{"id":"ff2c84bf-dab5-4628-b230-b386b19b16ce","arxiv_id":"2505.18770","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DPSPG trains a negative-prompt branch alongside the normal positive prompt generator for CLIP, improving domain generalization accuracy and reducing prompt variability across random seeds.","lead":"This paper introduces DPSPG, a prompt generation method for vision-language models that trains separate positive and negative prompt branches to reduce seed-to-seed variability. It reports consistent gains over existing methods on five domain generalization benchmarks and gives a margin-based argument for why negative prompts help.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Margin-enlargement proof hinges on Eq. (11), which no training loss enforces and no experiment measures; if the negative-score separation δ does not survive on target domains, the claimed αδ margin gain and exponential gradient bound do not follow.","rationale":"The reader's conditional verdict is sound, and my independent read converges on the same load-bearing assumption. Eq. (11) is the only link between the negative-prompt mechanism and the claimed margin/gradient guarantees; without it the theoretical part reduces to a softmax-margin statement that would hold for any classifier. I sharpen the reader's point: the paper not only fails to verify δ on target domains, it never measures δ even on source data, and neither loss in the actual pipeline (BCE for v− and MSE for G−) is an explicit ranking or separability objective in score space. Thus the assumption is not just optimistic extrapolation; it is also unenforced. I note in the paper's favor that the empirical comparisons are extensive, the code is provided, the transformer generator is measurably faster than SPG (Table VI), and the ablation (Table V) is consistent with negative prompts helping. Those facts justify keeping the verdict conditional rather than rejecting. The proposed test—computing δ(x) on source and target from the released checkpoints—would settle whether the theoretical claims actually hold; if they fail, the paper's remaining contribution would be an incremental empirical method with small average gains and no error bars, which would shift the verdict toward reject or unverified.","tokens_in":18896,"tokens_out":5940,"duration_ms":58273,"concrete_test":"Using the released code, train DPSPG on each multi-source DG split and, on the held-out target domain, compute for every test image x the per-sample separation δ(x) = min_{i≠y} (s−_i(x) − s−_y(x)) using the generated negative prompts and CLIP text encoder. Report the median and the fraction of images with δ>0 on both source and target domains, and compare the actual margin Δ_i with the positive-only margin Δ+_i. If a substantial fraction of target images have δ≤0, or if the measured Δ_i is not at least Δ+_i, Eqs. (12) and (17) are not in force and the theoretical explanation is refuted. A sharper version: add a hinge loss to stage 2 that explicitly enforces s−_i(x) ≥ s−_y(x)+δ on source; then re-measure δ on target — this shows whether the assumption is a reliable property of negative prompts or an artifact of the current objectives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's theoretical centerpiece (§III-C) is inequality (11): for every false class i, s−_i(x) ≥ s−_y(x)+δ. Substituting (11) into (10) gives the margin gain Δ_i ≥ Δ+_i+αδ (Eq. 12), and Eq. 17 then claims an exponentially tighter gradient-norm bound. Everything downstream rests on δ>0. The paper justifies (11) with the phrase 'by designing the negative prompts,' but neither training stage in §III-B actually enforces it. The negative prompt labels v−,d are fit with BCE (Eq. 5), and the generator G− is fit by MSE to those labels (Eq. 6); neither is a margin or ranking objective in the space of the inference-time scores s−_i(x), and δ is not a parameter of either loss. As a result, on test images the ordering can invert: if s−_i(x) − s−_y(x) is negative for even one false class, the bound (12) reverses direction and negative learning shrinks the margin instead of enlarging it. This is not merely a target-domain extrapolation concern: the paper reports no measurement of δ on source or target data, and no ablation checks whether the inequality holds for G−'s outputs rather than for the label prompts. Given that the theoretical claims are the paper's stated explanation for both stability and generalization, Eq. (11) is load-bearing and unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Dual-Path Stable Soft Prompt Generation (DPSPG), a two-stage method for domain generalization with CLIP-based prompt learning. In the first stage, domain-specific positive and negative soft-prompt labels are learned with cross-entropy and binary cross-entropy losses. In the second stage, two transformer-based generators produce instance-specific positive and negative prompts, trained by MSE alignment to those labels. At inference, the logit for class i is the positive-prompt similarity minus α times the negative-prompt similarity. The authors argue that this dual-path design reduces 'Prompt Variability' across random seeds, and they provide a theoretical analysis claiming that negative prompts enlarge the decision margin by at least αδ and tighten a gradient-norm bound. Experiments on PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet with ResNet-50 and ViT-B/16 backbones show consistent average accuracy improvements over prior prompt-based DG methods, along with lower standard deviations on several stability metrics. The code is publicly available.","tokens_in":19245,"tokens_out":5300,"duration_ms":47195,"significance":"If the empirical claims hold, DPSPG is a useful and well-engineered extension of dynamic prompt learning for DG: it introduces negative prompts inside a transformer-based generator, addresses a real phenomenon (seed-dependent prompt variability), and validates the design with ablations and stability analyses. The paper is commendable for releasing code, evaluating on five standard benchmarks with two backbones, and reporting training-time efficiency and stability comparisons. However, the theoretical centerpiece is not established: the margin-enlargement result is conditional on an asserted inequality that no training objective enforces and no experiment measures, and the 'gradient norm stability' bound concerns the input-output Jacobian rather than the parameter gradients actually optimized during training. The empirical gains, while consistently positive on average, are often small relative to the prior SPG baseline and are reported without per-run variance in the main tables. The conceptual contribution of 'Prompt Variability' is interesting, but the theoretical framing as currently written overstates what has been proven.","major_comments":[{"comment":"The entire margin-enlargement result (Eq. 12) and the subsequent exponential gradient bound (Eq. 17) rest on the inequality s_i^-(x) ≥ s_y^-(x) + δ for every incorrect class i and every input x. This inequality is introduced with the phrase 'by designing the negative prompts,' but none of the training objectives in §III-B enforces it: Eq. (5) is a BCE loss on label probabilities, and Eq. (6) is an MSE loss on prompt vectors. Neither is a ranking or margin loss defined on the inference-time negative scores. The paper presents no measurement of δ on source or target data, and no ablation checks whether the inequality holds for the generated negative prompts. If Eq. (11) fails for even one incorrect class, the direction of Eq. (12) reverses and negative learning can shrink the margin. This is a load-bearing assumption, not a derived result; please either enforce Eq. (11) with a suitable loss, empirically verify δ on both source and target splits, or explicitly reframe the theory as conditional on an assumption that is tested rather than asserted.","section":"§III-C, Eq. (11)"},{"comment":"The claim of 'gradient norm stability' during optimization does not follow from the presented calculation. Eq. (13)–(16) bound the Jacobian of the softmax output with respect to the input x, ∂f/∂x, after assuming an L-Lipschitz map from x to the logits g. But the training in §III-B optimizes the parameters of G+ and G−; the quantity relevant for optimization is ∂L/∂θ, which is not controlled by the input-output Jacobian bound. A smaller ∥J_f(x)∥ is a robustness statement about sensitivity to input perturbations, not a statement about smoother or more stable gradient descent over prompt-generator parameters. Please correct the object being bounded, or remove the claim that this result explains stable optimization.","section":"§III-C, Eqs. (13)–(17)"},{"comment":"The main results are reported as single numbers without standard deviations over random seeds, despite the paper's emphasis on seed variability. Some of the headline gains are small relative to the closest baseline (e.g., Tables II–IV: VLCS ResNet-50 improvement over SPG is 0.13, PACS ViT-B/16 improvement is 0.25 and ties with VPT at 96.93, DomainNet ResNet-50 improvement is 0.26), and several per-domain accuracies are actually lower than SPG (e.g., VLCS Pascal 84.80 vs. 85.37; VLCS Sun 80.41 vs. 82.40; OfficeHome Product 82.60 vs. 83.40 for ResNet-50). Without variance estimates or significance tests, the claim of 'consistently outperforms' is not fully supported in the main tables. Please add mean±std across at least three seeds for the main comparisons and discuss the significance of the average gains.","section":"Tables II–IV"}],"minor_comments":[{"comment":"The description of y_dj as a 'multi-label one-hot encoded vector (e.g., [1,0,1])' is confusing because all DG datasets used in the paper are single-label; please clarify whether this is a one-hot vector or a true multi-label vector, and how the BCE loss is applied in the single-label case.","section":"§III-C, Eq. (5)"},{"comment":"The row descriptions for the ablation are ambiguous. Row #3 is described in the text as 'DPSPG without incorporating negative learning,' but the checkmark pattern appears to include CGAN and exclude the transformer, which would make it an SPG variant rather than DPSPG without negative learning. Please clarify the exact component configuration of each row, for example with explicit '✓/✗' entries for all four components.","section":"Table V"},{"comment":"The symbol α is used both as the weight for the negative MSE loss in Eq. (6) and as the weight of the negative score in the inference logit in Eq. (7). The implementation section states α=0.2, but the sensitivity analysis (Fig. 6) is described as varying 'the combination weight α used in Equation 6.' Please clarify whether the same α is used in both places and whether the sensitivity analysis varies both simultaneously.","section":"§III-B, Eqs. (6) and (7)"},{"comment":"The abstract and introduction state that 'theoretical and empirical analyses demonstrate' that negative learning increases the effective margin and reduces the gradient-norm upper bound. Given the conditional nature of the theory (see major comment 1), please temper this phrasing, e.g., by saying the analysis shows the margin gain under an explicit separation condition that is then validated empirically.","section":"Abstract and Introduction"}],"recommendation":"major_revision","confidential_remarks":"The paper's main empirical contribution is plausible and the experiments are extensive, but the theoretical section is the stated justification for the method's name and claims. Since the key inequality (11) is neither enforced nor measured, and the gradient-norm bound is about input sensitivity rather than parameter updates, the theory should either be repaired or substantially de-emphasized. The reported empirical gains are often marginal, so the authors should also provide variance information in the main tables. With these revisions, the paper could be acceptable, but in its current form the central explanatory claim is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the empirical core of this paper is real: DPSPG—a dual-path generator producing positive and negative soft prompts for CLIP—consistently beats its immediate predecessor SPG on five DG benchmarks, with code released and a clean ablation showing the negative path matters. Second, the theoretical explanation they offer for why it works is not actually a proof. The margin enlargement and gradient-norm bounds are conditional on inequality (11), which the training losses (BCE and MSE) never enforce and the paper never measures on source or target data.\n\nWhat is genuinely new: the Prompt Variability diagnosis. The authors show that existing dynamic prompt methods (DPL, SPG) produce prompts that scatter across random seeds, and they propose a sensible fix: generate both positive and negative prompts with separate transformer generators trained to match domain-specific prompt labels, then subtract the negative logit from the positive one at inference. The stability improvement is substantiated with box plots and a lower intra-to-inter domain distance ratio. That is a useful, concrete contribution.\n\nThe soft spots are proportionate. The main tables give averages but no standard deviations, and some of the gains (e.g., PACS ViT) are under 0.3 points where the baseline is already 96.5. The TerraIncognita ResNet-50 gain of 7.89 points is striking, but it comes from a single run and no error bar. The theory is the bigger issue: Eq. (12) and (17) follow only if s−_i(x) − s−_y(x) ≥ δ for every false class i, yet no loss is designed to create that separation and no experiment reports whether it holds. Without that, the claims of “αδ margin gain” and exponentially smaller gradient norms are not justified by the presented evidence. The theory reads as a post-hoc heuristic.\n\nWho is this for? Anyone working on prompt learning or DG who wants a stable, efficient alternative to SPG and is willing to treat the margin story as motivation rather than proof. It deserves a serious referee: the empirical work is extensive and reproducible, and the theoretical overreach can be fixed by measuring δ or reframing. I would cite the method in a prompt-learning survey, and I would bring it to reading group as a good case study in separating empirical claims from theoretical framing.","headline":"A solid empirical recipe for stabilizing dynamic prompts via negative learning, but the margin theory rests on an unverified inequality the paper never measures.","tokens_in":19701,"tokens_out":3220,"would_cite":true,"duration_ms":28557,"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":"Negative prompts enlarge the decision margin and stabilize prompt-based domain generalization.","keywords":["domain generalization","vision-language models","prompt learning","soft prompt generation","negative learning","prompt variability","transformer prompt generator","CLIP"],"falsifier":"On any held-out target domain, compute $\\delta(x)=\\min_{i\\ne y}\\left(s^-_i(x)-s^-_y(x)\\right)$ using the trained negative prompt generator. If a substantial fraction of target images have $\\delta(x)\\le 0$, then the inequality underlying Equation (12) fails on the test distribution, and the claimed margin expansion and exponential gradient-norm bound cannot explain the reported accuracy gains.","tokens_in":1611,"feed_emoji":"🎯","tokens_out":1776,"duration_ms":56694,"temperature":0.7,"pith_summary":"The paper targets a problem it calls Prompt Variability: dynamic prompt generators for vision-language models often produce very different prompts for the same image across random seeds, and those prompts sit far from the optimal domain prompt. The proposed method, Dual-Path Stable Soft Prompt Generation (DPSPG), trains two transformer-based prompt generators, one producing positive prompts such as “a photo of a {class}” and one producing negative prompts such as “a photo without a {class}”, then subtracts the negative score from the positive score at inference. The paper argues that negative learning enlarges the effective classification margin by at least $\\alpha\\delta$ and exponentially tightens an upper bound on the gradient norm, which explains smoother optimization and more robust prompts. Experiments on five domain-generalization benchmarks report higher average accuracy than prior prompt-generation methods together with lower seed-to-seed variance, supporting the claim that negative prompts are a useful stabilizer.","feed_headline":"Negative prompts stabilize domain-generalization prompts","feed_subtitle":"A dual positive/negative prompt generator enlarges decision margins and cuts seed-to-seed variance on five benchmarks.","key_machinery":"The central object is the dual-path prompt generator: two transformer encoders, $G^+$ and $G^-$, each composed of four transformer layers and a linear layer, which map CLIP image embeddings to positive and negative soft prompts. These generators are trained with a mean-squared-error loss to match per-domain positive and negative prompt labels learned in a first stage, and at inference the two prompts define the combined logit $g_i(x)=s^+_i(x)-\\alpha s^-_i(x)$. The load-bearing identity is the separation inequality $s^-_i(x)\\ge s^-_y(x)+\\delta$ for every wrong class $i$; substituting it into the margin equation yields $\\Delta_i(x)\\ge \\Delta^+_i(x)+\\alpha\\delta$, which in turn produces the exponential Jacobian bound $\\|J_f(x)\\le \\frac{L}{\\tau}e^{-(\\Delta^+_i(x)+\\alpha\\delta)/\\tau}$ that carries the stability argument.","core_discovery":"The central claim is that negative learning stabilizes prompt generation. By generating a negative soft prompt alongside the positive one and using the combined logit $g_i(x)=s^+_i(x)-\\alpha s^-_i(x)$, the margin between the true class and each incorrect class becomes at least $\\Delta^+_i(x)+\\alpha\\delta$ whenever the negative-prompt separation condition $s^-_i(x)\\ge s^-_y(x)+\\delta$ holds for every incorrect class $i$. The paper further derives an exponential gradient-norm bound $\\|J_f(x)\\le \\frac{L}{\\tau}e^{-(\\Delta^+_i(x)+\\alpha\\delta)/\\tau}$, arguing that the enlarged margin yields smoother optimization and reduced sensitivity to input perturbations. Empirically, DPSPG reports state-of-the-art accuracy on PACS, VLCS, OfficeHome, TerraIncognita, and DomainNet while producing prompts that cluster tightly around the optimal prompt, directly addressing the Prompt Variability identified in prior dynamic prompt methods.","pith_inferences":["The paper never measures $\\delta$ on target data; a direct test is to compute $s^-_i(x)-s^-_y(x)$ for the learned negative prompts on held-out domains. If $\\delta\\le 0$ for a non-negligible fraction of test images, the claimed margin expansion is a source-domain artifact and the accuracy gains would need another explanation.","The dual-path stabilization mechanism is not specific to prompts: any generative adapter that conditions on image features could in principle gain seed stability from a secondary generator trained to reject the class-conditional template, although the paper does not explore that extension.","The robustness claim that a smaller Jacobian norm reduces sensitivity to input perturbations suggests DPSPG should exhibit improved adversarial robustness, but the paper runs no adversarial attacks; an adversarial-robustness evaluation would be a direct test of that corollary.","Fixing $\\alpha=0.2$ across all datasets may leave performance on the table for large-shift domains such as Sketch and TerraIncognita, where the optimal balance between positive and negative scores could differ."],"forward_implications":["If the margin result holds, dynamic prompt generation for domain generalization becomes reproducible: the same image will produce nearly identical prompts across seeds, removing seed dependence as a major source of performance variance.","The exponential gradient-norm bound implies smoother optimization, so DPSPG should train more stably and with less sensitivity to learning-rate and warm-up choices than GAN-based prompt generators such as SPG.","The dual-path recipe of training a positive and a negative generator on domain prompt labels and subtracting scores at inference is a general template that can be applied to other CLIP-based classification settings beyond the five DG benchmarks tested.","Because the margin gain scales with $\\alpha\\delta$, the balancing weight $\\alpha$ acts as a confidence lever: larger values widen margins when negative prompts are reliable but increase reliance on source-learned negative labels."],"supporting_citations":[{"why":"Supplies the frozen CLIP visual and text encoders that DPSPG adapts and serves as the zero-shot baseline.","marker":"[11]"},{"why":"Provides the learnable soft-prompt framework used to train the positive and negative domain prompt labels, and serves as a fixed-prompt baseline.","marker":"[13]"},{"why":"Is a dynamic domain-prompt baseline whose prompt variability motivates the Prompt Variability problem.","marker":"[16]"},{"why":"Is the main dynamic prompt-generation baseline; DPSPG replaces its GAN generator with transformer-based dual-path generators and compares directly against it.","marker":"[17]"},{"why":"Is the instance-specific prompt learning baseline that also exhibits prompt variability and is used for comparison.","marker":"[42]"},{"why":"Is a multi-modal prompt learning baseline used in the comparison tables for domain generalization performance.","marker":"[15]"}],"fun_headline_variants":["Negative prompts steady domain-generalization training","Stable soft prompts via negative learning","Dual-path prompt generator cuts seed variance","Negative learning enlarges prompt margins","Prompt variability tamed by negative prompts"],"cache_read_input_tokens":21888,"weakest_assumption_plain":"The margin and gradient-stability theorems rely on the assumption that for every wrong class, the negative prompt gives the true class a score at least $\\delta$ lower than the wrong class, but this separation is learned only on source domains and is never verified on the unseen target domains where it must hold.","fun_headline_variants_meta":{"raw":{"variants":["Negative prompts steady domain-generalization training","Stable soft prompts via negative learning","Dual-path prompt generator cuts seed variance","Negative learning enlarges prompt margins","Prompt variability tamed by negative prompts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1369,"prompt_tokens":989,"completion_tokens":380,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":605,"completion_tokens_details":{"reasoning_tokens":319}},"tokens_in":605,"tokens_out":380,"duration_ms":3608,"temperature":1.0,"reasoning_tokens":319,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:25:18.617916+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On any held-out target domain, compute $\\delta(x)=\\min_{i\\ne y}\\left(s^-_i(x)-s^-_y(x)\\right)$ using the trained negative prompt generator. If a substantial fraction of target images have $\\delta(x)\\le 0$, then the inequality underlying Equation (12) fails on the test distribution, and the claimed margin expansion and exponential gradient-norm bound cannot explain the reported accuracy gains.","supporting_citations":[{"cited_title":"Domain prompt learning for efficiently adapting clip to unseen domains,","cited_arxiv_id":null,"evidence_quote":"Is a dynamic domain-prompt baseline whose prompt variability motivates the Prompt Variability problem."},{"cited_title":"Soft prompt generation for domain generalization,","cited_arxiv_id":null,"evidence_quote":"Is the main dynamic prompt-generation baseline; DPSPG replaces its GAN generator with transformer-based dual-path generators and compares directly against it."},{"cited_title":"Conditional prompt learning for vision-language models,","cited_arxiv_id":null,"evidence_quote":"Is the instance-specific prompt learning baseline that also exhibits prompt variability and is used for comparison."}],"review_version":1}