5% (Bing, CelebA), 42.4% (CC12M, CelebA), 42.8% (Bing, ImageNet-X), and 38.1% (CC12M, ImageNet-X). Thus, for a majority of target-bias cells, the retrieved images do not contain the intended combination, and the per-class accuracy differences in Eq. (1) can reflect retrieval difficulty or off-target content (e.g., 'spherical birdhouse' images being rarer and harder to retrieve) rather than classifier bias. The central claim therefore requires either a demonstration that bias scores are stable when computed on VQA-confirmed subsets only, or a correction/calibration procedure that accounts for retrieval noise.","section":"Sec. 3.4, Eq. (5); Appx. F, Tables 17-18"},{"comment":"The quantitative grounding for the main claim is weak in several connected ways. Ground-truth matching recovers only 10.8-12.3% of CelebA biases and 7.5-11.2% of ImageNet-X biases, with Detected->GT precision of 12.8-14.2% and 2.5-3.0% respectively. The controlled-retrieval experiments in Appendix B show that retrieving from the labeled ImageNet subset (IMAGENET-GT) lowers the ground-truth hit rate while raising VQA agreement, which is consistent with external-setting scores being dominated by retrieval noise. Moreover, all reported scores use k=20 images per cell, and no confidence intervals or significance tests are provided, so the accuracy differences behind phi are high-variance. The authors should report bootstrap or repeated-retrieval confidence intervals and show that the ranking of detected biases is stable under resampling, especially before claiming superiority over B2T.","section":"Sec. 4.2.1, Tables 1-2; Appx. B, Tables 5-6"},{"comment":"The discovery claim is constrained by the proposal mechanism in a way that is not fully accounted for. The pipeline proposes biases with an LLM, generates captions from those proposals, retrieves images for them, and then scores the same proposals; Appendix E shows the LLM misses roughly 67-71% of ground-truth biases, so the detected set is limited to what the LLM can verbalize. The 'new bias' percentages (68-83% on ImageNet-X) count biases absent from the annotation set, but the paper does not independently verify that these are real classifier biases rather than retrieval artifacts or semantically similar but off-target proposals. The authors should provide a validation step for novel detected biases, for example human evaluation on a sample of VQA-confirmed retrieved images, before interpreting the open-set results as evidence of discovering biases beyond the annotated sets.","section":"Sec. 3.3-3.5; Appx. E, Tables 15-16"}],"minor_comments":[{"comment":"The text says 'On ImageNet-X (Tab. 1),' but the ImageNet-X results are in Table 2; the cross-reference should be corrected.","section":"Sec. 4.2.1, paragraph after Tables 1-2"},{"comment":"The caption of Table 6 states 'Agreement between detected biases and VQA on CelebA,' but the table reports ImageNet-X results; this appears to be a copy-paste error.","section":"Appx. B, Table 6 caption"},{"comment":"The definition of Ay(f,b_i,j) is not fully specified: the authors should clarify whether this is the accuracy on the subset of retrieved images for that target-bias pair, and how the per-class accuracy is computed when the target class is an attribute (e.g., 'smiling' vs. 'not smiling') rather than a single category.","section":"Sec. 3.1, Eq. (1)"},{"comment":"The paper reports per-class LLM generation times in Appendix J but does not report the total wall-clock cost of running the full pipeline over all 1000 ImageNet classes and all retrieved images; a brief cost estimate would help readers judge practical usability.","section":"Sec. 4.1, Implementation details"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about retrieved-image validity is genuine and is explicitly acknowledged in the paper's own Appendix A. I would not reject the paper outright, because the framework is modular and the weakness is identifiable and potentially addressable by filtering retrieved images with VQA, restricting analysis to confirmed target-bias cells, and adding uncertainty quantification. However, as it stands, the headline claim that C2B 'identifies biases' is not supported by the evidence, and the authors need to either fix the measurement or substantially temper the claim. The 'first framework' novelty claim also deserves a careful comparison with annotation-free failure-mode discovery methods such as UDIS and Domino, which the paper cites but does not position against in the experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this paper defines a genuinely new problem—bias detection for visual classifiers with no labeled data, just a text description—and builds a sensible first pipeline for it. The evaluation, though, is not yet strong enough to support the central claim that the detected biases are real model biases. I'd send it to reviewers, but I'd expect major revisions.\n\nThe new piece is real: prior bias detectors need task labels or annotated failure sets; C2B only needs the task description and a pre-trained model. The pipeline is straightforward—LLM proposes bias attributes/classes, captions, retrieval, accuracy differences—and the authors are unusually honest about its limits. They include four evaluation angles (ground-truth matching, VQA agreement, retrieval-quality analysis, controlled retrieval), release code, and discuss failure cases. That's good practice.\n\nThe soft spot is exactly where the stress-test points: the retrieved images are the test set, and fewer than half of them actually contain both the intended target and bias class (Tables 17-18, acknowledged in Appendix A). With k=20 images per cell and no error bars, accuracy differences between bias classes can be driven by retrieval artifacts rather than model bias. The low absolute detection rates (10-14% of ground-truth biases, with false-hit rates about half as large) reinforce the concern. The 0.9 SBERT threshold and 0.05 bias threshold are empirically set; the results do hold across threshold variations, which helps a bit.\n\nThat said, the stress-test overstates one point. The controlled experiments in Appendix B show that when the target class is fixed (IMAGENET-GT), precision improves and VQA agreement rises, even though hit rate drops. That pattern is consistent with retrieval noise being a real factor, but it does not show noise dominates—the remaining detections are more precise, which is what you would expect if the method is capturing something real. The VQA-based evaluation on the original validation sets (with true task labels) is also a partial counterweight to the retrieval-quality problem. Still, the agreement scores there are low (0.22-0.32), so the evidence for real bias detection is suggestive, not conclusive.\n\nWho is this for? Anyone working on model debugging, fairness auditing, or retrieval-based evaluation. The paper is a useful scoping result and a cautionary example of how easy it is for retrieval noise to sneak into pseudo-labeled evaluations. I would not reject it; I would give it a conditional accept with the demand for error bars, a sensitivity analysis of k, and a small human-verified set of retrieved images to show the bias scores are not an artifact. And I would check whether the 'first' claim holds against a broader literature search, because that is the kind of claim reviewers should verify.\n\nRecommendation: send to peer review. The setting is new and the authors have done enough honestly to warrant referee time.","headline":"Genuinely new unsupervised setting, honest first pipeline, but retrieval quality and small samples keep the central claim from being proven; worth a serious referee.","tokens_in":33285,"tokens_out":3649,"would_cite":true,"duration_ms":40928,"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":"A pre-trained classifier's biases can be detected with no labeled data, using only a textual description of its task.","keywords":["unsupervised bias detection","visual classifier auditing","large language models","text-to-image retrieval","bias score","pseudo-labeled datasets","open-set bias discovery","model fairness"],"falsifier":"Take a classification task with ground-truth labels for both target classes and bias factors, run C2B on the unlabeled pool, and then recompute the bias scores using only the ground-truth-labeled images that satisfy each caption. If the ranking of biases changes substantially when retrieval noise is removed, the pipeline's outputs are driven by retrieval artifacts rather than by the classifier's actual per-group accuracy differences.","tokens_in":32243,"feed_emoji":"🔍","tokens_out":8890,"duration_ms":84932,"temperature":0.7,"pith_summary":"A person who downloads a pre-trained visual classifier usually has no labeled validation set and no ready-made list of failure modes. This paper proposes C2B, a framework that claims to recover the classifier's biases from nothing more than a plain-language description of the task it was built for. An LLM proposes candidate bias attributes and classes for each target class, converts them into captions like 'a photo of a rectangular birdhouse', and a retrieval engine pulls images for those captions from an unlabeled pool; accuracy gaps on those images become bias scores. On the two tested benchmarks the paper reports that C2B finds more ground-truth-annotated biases than a supervised open-set baseline, with lower miss rates, and that roughly two-thirds to four-fifths of the biases it detects are not present in the benchmark annotations. The framework is training-free, so the same pipeline can be re-run with stronger language models or retrievers as they become available.","feed_headline":"A text prompt alone uncovers a classifier's biases","feed_subtitle":"C2B generates bias hypotheses, retrieves test images, and scores accuracy gaps—all without task annotations.","key_machinery":"The object doing the work is the signed bias score of Eq. (1), which turns a retrieved image set into a number between $-1$ and $1$: for a fixed target class and bias attribute, it is the model's accuracy on images showing one bias class minus its average accuracy on the other bias classes of the same attribute. This score is what lets an unlabeled retrieval step yield a ranked list of biases, with positive values meaning the model does better on that bias class and negative values meaning it does worse. Two swappable components carry the rest: an LLM that proposes the bias vocabulary and generates captions, and a text-to-image retriever that supplies the pseudo-labeled images; in the reported experiments these are a lightweight instruction-tuned LLM, a contrastive language-image model, a 12-million-image corpus, and a web search engine. Because both components are task-agnostic and training-free, the same pipeline can be aimed at any classification task described in natural language.","core_discovery":"The central claim is that bias discovery for visual classifiers can be reformulated as a language-and-retrieval problem, removing the need for task-specific labels. C2B starts with a task description and an output class set; a large language model produces, for each class, a list of bias attributes and bias classes. A second prompting stage turns each target-bias pair into a caption, and a contrastive vision-language model retrieves the top-$k$ images for that caption from a large unlabeled database. Treating the captions as pseudo-labels, the paper computes the bias score $\\phi_{y,i,j}=A_y(f,b_{i,j})-\\frac{1}{n_i-1}\\sum_{k\\neq j}A_y(f,b_{i,k})$, the per-class accuracy gap between one bias class and its siblings. The paper reports that this pipeline outperforms a supervised open-set bias detector at recalling ground-truth biases on CelebA and ImageNet-X, while also flagging many biases absent from those annotated sets.","pith_inferences":["The paper's own VQA check shows fewer than half of retrieved images contain both the intended target and bias class; if a retriever with stronger compositional understanding replaces the current one, hit rates should rise and false-hit rates fall. This can be tested by re-running C2B's two evaluation protocols with a different retriever.","The top-down LLM proposal step means biases that are not linguistically expressible, or outside the LLM's knowledge, will likely stay invisible; a hybrid that clusters the classifier's errors on retrieved images first and then uses the LLM to name each cluster could combine this approach with bottom-up discovery.","Because the framework needs only a text description, it could in principle audit classifiers in specialized domains such as medical imaging, where the user has no labels and biases are tied to acquisition devices or hospital settings; the paper lists this as a direction but does not demonstrate it."],"forward_implications":["A non-expert can audit any downloaded classifier with a one-sentence task description, removing the annotation bottleneck that restricts bias detection to datasets someone has already labeled.","Bias detection is no longer limited to a predefined list: C2B proposes class-specific attributes, so it can name biases like 'camera angle: three-quarter' that benchmark annotations never record.","In the paper's experiments, C2B misses fewer known ground-truth biases than the supervised baseline (76.9–85.1% vs. 92.3–95.9% on the tested models and tasks), while detecting biases the annotations do not contain.","Because C2B is training-free and modular, improving the LLM or the retriever should translate directly into better bias discovery without retraining the pipeline."],"supporting_citations":[{"why":"Supplies the large language model that proposes bias attributes, classes, and captions from the task description.","marker":"[18]"},{"why":"Provides the embedding-based retrieval implementation used to score caption-image similarity.","marker":"[7]"},{"why":"Is the large unlabeled image database from which caption-matched images are retrieved.","marker":"[12]"},{"why":"Defines the contrastive vision-language model whose embedding similarity ranks images against captions.","marker":"[51]"},{"why":"Is the supervised open-set bias detector used as the baseline that C2B must match or beat without labels.","marker":"[33]"},{"why":"Provides the factor-annotated ImageNet validation set used as ground-truth bias evaluation on ImageNet.","marker":"[28]"},{"why":"Provides the face-attribute dataset whose remaining attributes serve as ground-truth biases for face tasks.","marker":"[41]"},{"why":"Is the face classifier whose biases the paper audits on CelebA.","marker":"[48]"}],"fun_headline_variants":["Bias hunting without labels: just a text prompt","No labels needed: LLM+retrieval expose classifier bias","Classifier bias detection goes annotation-free with C2B","C2B: unsupervised bias spotting via text and image retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the retrieved images really show the combination described in each caption: if 'a photo of a rectangular birdhouse' returns images that are not rectangular birdhouses, the accuracy gaps C2B reports measure retrieval noise rather than model bias, and the paper's own VQA check finds fewer than 50% of retrieved images match both the intended target and bias class.","fun_headline_variants_meta":{"raw":{"variants":["Bias hunting without labels: just a text prompt","No labels needed: LLM+retrieval expose classifier bias","Classifier bias detection goes annotation-free with C2B","C2B: unsupervised bias spotting via text and image retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00021,"raw_usage":{"total_tokens":1430,"prompt_tokens":986,"completion_tokens":444,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":602,"completion_tokens_details":{"reasoning_tokens":376}},"tokens_in":602,"tokens_out":444,"duration_ms":4378,"temperature":1.0,"reasoning_tokens":376,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:16:01.191807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a classification task with ground-truth labels for both target classes and bias factors, run C2B on the unlabeled pool, and then recompute the bias scores using only the ground-truth-labeled images that satisfy each caption. If the ranking of biases changes substantially when retrieval noise is removed, the pipeline's outputs are driven by retrieval artifacts rather than by the classifier's actual per-group accuracy differences.","supporting_citations":[{"cited_title":"Clip retrieval: Easily compute clip embeddings and build a clip retrieval system with them","cited_arxiv_id":null,"evidence_quote":"Provides the embedding-based retrieval implementation used to score caption-image similarity."},{"cited_title":"Discovering and miti- gating visual biases through keyword explanation","cited_arxiv_id":null,"evidence_quote":"Is the supervised open-set bias detector used as the baseline that C2B must match or beat without labels."},{"cited_title":"Imagenet-x: Understanding model mistakes with factor of variation annotations","cited_arxiv_id":null,"evidence_quote":"Provides the factor-annotated ImageNet validation set used as ground-truth bias evaluation on ImageNet."}],"review_version":1}