{"id":"22d0af52-c82e-4785-a6b6-45cc09d603f5","arxiv_id":"2505.22792","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Rhet2Pix combines staged LLM prompt decomposition with a discounted PPO fine-tuning scheme for Stable Diffusion, claiming strong rhetorical text-to-image generation, but the quantitative evidence is circular and undefined.","lead":"A new framework, Rhet2Pix, tries to make text-to-image models understand metaphors and similes by breaking prompts into stages and fine-tuning Stable Diffusion with reinforcement learning. The paper claims it beats GPT-4o and Grok-3, but the evaluation uses metrics that appear to be the same as the training reward and are never defined.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's alignment metrics are undefined and coincide with the training rewards, making the SOTA claim circular.","rationale":"The load-bearing assumption is that the Table 1 metrics measure rhetorical fidelity. The reader's weakest assumption identifies exactly this, and I agree. The paper defines neither the metrics nor any external validation; the metric names coincide with the training rewards, making the evaluation circular. The qualitative examples are cherry-picked and do not compensate. The only independent evidence would be a publicly released, reproducible evaluation script and a human study; neither is present. Therefore the central claim—outperforming GPT-4o and Grok-3—is unsupported. The verdict should remain as the reader set it. I considered whether the 'two-layer MDP' being a fixed-LLM outer layer is a separate load-bearing issue, but the empirical claim fails first on evaluation validity; even the engineering novelty would not rescue an invalid comparison. No change to the reader's verdict is needed.","tokens_in":16189,"tokens_out":3895,"duration_ms":38799,"concrete_test":"Pre-register a human pairwise preference study on a held-out set of 100 rhetorical prompts with at least 50 raters, comparing Rhet2Pix against GPT-4o and Grok-3 on whether the image conveys the intended rhetorical meaning rather than the literal vehicle. If human raters do not significantly prefer Rhet2Pix (e.g., binomial test, p < 0.05), or if their judgments do not correlate with the Table 1 alignment scores, the metric is not a valid proxy for rhetorical fidelity and the SOTA claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1's headline numbers are not a valid basis for the SOTA claim. The columns 'Semantic-alignment' and 'Elemental-alignment' are never defined in Section 5.2, yet their names exactly match the two reward terms in Section 4.5: r_stage = sum_j w_j <v, u_j> (staged semantic alignment) and r_subject = sum_k <v, f_text(k)/||f_text(k)||> with a vehicle penalty (elemental alignment). No human annotation, external benchmark (e.g., MetaCLUE or T2I-CompBench), or independent metric definition is provided. Since Rhet2Pix is trained via PPO to maximize these exact CLIP cosine similarities, its higher Sum-alignment (0.6671 vs Grok-3's 0.4192) is an expected consequence of optimizing the evaluation metric, not evidence of better rhetorical understanding. The paper's own Appendix C concedes that true rhetorical generation 'may necessitate a fundamentally new backbone' beyond conventional text-image embedding similarity, further undermining the metric's validity. Without a definition of the metric and external validation, the central comparison to GPT-4o and Grok-3 collapses.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Rhet2Pix, a framework for rhetorical text-to-image generation that formulates the task as a two-layer Markov Decision Process and fine-tunes a Stable Diffusion backbone with a PPO-style policy optimization. In the outer layer, an LLM decomposes a rhetorical prompt into staged sub-prompts whose images are evaluated by a rhetorical reward; in the inner layer, the final reward is discounted backward along the denoising trajectory to address reward sparsity. The paper reports quantitative results in Table 1 claiming that Rhet2Pix outperforms GPT-4o, Grok-3, and several academic baselines on 'semantic-alignment' and 'elemental-alignment' metrics, and provides qualitative examples and ablations.","tokens_in":16588,"tokens_out":5795,"duration_ms":64275,"significance":"If the claims were supported, the paper would make a useful contribution to an underexplored task: generating images that preserve figurative meaning rather than literal vehicle depiction. The paper is also commendable for planning to release code and data. However, the headline quantitative result is not currently supported because the evaluation metrics in Table 1 are not defined independently of the training reward in Section 4.5. The paper's own limitation statement in Appendix C further weakens the validity of CLIP-based alignment as the sole measure of rhetorical fidelity. As a result, the empirical contribution is unverified.","major_comments":[{"comment":"The evaluation metrics are never defined in Section 5.2, and their names ('Semantic-alignment', 'Elemental-alignment') exactly match the reward terms 'staged semantic alignment' and 'elemental alignment' in Section 4.5. The formulas there are CLIP cosine similarities, including the manually tuned vehicle penalty threshold tau. Because Rhet2Pix is trained via PPO to maximize these same quantities, its higher scores on this table are a direct consequence of optimizing the evaluation metric rather than evidence of better rhetorical understanding. The comparison to GPT-4o and Grok-3 is therefore circular. The authors must either define the evaluation metrics as external, independent measures, or replace this table with human evaluation or a standard benchmark (e.g., MetaCLUE or T2I-CompBench) applied to all models under identical conditions.","section":"Section 5.2, Table 1"},{"comment":"The paper explicitly concedes that fully addressing rhetorical text-to-image generation 'may necessitate a fundamentally new backbone—one that surpasses conventional text-image embedding similarity.' This statement directly contradicts the validity of using CLIP-based embedding similarities as the sole evaluation metric for rhetorical fidelity in Table 1. If the authors believe this limitation statement, then the quantitative evaluation cannot be used to substantiate the claim of outperforming SOTA MLLMs. The paper should either justify how the CLIP metrics capture rhetorical meaning despite this acknowledged limitation, or adopt a more appropriate evaluation protocol.","section":"Appendix C"},{"comment":"The ablation study is presented only qualitatively, with a few image examples. The claims that 'scene modeling' and 'reinforcement learning' are crucial for performance are not supported by any quantitative measurement. Without quantitative ablations using the same (independent) evaluation protocol as the main comparison, it is impossible to assess the contribution of each component or to rule out that the observed qualitative differences are due to cherry-picked examples.","section":"Section 5.3, Figure 5"},{"comment":"The reward function contains several manually set free parameters that are not analyzed: the monotonic weight vector w(j) for staged alignment, the vehicle penalty threshold tau, and the discount factors gamma, lambda, and gamma_denoise. In particular, tau is described as 'manually determined by empirical analysis of CLIP similarity scores'—the same score distribution used in the evaluation table. This creates a risk that the reported improvements reflect tuning of the evaluation-aligned reward rather than a general advance. A sensitivity analysis of these parameters is needed to establish robustness.","section":"Section 4.5"}],"minor_comments":[{"comment":"The inner-layer MDP defines the action as a_t = x_{t-1}, but Section 4.4 repeatedly refers to 'adjacent action pairs (s_t, s_{t+1})'. This is inconsistent: the transition pair should be (s_t, a_t) or (s_t, s_{t-1}), depending on the time indexing. Please clarify the notation.","section":"Section 4.3/4.4"},{"comment":"Line 49 defines the critic loss as E[(V_phi(s_t) - Ahat)^2], which is not the standard value loss used in PPO (typically (V_phi(s_t) - (Ahat + V_old(s_t)))^2 or a similar target). This may lead to biased value estimation and could affect the reported training behavior.","section":"Algorithm 1"},{"comment":"The staged semantic alignment reward assumes each prompt P^j_i is decomposed into sub-sentences S^k_i, but the paper does not specify how this decomposition is obtained. A reproducible description (e.g., an LLM prompt or parsing rule) is needed.","section":"Section 4.5"},{"comment":"The hyperparameter table describes a ViT-based critic and its architectural components, but the main text does not describe the critic architecture. Consider moving a brief description into the main text or at least referencing the table when the critic is introduced in Section 4.4.","section":"Table 3"},{"comment":"There are several minor grammatical issues and awkward phrasings (e.g., 'environmental steps' in Figure 8, inconsistent capitalization of 'state-of-the-art'). A careful proofreading pass would improve readability.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The reader's central concern is well founded: Table 1's metrics appear to be exactly the training rewards from Section 4.5, making the SOTA comparison circular. This is the key issue that must be resolved before publication. The paper also lacks human evaluation and external benchmarks, and its own Appendix C weakens the validity of CLIP-based evaluation. If a major revision is invited, the authors should be asked to (i) provide a properly defined, independent evaluation protocol and, if necessary, rerun all comparisons under it; and (ii) include quantitative ablations and a sensitivity analysis of the reward parameters. Given the fundamental role of the circular evaluation, I cannot recommend acceptance at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What should you know about Rhet2Pix? It's a decent engineering recipe for steering a diffusion model away from literal vehicles in metaphor and simile prompts, but the headline numbers in Table 1 are not evidence of superiority over GPT-4o or Grok-3. The 'Semantic-alignment' and 'Elemental-alignment' scores look exactly like the staged-semantic and elemental reward terms from Sec. 4.5—same names, same CLIP-cosine formulas, same manually tuned vehicle penalty threshold. The paper never defines these metrics independently, so the comparison is effectively 'how well did each model optimize Rhet2Pix's own reward.' That collapses the central claim.\n\nWhat's genuinely new: the specific combination of staged LLM prompt expansion, outer-layer reward shaping with GAE, and per-timestep discounted PPO over denoising pairs is a reasonable contribution to RL fine-tuning of text-to-image models. The paper writes out the algorithm, gives pseudo-code, hyperparameters, and qualitative samples. It also identifies a real failure mode: object-level word embedding alignment drags generation toward literal vehicles. The qualitative ablation (Fig. 5) suggests staged prompting and RL both help, though it's only qualitative.\n\nSoft spots beyond the circular metric. First, the 'two-layer MDP' framing overstates the hierarchy: the outer-layer actions are generated by a fixed LLM with a verification loop; only the denoiser is trained. That's not hierarchical policy optimization in any learned sense. Second, there is no human evaluation or external benchmark (MetaCLUE, T2I-CompBench) to check rhetorical fidelity. Third, the metric definitions and key hyperparameters for the reward weights w(j), threshold τ, and discount factors are missing from the main text, and the appendix doesn't fill that gap either. Minor point: the related work and limitations are honest; Appendix C concedes a fundamentally new backbone may be needed, which partly undercuts the paper's own framing.\n\nVerdict: as submitted, this shouldn't be accepted on its quantitative claims. But the problem is real and the pipeline is a credible baseline for future work. The authors could fix the evaluation with human ratings and independent metrics, and then the method might stand. I'd send it to peer review with a strong request to redo the evaluation, because the core idea deserves refereed scrutiny. For your own work, I'd cite it only for the problem framing, not the results.\n\nRecommendation: engage with it as a reviewer if asked, but don't build on the SOTA claim.","headline":"Rhet2Pix is a plausible engineering recipe for steering diffusion models away from literal metaphor vehicles, but its headline numbers are circular: Table 1's metrics appear to be the training rewards, so the SOTA claim over GPT-4o and Grok-3 is unsupported.","tokens_in":16972,"tokens_out":1855,"would_cite":false,"duration_ms":18880,"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":"Rhet2Pix outperforms GPT-4o and Grok-3 at generating images from rhetorical language.","keywords":["rhetorical text-to-image generation","metaphor visualization","diffusion models","reinforcement learning","Markov decision process","reward sparsity","PPO","prompt decomposition"],"falsifier":"Run a human study where annotators choose which image best captures the figurative meaning of a set of rhetorical prompts, comparing Rhet2Pix against GPT-4o and Grok-3. If annotators do not prefer Rhet2Pix images at a clear rate, or if the reported alignment scores do not predict those choices, the central claim would be refuted.","tokens_in":15978,"feed_emoji":"🎨","tokens_out":4320,"duration_ms":41210,"temperature":0.7,"pith_summary":"The paper claims that rhetorical text-to-image generation fails because models align images to literal word embeddings of the metaphorical vehicle, and proposes Rhet2Pix, a framework that recasts the task as a two-layer Markov decision process. The outer layer decomposes a rhetorical prompt into progressively enriched sub-prompts, while the inner layer treats diffusion denoising as a policy optimization problem with discounts that spread a final reward across all denoising steps. On a filtered FLUTE metaphor and simile benchmark, Rhet2Pix reports higher semantic and elemental alignment scores than GPT-4o, Grok-3, and academic baselines. The claim matters because it suggests that a comparatively small fine-tuned Stable Diffusion model can outperform much larger multimodal models on figurative language.","feed_headline":"Diffusion RL model beats GPT-4o and Grok-3 at metaphors","feed_subtitle":"Rhet2Pix keeps images figurative—“my bedroom is a pig sty” no longer shows literal pigs.","key_machinery":"The central object is a two-layer Markov decision process. The outer MDP models the staged progression of prompts: the state is the image generated at a stage and the action is the semantic increment added to the prompt; advantages are computed with Generalized Advantage Estimation. The inner MDP models diffusion denoising: the state is the latent $x_t$ with prompt and timestep, the action is the next denoised latent $x_{t-1}$, and the terminal reward is propagated to each step via $\\hat{A}^{(j)}_t = \\gamma_{\\text{denoise}}^t \\hat{A}_j$. This discounting is what the paper claims solves reward sparsity, because every adjacent action pair gets a nonzero advantage. Policy updates use a PPO clipped surrogate over shuffled action pairs.","core_discovery":"Rhet2Pix establishes that the literal-visual bias in rhetorical image generation can be corrected by separating semantic planning from pixel synthesis. In the outer layer, an LLM extracts key factors (subject, vehicle, theme, emotion) and builds a sequence of staged prompts, each adding context while preserving the subject. In the inner layer, the denoising trajectory of a Stable Diffusion model is optimized with PPO, where each outer-layer advantage is discounted backward across denoising steps so that early actions receive credit for the final rhetorical alignment. A combined reward penalizes the presence of the metaphorical vehicle and rewards staged semantic alignment and aesthetic quality. The paper reports that on its benchmark Rhet2Pix achieves a Sum-alignment of 0.6671 versus 0.4192 for Grok-3, and qualitatively produces images that depict the intended subject rather than the literal vehicle.","pith_inferences":["The paper's own limitation note admits that a fundamentally new backbone beyond text-image embedding similarity may be needed; this suggests the gains may be bounded by the CLIP embedding space used for both reward and evaluation.","A natural extension would be to test Rhet2Pix on other figures of speech such as hyperbole or irony, which may not map to concrete visual subjects as directly as metaphor and simile.","The reported elemental-alignment advantage for Rhet2Pix (0.4017 vs 0.1741) may partly reflect that the vehicle penalty is hand-tuned on the same distribution, so a distribution shift in rhetorical prompts could erode the gap."],"forward_implications":["If correct, small diffusion models fine-tuned on rhetorical data can outperform general MLLMs on figurative language without architectural changes.","The two-layer MDP with discounted rewards could be applied to other long-horizon generation tasks where feedback is only available at the end, such as text-to-video or layout generation.","The vehicle-penalty reward offers a reusable recipe for preventing literal over-alignment in any task with metaphorical or abstract semantics.","Because only the final prompt is used at inference, the method adds no extra generation cost, only a fine-tuning cost."],"supporting_citations":[{"why":"Provides the Stable Diffusion backbone that Rhet2Pix fine-tunes.","marker":"Rombach et al. 2022b"},{"why":"Supplies the DDIM sampling algorithm used for denoising trajectories.","marker":"Song, Meng, and Ermon 2020"},{"why":"Contributes the FLUTE dataset, filtered for metaphor and simile samples.","marker":"Chakrabarty et al. 2022b"},{"why":"Defines DDPO, a key RL-for-diffusion baseline and the reward-sparsity setup Rhet2Pix extends.","marker":"Black et al. 2024"},{"why":"Introduces DPOK, another RL fine-tuning baseline used for comparison.","marker":"Fan et al. 2023"},{"why":"Provides LoRA, the parameter-efficient fine-tuning method for the UNet.","marker":"Hu et al. 2022"},{"why":"Powers the LAION aesthetics predictor used in the reward function.","marker":"Schuhmann et al. 2022"},{"why":"Supplies the diffusion-policy optimization view that motivates the agentic multi-step formulation.","marker":"Ren et al. 2024"},{"why":"Supports the claim that denoising steps at different timesteps have unequal effects, motivating discounted reward propagation.","marker":"Hu et al. 2025"},{"why":"Provides GPT-4o, a primary MLLM baseline that Rhet2Pix claims to outperform.","marker":"OpenAI et al. 2024"}],"fun_headline_variants":["Rhet2Pix uses two-layer RL to nail rhetorical images","Beats GPT-4o, Grok-3 on figurative image generation","No literal pigs: diffusion RL maps 'pig sty' to mess","Two-layer MDP diffusion beats SOTA MLLMs on rhetoric","Rhetorical text-to-image tamed by policy-optimized diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that the CLIP-similarity scores used to measure alignment are valid measures of rhetorical fidelity, but these scores appear to be the same terms used to train the model, and no human or external benchmark verifies that they correspond to correct metaphorical images.","fun_headline_variants_meta":{"raw":{"variants":["Rhet2Pix uses two-layer RL to nail rhetorical images","Beats GPT-4o, Grok-3 on figurative image generation","No literal pigs: diffusion RL maps 'pig sty' to mess","Two-layer MDP diffusion beats SOTA MLLMs on rhetoric","Rhetorical text-to-image tamed by policy-optimized diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000805,"raw_usage":{"total_tokens":3542,"prompt_tokens":958,"completion_tokens":2584,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":2490}},"tokens_in":574,"tokens_out":2584,"duration_ms":20825,"temperature":1.0,"reasoning_tokens":2490,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:00:11.859560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a human study where annotators choose which image best captures the figurative meaning of a set of rhetorical prompts, comparing Rhet2Pix against GPT-4o and Grok-3. If annotators do not prefer Rhet2Pix images at a clear rate, or if the reported alignment scores do not predict those choices, the central claim would be refuted.","supporting_citations":[],"review_version":1}