{"id":"88758b66-dca4-4358-932a-1a8c05477fea","arxiv_id":"2504.18204","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"The VCA framework uses diversity, consistency, and preference rewards to fine-tune a diffusion model with LoRA over multi-round dialogues, reporting improved intent alignment.","lead":"This paper introduces a multi-round dialogue framework for image generation that fine-tunes Stable Diffusion with LoRA using a preference reward model, claiming better alignment with user intent. It reports higher user satisfaction and consistency than several existing systems in multi-turn image editing tests.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Load-bearing premise: R_MI is a faithful reward for human intent. It is not validated against independent human judgments, its DPO negatives are heuristic mismatches, and the headline preference comparisons use R_MI itself.","rationale":"The reader's weakest assumption identifies the same load-bearing premise: the DPO-trained reward model must be a faithful and stable proxy for human preference. My stress-test concurs and sharpens the evidentiary issue. The paper uses R_MI for three distinct purposes: as the objective of PPO/LoRA optimization, as the evaluation metric for preference scores and win rates, and as the basis for claims about human intent alignment. When the same unvalidated model is used for both training and evaluation, agreement with itself is unsurprising and does not support the central claim. A direct validation of R_MI against independent human labels would settle whether the reward is genuinely aligned with user intent. The theoretical flaws in Theorem 3.1 and 3.2 are also serious, but they undermine the paper's 'rigorous foundation' claim rather than the empirical headline; the reward-model validation is more directly load-bearing for the claimed user-satisfaction results. The method might still work empirically, but the current evidence is insufficient, and the lack of statistical rigor in Table 1 strengthens the concern. Therefore the reader's REJECT verdict stands unchanged.","tokens_in":17531,"tokens_out":4163,"duration_ms":41970,"concrete_test":"Conduct a preregistered human evaluation using independent raters on a held-out set of at least 500 image pairs from multi-turn dialogues, including VCA outputs and baseline outputs. Compute (a) Spearman correlation and top-1 accuracy of R_MI against human preference labels, compared against CLIP and ImageReward, and (b) a paired significance test (e.g., Wilcoxon signed-rank or bootstrap CI) for VCA versus each baseline on the same human labels. If R_MI is not significantly better than CLIP at predicting human judgments, or if VCA is not significantly preferred over baselines, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim—that VCA consistently surpasses baselines in user satisfaction, especially in multi-turn dialogue—depends on R_MI being a faithful and stable proxy for human preference. The paper does not establish this. R_MI (Section 3.2, Eq. (10)) is not mutual information; it is the mean logit of a Qwen-VL model fine-tuned with DPO. Its training labels come from heuristic mismatching: random image mismatches in FashionIQ/DeepFashion, opposite or random text prompts in the custom dataset, and lower-scoring ImageReward images (Section 4.2). These negative examples do not necessarily reflect real user dissatisfaction in multi-round correction, and the reward model is never scored against held-out human preference judgments. The evaluation compounds the problem: Figure 4 and Figure 5 report 'preference scores' that are R_MI or derived from it, so comparing methods by this metric is partially circular; Figure 6 compares selection accuracy using R_MI as the yardstick. Table 1 reports 'Human Eval' only as a rank/win count, without confidence intervals, significance tests, or details sufficient to judge the 2706-user procedure. If R_MI is biased or miscalibrated, the PPO/LoRA updates optimize the wrong objective and the claimed 'consistent surpass' is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Visual Co-Adaptation (VCA), a framework for multi-round, human-in-the-loop text-to-image generation. It constructs multi-turn dialogue prompt-image datasets, trains a reward model by applying DPO to Qwen-VL, and fine-tunes Stable Diffusion v2.1 with LoRA using a total reward that combines diversity, consistency, and a so-called mutual information term. The authors state two theorems: convergence of the multi-round latent distribution to a target distribution in total variation (Theorem 3.1), and convergence of the dynamically weighted reward optimization to the Pareto optimal set (Theorem 3.2). The experimental section compares VCA with several text-to-image baselines on preference scores, CLIP/BLIP/LPIPS, modality-switching tasks, ablations, and a 2706-user human evaluation, concluding that VCA is consistently superior, especially in multi-turn dialogue scenarios.","tokens_in":17909,"tokens_out":4587,"duration_ms":47290,"significance":"If the claims were substantiated, the framework would be a practically useful contribution to interactive text-to-image generation, and the described multi-turn dialogue dataset could be a reusable resource. The paper has concrete strengths: a large constructed dataset, a systematic combination of three reward terms, ablations of the reward components, and a clear system design. However, the evaluation is compromised by the fact that the headline 'preference score' is the same DPO-trained reward used as the training objective, the human study lacks statistical detail, and the theoretical results are not correct as stated. In its current form, the paper does not establish its central claims; the useful components could be salvaged only with substantially revised evaluation and theory.","major_comments":[{"comment":"Theorem 3.1 is mathematically invalid as stated. The proof explicitly identifies the limiting object p_target with a Dirac delta, but each iterate p_T is an absolutely continuous Gaussian density. Total variation distance between an absolutely continuous probability measure and a Dirac measure is 1, not 0, and Scheffé's lemma cannot be invoked because the limit is not a density. The argument that pointwise convergence p_T(z) to 0 for z ≠ z_target combined with distributional convergence to delta implies convergence in TV is false. Because this theorem is listed as a key contribution, the error is load-bearing.","section":"Theoretical Analysis in §3.1 and Appendix A.1 (Theorem 3.1, Eq. 7)"},{"comment":"The proof of Theorem 3.2 relies on incorrect geometric and variational claims. Cosine similarity on the unit sphere is not an affine function, and the same normalized inner product cannot be simultaneously concave for R_div and convex for R_cons as asserted. The step that dV/dt tends to 0 does not imply that the sequence z_t* converges to a maximizer, and the negative-definiteness of d^2V/dt^2 is asserted without checking the signs of the reward terms. The scalarization argument shows at most that each z_t* is Pareto optimal for its own fixed weights, not that the sequence converges to a balanced Pareto point as t tends to infinity. Thus Theorem 3.2 is unproven.","section":"Appendix A.2 (Theorem A.10 / Theorem 3.2)"},{"comment":"Eq. (10) defines R_MI = I(X;Y), but the implementation is a DPO-trained Qwen-VL model whose reward is the mean logit; no identity with mutual information is established. Figures 4, 5, and 6 evaluate methods using this same 'preference score' (R_MI). Since VCA is optimized against R_MI, comparing methods on this metric is partly circular for the claim of improved preference understanding, and the metric is never validated against independent human preference judgments. The DPO negative examples are constructed by heuristic mismatch, which may not reflect genuine user dissatisfaction in multi-turn correction. This leaves the central empirical claim unsubstantiated.","section":"§3.2 (Eq. 10) and §4.3 (Figs. 4-6)"},{"comment":"The human evaluation is reported only as ranks and win counts for 2706 users. There are no confidence intervals, no significance tests, no inter-rater agreement measure, and the protocol description ('blind cross-over design') lacks essential details: how prompts were sampled, how many images each user saw, how the composite 0-5 scores for response time, aesthetics, intent reflection, and round number were aggregated, and whether the comparison was paired. Without this information, the conclusion that VCA 'consistently surpasses' competing models in user satisfaction is not supported by the reported evidence.","section":"Table 1 (Human Eval)"}],"minor_comments":[{"comment":"Section 4.5 appears to duplicate the ablation paragraph from Section 4.4 ('Ablation Study on Reward Coefficient Impact'); this is likely a copy-paste error.","section":"§4.5"},{"comment":"Notation is inconsistent: Eq. (2) writes DM(t,τ), while Eq. (6) writes DM(t); the indexing should be made uniform and precise.","section":"§3.1 (Eq. 2 vs. Eq. 6)"},{"comment":"The header 'Lpips [38]↓ Aesthetic Score CLIP [20] ↑ BLIP [14]↑ Round↓' is confusing because the row entries alternate 'Rank' and 'Score'; clarifying which columns are ranks and which are scores would improve readability.","section":"Table 1 header"},{"comment":"Figure 8 reports user satisfaction, BLIP, CLIP, and LPIPS curves without error bars; given the plateau claim between 10k and 15k samples, variance information is needed.","section":"Figure 8"},{"comment":"The same quantity is referred to as both 'mutual information reward' and 'preference score'; this conflation should be resolved and the terminology aligned.","section":"§3.2 (Eq. 10) and §4.3"},{"comment":"Reference [6] ('Enhancing Intent Understanding for Ambiguous prompt: A Human-Machine Co-Adaption Strategy') lacks a venue, arXiv identifier, or year, and should be completed.","section":"References"}],"recommendation":"reject","confidential_remarks":"The two theorems contain false mathematical statements, and the principal evaluation uses the training reward as the metric, making the central claim circular. The dataset construction and system design may be salvageable, but in the current form the load-bearing errors are not local fixes; I would not encourage a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Kun Li et al. put together a sensible-looking pipeline for multi-round text-to-image generation: a QLoRA-tuned Qwen-VL reward model, PPO updates on LoRA adapters, and dynamic weights over diversity, consistency, and that reward. The assembled dataset—30/35/18% subsamples of ImageReward/DeepFashion/FashionIQ plus 38.8k custom items structured into multi-turn dialogues—is a genuinely useful resource, and the idea of weighting diversity early and consistency later is a reasonable heuristic. Credit where due: the experimental infrastructure is non-trivial, and the paper is clearly trying to solve a real problem.\n\nThe problems start with the theory. Theorem 3.1 claims TV convergence to a target density, but the proof explicitly builds a sequence of Gaussians collapsing to a Dirac delta; Scheffé's lemma does not apply to limit objects that aren't densities, and TV distance to a delta is 1, not 0. This is not a technical gap, it's a false theorem as stated. Theorem 3.2 is similarly hand-wavy: the concavity/convexity claims rest on unjustified assumptions about the feature map, and the convergence argument reduces to 'dV/dt -> 0, second derivative negative, so it converges'—that's not a proof. The paper calls R_MI 'mutual information' and then defines it as the mean logit of a DPO-trained Qwen-VL; those are different things.\n\nThe evaluation has a circularity problem: Figures 4-6 use 'preference score' from that same DPO reward model, so the model is being judged by the metric it was trained to optimize. The human study—2706 users, no error bars, no significance tests, no protocol details—cannot carry the headline 'consistently surpasses.' No code or data are released, so the dataset contribution is unverifiable.\n\nI'm not saying the method can't work; the components are all standard and the multi-round setup is plausible. But as submitted, the central claims are not credible. A serious editor should desk reject this. If the authors replace the theory with honest statements, validate R_MI against independent human judgments, report the user study properly, and release the dataset, it's worth another look. For now, it's a pass.","headline":"A plausible multi-turn reward-tuning pipeline undermined by invalid theory, a mislabeled reward model, and circular evaluation.","tokens_in":18368,"tokens_out":3391,"would_cite":false,"duration_ms":36313,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A multi-round dialogue loop with diversity, consistency, and a DPO-trained preference reward can steer diffusion image generation to match user intent better than single-shot models.","keywords":["multi-round dialogue","text-to-image generation","diffusion models","human preference alignment","reward model","LoRA fine-tuning","mutual information","human-in-the-loop"],"falsifier":"Collect independent human preference judgments on held-out multi-turn dialogues and compare the orderings produced by the $R_{\\mathrm{MI}}$ preference score, CLIP, aesthetic, and BLIP scores; if the preference score does not match human rankings better than the other metrics, or if PPO updates that raise $R_{\\mathrm{MI}}$ lower human satisfaction, the central claim fails.","tokens_in":17338,"feed_emoji":"🎨","tokens_out":9116,"duration_ms":86066,"temperature":0.7,"pith_summary":"The paper tries to establish that text-to-image generation gets better when optimization spans the whole multi-round conversation rather than a single prompt. It introduces Visual Co-Adaptation (VCA): an LLM refines the prompt at each round, a DPO-trained reward model built on Qwen-VL scores how well each candidate image matches user intent, and PPO tunes LoRA adapters inside Stable Diffusion to maximize a reward that combines diversity, inter-round consistency, and that preference score. The authors construct a large multi-turn dialogue dataset of prompt-image pairs and report that VCA surpasses existing systems in user satisfaction, win rates, modality-switching accuracy, and visual consistency, especially in multi-turn settings. If the paper is right, iterative user feedback becomes usable as a direct training signal for diffusion models without retraining the full backbone.","feed_headline":"Feedback loops teach diffusion models what users actually want","feed_subtitle":"Diversity, consistency, and intent-alignment rewards beat single-shot text-to-image baselines in user tests.","key_machinery":"The load-bearing object is the dynamic total reward $R_{\\mathrm{total}}(t) = \\lambda_{\\mathrm{div}}(t) R_{\\mathrm{div}} + \\lambda_{\\mathrm{cons}}(t) R_{\\mathrm{cons}} + \\lambda_{\\mathrm{MI}}(t) R_{\\mathrm{MI}}$, with $\\lambda_{\\mathrm{div}}(t)=e^{-\\alpha t}$, $\\lambda_{\\mathrm{cons}}(t)=1-e^{-\\beta t}$, and $\\lambda_{\\mathrm{MI}}(t)=\\frac12 e^{-\\gamma t}$. $R_{\\mathrm{div}}$ pushes early-round samples apart in the U-Net feature space, $R_{\\mathrm{cons}}$ rewards cosine similarity between consecutive dialogue rounds, and $R_{\\mathrm{MI}}$, called the preference score, comes from a Qwen-VL model fine-tuned with DPO on positive/negative image pairs. PPO converts $R_{\\mathrm{total}}$ into gradient updates for LoRA adapters inserted in the attention layers, while the base diffusion weights stay fixed. The time-dependent weights shift training from exploration (diversity) to stability (consistency) and intent alignment as dialogue rounds proceed.","core_discovery":"The central claim is that mutual-information maximization—implemented as a preference score from a Qwen-VL reward model trained with DPO—captures user intent more faithfully than both standard reinforcement learning and CLIP-style metrics, and that this reward can guide a multi-round diffusion process. The paper argues that optimizing $R_{\\mathrm{total}}$ with PPO while updating only LoRA adapters produces images that are diverse in early rounds, consistent across rounds, and increasingly aligned with the user's stated intent. In the reported evaluations, the method ranks first on human satisfaction, BLIP and CLIP scores, and the number of dialogue rounds needed, and it outperforms baseline models in win rates across eight dialogue rounds.","pith_inferences":["The same reward decomposition could transfer to other iterative generative settings—image editing, video generation, or embodied instruction following—with the dynamic weighting schedule as the reusable component.","Because the reward model's negative examples are built by heuristics (low-scoring images, random mismatches, LLM-generated unrelated images), a testable prediction is that the preference score is vulnerable to shortcuts based on surface style rather than genuine intent; independent human labels would settle this.","An extension the paper does not explore is replacing the fixed DPO-trained reward with the user's own in-dialogue selections, which would make co-adaptation fully personal and remove the need for a separate reward model.","The convergence results assume the prompt-refinement process reaches an ideal fixed point; taken literally, this predicts that remaining failures will concentrate in the LLM's prompt-update step rather than in the diffusion sampler."],"forward_implications":["Interactive text-to-image systems can be trained directly on multi-round dialogue data, so each user refinement becomes part of the learning signal rather than a one-off edit.","A time-dependent reward schedule offers a general recipe for balancing exploration and convergence in iterative generative tasks.","A DPO-trained vision-language reward model can serve as a reusable preference oracle that is more discriminative than CLIP for selecting images that match intent.","LoRA-based reward-driven updates let a diffusion model be adapted to new preferences without retraining the full backbone, making per-user or per-dialogue personalization feasible."],"supporting_citations":[{"why":"supplies the Stable Diffusion v2.1 base model that VCA fine-tunes with LoRA.","marker":"[25]"},{"why":"is the vision-language backbone from which the mutual-information reward model is built.","marker":"[2]"},{"why":"provides the Direct Preference Optimization objective used to train the reward model on preference pairs.","marker":"[21]"},{"why":"provides the Proximal Policy Optimization update that adjusts LoRA parameters from the reward signal.","marker":"[28]"},{"why":"provides the low-rank adaptation machinery that keeps fine-tuning parameter-efficient.","marker":"[9]"},{"why":"provides part of the dialogue dataset and the ImageReward preference baseline that VCA is compared against.","marker":"[36]"},{"why":"defines the multi-turn text-to-image dialogue task and is the main comparison for modality-switching performance.","marker":"[11]"},{"why":"is the prompt-to-prompt editing baseline that VCA compares against for consistency and alignment.","marker":"[7]"},{"why":"supplies the human-machine co-adaptation idea, extended here to multi-round visual dialogues.","marker":"[6]"}],"fun_headline_variants":["Multi-round rewards teach diffusion models user intent","DPO-trained reward guides diffusion across dialogue rounds","Intent-aligned diffusion: rewards beat CLIP in user tests","Co-adaptive rewards improve diffusion image alignment","Diffusion learns to match user intent via multi-turn feedback"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the DPO-trained Qwen-VL reward model's preference score is a faithful and stable proxy for what human users actually want across dialogue rounds, so that PPO updates driven by it improve real intent rather than a heuristic artifact.","fun_headline_variants_meta":{"raw":{"variants":["Multi-round rewards teach diffusion models user intent","DPO-trained reward guides diffusion across dialogue rounds","Intent-aligned diffusion: rewards beat CLIP in user tests","Co-adaptive rewards improve diffusion image alignment","Diffusion learns to match user intent via multi-turn feedback"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1361,"prompt_tokens":856,"completion_tokens":505,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":472,"completion_tokens_details":{"reasoning_tokens":431}},"tokens_in":472,"tokens_out":505,"duration_ms":5142,"temperature":1.0,"reasoning_tokens":431,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:22:18.001854+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect independent human preference judgments on held-out multi-turn dialogues and compare the orderings produced by the $R_{\\mathrm{MI}}$ preference score, CLIP, aesthetic, and BLIP scores; if the preference score does not match human rankings better than the other metrics, or if PPO updates that raise $R_{\\mathrm{MI}}$ lower human satisfaction, the central claim fails.","supporting_citations":[{"cited_title":", 3 pages","cited_arxiv_id":null,"evidence_quote":"supplies the Stable Diffusion v2.1 base model that VCA fine-tunes with LoRA."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides part of the dialogue dataset and the ImageReward preference baseline that VCA is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the multi-turn text-to-image dialogue task and is the main comparison for modality-switching performance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the prompt-to-prompt editing baseline that VCA compares against for consistency and alignment."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the human-machine co-adaptation idea, extended here to multi-round visual dialogues."}],"review_version":1}