{"id":"16361549-6cbb-4355-9f7a-1f353f5b5a64","arxiv_id":"2507.08979","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PRISM debiases CLIP by using an LLM to generate biased scene descriptions and then learning a linear projection of the embedding space that reduces spurious correlations, yielding higher worst-group accuracy on Waterbirds and CelebA.","lead":"PRISM is a new method that debiases CLIP by asking a large language model to list spurious attributes for each class, then learning a projection of the embedding space that makes embeddings less sensitive to those attributes. It reports state-of-the-art worst-group accuracy on Waterbirds and CelebA without using external image data, which matters because such a cheap, task-agnostic debiasing step could be applied before deploying zero-shot classifiers.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 7 defines LLD with no dependence on the learned projection P, so Algorithm 1's gradient update is ill-posed as written; this must be clarified before the central claim can be assessed.","rationale":"The reader's weakest_assumption is the text-to-image transfer hypothesis, an important empirical risk. However, the single most load-bearing issue is closer to the core: the paper's training objective, Eq. 7, does not mathematically depend on the only trainable parameter P. If taken literally, Algorithm 1 cannot learn anything, making the strongest_claim 'PRISM learns a linear projection P ... with the LD loss (Eq. 7)' formally undefined. This is not an attack on the authors or an external-consensus dispute; it is an internal inconsistency in the written method. The empirical results are strong and the code is public, so the issue is eminently checkable. If the code confirms P is applied inside the loss, the paper needs only a correction to Eq. 7 and the reader's original transfer concern regains prominence; if the code does not apply P, the reported results are unexplained by the stated method. I therefore recommend keeping the conditional verdict: the paper should be accepted only after the loss definition is clarified and verified against the released implementation. I also flag the 'Table X' placeholder in Appendix C as a minor completeness issue, in line with the instruction to surface missing references. My disagreement with the reader's weakest_assumption is not a dismissal of the transfer concern; it reflects that the optimization-objective gap is logically prior and, if unresolvable, would invalidate the central claim even before considering modality transfer.","tokens_in":17041,"tokens_out":7565,"duration_ms":95632,"concrete_test":"Inspect the public repository (https://github.com/MahdiyarMM/PRISM) training loop for Stage 2: verify whether the learnable projection P is applied to phi_T(T) before computing the cosine similarities that enter Eq. 7. Concretely, confirm that a gradient with respect to P flows into the loss (e.g., the code computes cos_sim(P(phi_T(T_a)), P(phi_T(T_b))) rather than cos_sim(phi_T(T_a), phi_T(T_b))). As a complementary check, set P = I (identity) at inference and re-evaluate Waterbirds WG; if WG stays near 84.2% rather than falling to the zero-shot baseline 36.4%, the reported effect is not attributable to the learned projection. If the code applies P inside the loss, then Eq. 7 needs a correcting revision; if it does not, the paper's central algorithm is not executable as described.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism of PRISM is the learned linear projection P, trained by minimizing the Latent space Debiasing loss LLD (Eq. 7) on LLM-generated scene descriptions (Algorithm 1, steps 6-8). However, as written in Eq. 7, every term is an inner product of the form <phi_T(T_{a,y}), phi_T(T_{a',y'})>, with no P appearing anywhere in the loss. Consequently, for any projection P, LLD(P) = LLD(I), so grad_P LLD = 0 identically. Taken literally, the optimization loop in Algorithm 1 cannot update P, and the reported improvements (Waterbirds WG 36.4% to 84.2%) cannot arise from the described training objective. This is an internal inconsistency in the method's formal definition, and it is more fundamental than the text-to-image transfer assumption: even if transfer holds, there is no well-defined objective by which the projection is learned. The most plausible resolution is that the authors intend the inner products in Eq. 7 to be computed after applying P, i.e., <P(phi_T(T_{a,y})), P(phi_T(T_{a',y'}))>, but the paper does not state this. The released code should settle the issue. Separately, Appendix C contains an unresolved 'Table X' reference, a completeness defect but not a correctness threat, and Table 2 validates the text-image bias-transfer hypothesis only on waterbird-style scene descriptions, not on CelebA, leaving the transfer assumption thinner than the headline results imply.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PRISM, a data-free method to reduce spurious correlations in CLIP's embedding space. In Stage 1 an LLM generates scene descriptions conditioned on class labels and discovered spurious attributes; in Stage 2 a linear projection P is learned by minimizing a Latent space Debiasing loss (Eq. 7) over text embeddings of these descriptions, and zero-shot classification is performed by applying P to both image and text embeddings (Eq. 8). A lighter variant, PRISM-mini, removes spurious directions via an orthogonal projection. Experiments on Waterbirds and CelebA with CLIP-ViT-L/14 report worst-group accuracy 84.2% and 84.0%, respectively, outperforming data-free baselines, with additional RN50 results in Appendix C.","tokens_in":17376,"tokens_out":4224,"duration_ms":49523,"significance":"If the mechanism is as claimed, PRISM would be a valuable contribution: it avoids predefined bias categories and external image data, uses only LLM-generated text during training, preserves overall accuracy, and ships public code. The reported gains are large and the design is simple to reproduce. However, the formal definition of the loss and the evidence for text-to-image transfer need strengthening before the empirical claims can be assessed.","major_comments":[{"comment":"As written, every term in LLD is an inner product of the form <phi_T(...), phi_T(...)>, with no occurrence of the learned projection P. Therefore LLD is constant in P and the gradient update in Algorithm 1 step 8 is identically zero; the reported WG improvements cannot arise from the described objective. Please define the loss on the projected embeddings, e.g. replace phi_T(...) with P(phi_T(...)) throughout Eq. (7), and align Algorithm 1 and Figure 2 with this definition.","section":"Section 3.2, Eq. (7), Algorithm 1"},{"comment":"The paper's transfer hypothesis (Section 3.1) is that spurious correlations in text mirror those in images, so a projection trained on text embeddings will debias image embeddings. Table 2 only demonstrates that raw CLIP text embeddings of scene descriptions exhibit a similar worst-group gap to image embeddings; it does not test whether the learned P transfers to images. Please add a direct experiment: train P on scene descriptions only, then apply it to image embeddings and report WG/Acc on the actual benchmarks, including CelebA.","section":"Section 5, \"Bias text\", Table 2"},{"comment":"The margin m and the number of scene descriptions are tuned on worst-group accuracy, but no validation protocol is described; if these values are selected using the test WG, the headline numbers are partially fitted. Please specify how m and the number of descriptions were chosen, whether a validation split was used, and report per-seed WG/Acc variability for the main results.","section":"Section 4, Figures 5 and 6"}],"minor_comments":[{"comment":"Appendix C contains an unresolved \"Table X\" reference in the comparison with BendVLM; please replace it with the actual table number (Table A2).","section":"Appendix C"},{"comment":"The symbol A is used both for the set of spurious attributes and for the matrix of their embeddings; please use distinct notation to avoid ambiguity.","section":"Section 3.2, Eqs. (9) and (10)"},{"comment":"With Llama 3.2, the CelebA WG of 58.9% is below the zero-shot baseline of 72.8%, so PRISM is not uniformly beneficial; a sentence acknowledging this failure mode would improve the presentation.","section":"Table 3"},{"comment":"The abstract's claim that PRISM outperforms current debiasing methods on Waterbirds and CelebA is not true for the RN50 backbone on Waterbirds, where FairerCLIP and Orth-Cali achieve higher WG; please qualify the claim by backbone.","section":"Abstract and Appendix C, Table A1"},{"comment":"Reference [24] concerns object hallucination and does not directly support the claim that spurious correlations in CLIP's text domain mirror those in the image domain; please cite a more appropriate source.","section":"Section 5, \"Bias text\", reference [24]"},{"comment":"There are several typographical errors, including \"orthogonlizing\" (Section 3.2), \"attibutes\" (Appendix B), and \"effectivenss\" (Appendix C); a careful proofread is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the empirical results are strong, but the formal inconsistency in Eq. 7 must be resolved before the central mechanism can be evaluated. The fix appears localized, so I recommend major revision rather than rejection. It would also be useful to check the released code to confirm which loss was actually optimized during training."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. The first is good: PRISM is a genuinely new combination—an LLM generates scene descriptions that expose spurious correlations, and a linear projection of CLIP's embedding space is trained (allegedly) with a contrastive-style loss on those text descriptions alone. If it works, it is a cheap, task-agnostic debiasing tool that competes with methods that need labeled images. The reported numbers are large: Waterbirds worst-group from 36.4 to 84.2, CelebA from 72.8 to 84.0, with code public.\n\nThe second thing is not good: Eq. 7, the loss that is supposed to train the projection, has no P in it. Every term is an inner product of text embeddings like <phi_T(T), phi_T(T')>. The gradient of that loss with respect to P is identically zero, so Algorithm 1's 'update P via gradient descent' is ill-posed as written. I suspect the authors intend to apply P before taking inner products—the figure and the text support that reading—but the paper never says so. This is a central formal flaw, not a typo in a footnote, and it has to be fixed before the results can be taken at face value.\n\nThe text-to-image transfer assumption is thinner than the headline. The paper hypothesizes that biases in CLIP's text embeddings mirror biases in image embeddings, and checks this with a single classification experiment on bird scene descriptions (Table 2), not on CelebA. That is suggestive, not conclusive. It is the load-bearing bridge between training on text and debiasing images, and it deserves more evidence.\n\nThere are also standard experimental hygiene issues: the margin m and the number of scene descriptions are studied in sensitivity plots, but it is not clear those choices are made on a validation set rather than the test set, and no variance is reported despite three seeds. The Appendix has an unresolved 'Table X' reference—minor, but sloppy.\n\nTo be fair, the central claim is not circular in the obvious way: the projection is learned from LLM-generated text, not from image labels, so the worst-group gains are not fitted values. And the cited prior work is handled honestly.\n\nMy verdict: the idea deserves a serious referee, and I would send it out. But the current version needs major revision. Fix Eq. 7, validate the transfer hypothesis on a second dataset, and report selection and variance honestly. Right now the paper is a promising sketch with an undefined training signal.","headline":"A promising data-free debiasing idea undercut by a central equation that, as written, does not depend on the learned projection; likely a typo, but must be fixed before the claims can be assessed.","tokens_in":17883,"tokens_out":3061,"would_cite":false,"duration_ms":35943,"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":"PRISM claims a text-trained projection removes CLIP's spurious biases, lifting worst-group accuracy from 36.4% to 84.2% (Waterbirds) and 72.8% to 84.0% (CelebA).","keywords":["spurious correlations","vision-language models","CLIP","LLM-guided debiasing","embedding projection","contrastive loss","worst-group accuracy","zero-shot classification"],"falsifier":"Take a dataset whose known spurious correlation has no natural textual counterpart—for example, classify two bird species where the confound is image resolution or lighting angle, attributes an LLM prompted for scene descriptions will not name—and run PRISM exactly as described. If worst-group accuracy does not improve over the zero-shot baseline, the text-to-image transfer hypothesis fails. A cheaper check on Waterbirds: strip the words for the spurious attributes (land, water, and their synonyms) out of the LLM-generated scene descriptions and retrain the projection; if the worst-group gain persists, the LD loss is removing something other than the stated spurious signal.","tokens_in":16882,"feed_emoji":"⚖️","tokens_out":13693,"duration_ms":122118,"temperature":0.7,"pith_summary":"PRISM claims that CLIP's tendency to lean on spurious correlations—a bird's background, a face's incidental context—can be largely neutralized without retraining the model, collecting new images, or pre-specifying which biases to hunt for. The recipe is to prompt a large language model with plain class names, get back scene descriptions that pair each class with its stereotyped surroundings, encode those descriptions with CLIP's text encoder, and learn a single linear projection of the shared embedding space using a contrastive loss that pulls same-class descriptions together and pushes different-class descriptions apart. The same projection, applied to image and text embeddings at test time, is reported to raise worst-group accuracy (the accuracy of the model's most poorly predicted subgroup) from 36.4% to 84.2% on Waterbirds and from 72.8% to 84.0% on CelebA, while overall accuracy stays roughly level. If the claim holds, it means debiasing a vision-language model can be done in a data-free, task-agnostic way: any new task needs only an LLM call and one short projection-training pass.","feed_headline":"Text-only projection lifts CLIP worst-group accuracy from 36% to 84%","feed_subtitle":"PRISM learns the map from LLM-written scene descriptions and applies it to images, keeping overall accuracy intact.","key_machinery":"The central object is the learned linear projection $P$ of CLIP's shared embedding space, trained on text only. It is driven by two components: an LLM that, given simple class prompts such as 'A photo of a duck', returns likely spurious attributes (e.g., a lake or a jungle background) and then generates controlled scene descriptions per group; and the Latent space Debiasing loss (LD, Eq. 7), a contrastive-style objective whose intra-class, inter-attribute term minimizes $1 - \\langle \\phi_T(T_{a,y}), \\phi_T(T_{a',y}) \\rangle$ for same-class pairs with different attributes, and whose inter-class, intra-attribute term penalizes similarity beyond $m$ for different classes sharing an attribute. Optimizing $P$ on these text embeddings alone is claimed to strip the spurious subspace out of the shared representation, so applying $P$ to both image and text embeddings at test time debiases zero-shot classification without fine-tuning CLIP or using external data. A cheaper variant, PRISM-mini, skips optimization and sets $P$ by orthogonally projecting against the embedding matrix of the LLM-suggested attributes (Eq. 10).","core_discovery":"On the paper's own terms, the discovery is that spurious biases in CLIP occupy a subspace of the joint embedding space that can be found and excised using text alone. Because CLIP aligns images and text in one space, the paper argues that the same spurious correlations visible in images leave detectable traces in text embeddings: a picture of a duck over water and a picture of a duck in a jungle are encoded closer to 'water' or 'jungle' than to each other. PRISM harvests those traces by prompting an LLM with class prompts to generate scene descriptions for every (class, spurious-attribute) group, then learns a projection $P$ that minimizes the Latent space Debiasing loss (Eq. 7), which pushes embeddings of the same class across different attributes together and pushes embeddings of different classes sharing an attribute apart by a margin $m$. At inference the same $P$ is applied to both image and text embeddings (Eq. 8), so the alignment between modalities is preserved while the spurious direction is suppressed. The reported outcome is a large gain in worst-group accuracy on both benchmarks with essentially unchanged overall accuracy, and a representation map in which the four (class, background) groups form visibly separated clusters.","pith_inferences":["[Inference] If the text-to-image transfer hypothesis holds generally, the same projection recipe could debias other aligned multimodal models (for example SigLIP or ALIGN) by swapping the encoders, making LLM-guided projection a general debiasing primitive rather than a CLIP-specific fix.","[Inference] Because the LLM's world knowledge is the sole source of bias attributes, purely visual confounds that have no natural language counterpart—such as camera angle, resolution, or colour balance—would be invisible to PRISM; testing on a confound of that kind would probe the method's true boundary.","[Inference] The margin $m$ between different classes sharing an attribute is a measurable quantity before and after projection, so the LD loss could double as a lightweight bias meter that reports how much spurious signal remains in the embedding space after debiasing.","[Inference] The paper's single-epoch, batch-of-64 training recipe hints that the debiasing direction is a low-dimensional subspace; if a closed-form whitening or orthogonalization matches the gradient-trained $P$, optimization could be removed entirely without losing the worst-group gains."],"forward_implications":["Debiasing becomes a data-free, task-agnostic step: for any new set of class names, the user needs only an LLM call to generate scene descriptions and a single quick projection-training pass, with no images, no task labels, and no fixed list of bias categories.","Because only the projector $P$ is learned and CLIP's weights stay frozen, the method preserves CLIP's general zero-shot ability; the paper reports overall accuracy rising 4.3% on Waterbirds and dipping 0.7% on CelebA while worst-group accuracy jumps 47.8 and 11.2 points respectively.","PRISM beats all compared data-free debiasing methods on both benchmarks and outperforms the image-using baselines on Waterbirds worst-group accuracy, suggesting text-only guidance can substitute for image-side supervision when removing spurious correlations.","The main controls are the margin $m$ in the LD loss and the number of LLM-generated scene descriptions; the paper finds an optimal range for both, with $m = 0.6$ maximizing worst-group accuracy on Waterbirds and CelebA.","The choice of LLM sets the method's ceiling: in the paper's comparison, GPT-4o (the largest model tested) identifies spurious correlations best and yields the highest worst-group accuracy, while a smaller LLM like Llama 3.2 70B drops worst-group accuracy to 58.9% on CelebA."],"supporting_citations":[{"why":"CLIP itself; supplies the frozen image and text encoders whose shared embedding space PRISM projects.","marker":"[20]"},{"why":"GPT-4o system card; the LLM used to generate scene descriptions in the main experiments, whose bias knowledge sets the method's effectiveness.","marker":"[13]"},{"why":"Waterbirds dataset; the primary benchmark where worst-group accuracy rises from 36.4% to 84.2%.","marker":"[23]"},{"why":"CelebA dataset; the second benchmark where worst-group accuracy rises from 72.8% to 84.0%.","marker":"[18]"},{"why":"Introduces Orth-Proj and Orth-Cali, the prompt-based debiasing approach PRISM extends and the origin of the orthogonal projection used in PRISM-mini.","marker":"[5]"},{"why":"FairerCLIP, the strongest image-based debiasing baseline PRISM is compared against and outperforms on Waterbirds worst-group accuracy.","marker":"[7]"},{"why":"WiSE-FT, an image-based robust fine-tuning baseline that PRISM beats on Waterbirds worst-group accuracy.","marker":"[29]"}],"fun_headline_variants":["CLIP debiased with a projection learned from pure text","Text-only projection teaches CLIP to ignore spurious backgrounds","PRISM debiases CLIP using LLM-generated scene descriptions, no extra data","Data-free debiasing: LLM prompts teach CLIP to ignore spurious cues","PRISM projects out spurious correlations using only text prompts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise, which the paper states as a hypothesis in Section 3.1 and verifies only in the 'Bias text' experiment (Table 2), is that spurious correlations in images leave matching traces in the text domain, so a debiasing projection learned from text descriptions alone will also remove the bias from image embeddings.","fun_headline_variants_meta":{"raw":{"variants":["CLIP debiased with a projection learned from pure text","Text-only projection teaches CLIP to ignore spurious backgrounds","PRISM debiases CLIP using LLM-generated scene descriptions, no extra data","Data-free debiasing: LLM prompts teach CLIP to ignore spurious cues","PRISM projects out spurious correlations using only text prompts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001571,"raw_usage":{"total_tokens":6282,"prompt_tokens":969,"completion_tokens":5313,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":5219}},"tokens_in":585,"tokens_out":5313,"duration_ms":42552,"temperature":1.0,"reasoning_tokens":5219,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:08:05.713291+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset whose known spurious correlation has no natural textual counterpart—for example, classify two bird species where the confound is image resolution or lighting angle, attributes an LLM prompted for scene descriptions will not name—and run PRISM exactly as described. If worst-group accuracy does not improve over the zero-shot baseline, the text-to-image transfer hypothesis fails. A cheaper check on Waterbirds: strip the words for the spurious attributes (land, water, and their synonyms) out of the LLM-generated scene descriptions and retrain the projection; if the worst-group gain persists, the LD loss is removing something other than the stated spurious signal.","supporting_citations":[{"cited_title":"Learning transferable visual models from natural language supervision","cited_arxiv_id":null,"evidence_quote":"CLIP itself; supplies the frozen image and text encoders whose shared embedding space PRISM projects."},{"cited_title":"Distributionally robust neural networks for group shifts: On the 9 importance of regularization for worst-case gener- alization","cited_arxiv_id":null,"evidence_quote":"Waterbirds dataset; the primary benchmark where worst-group accuracy rises from 36.4% to 84.2%."},{"cited_title":"Deep learning face attributes in the wild","cited_arxiv_id":null,"evidence_quote":"CelebA dataset; the second benchmark where worst-group accuracy rises from 72.8% to 84.0%."},{"cited_title":"FairerCLIP: Debiasing CLIP's Zero-Shot Predictions using Functions in RKHSs","cited_arxiv_id":"2403.15593","evidence_quote":"FairerCLIP, the strongest image-based debiasing baseline PRISM is compared against and outperforms on Waterbirds worst-group accuracy."},{"cited_title":"Robust fine- tuning of zero-shot models","cited_arxiv_id":null,"evidence_quote":"WiSE-FT, an image-based robust fine-tuning baseline that PRISM beats on Waterbirds worst-group accuracy."}],"review_version":1}