{"id":"2fd24584-2ff1-47b2-baa5-bcba4df500cc","arxiv_id":"2508.18032","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A stage-aware reinforcement learning recipe for text-to-image generation that reports benchmark gains, but whose central reward formulas are inverted and whose reasoning reward is coupled to the final outcome reward.","lead":"Visual-CoG trains an autoregressive text-to-image model with three reinforcement rewards, one for reasoning about the prompt, one for refining intermediate sketches, and one for checking the final image. The paper reports large benchmark gains, but the reward equations as written reward the wrong behavior and the reasoning reward is not independent of the final check.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 5 and Eq. 7 define rewards that increase with error, inverting the optimization; teacher model τ is unspecified, so the stage-aware mechanism is not supported as written.","rationale":"The reader's weakest assumption focuses on Eq. 5 and the undefined teacher model τ. I agree that the process-refining reward is a load-bearing, unspecified component. However, I identify the sign inversion in both Eq. 5 and Eq. 7 as the more fundamental issue: as written, the rewards are monotonically increasing functions of the very errors they are meant to penalize. This is not a missing detail but a mathematical inconsistency that would make the optimization actively harmful if implemented literally. The undefined teacher model compounds this by making the reward non-computable and non-reproducible. Since the central claim is that stage-aware rewards provide effective guidance, and the stated rewards are inverted, the empirical results cannot be explained by the proposed mechanism. The reader's verdict of REJECT is therefore appropriate: the method is not validated as described. I find no reason to change the verdict, and the concrete test I propose would definitively resolve whether the sign error is a typo or the actual implementation.","tokens_in":12710,"tokens_out":4396,"duration_ms":42629,"concrete_test":"Analytical sign check: for a single object with correct count (deviation 0), Eq. 7 gives exp(0)=1; for one extra object (deviation 1), it gives exp(1/τ)>1. Thus the reward is higher for the wrong count. The same holds for Eq. 5: as ||G(pθ)-G(pτ)|| increases, Rp increases. If the authors' implementation contains a minus sign inside the exponential (as any sensible reward would), then the published Eqs. 5 and 7 do not describe the training objective. Ask the authors to release the exact code for computing Rp and Rn and to specify the teacher model τ (checkpoint, architecture, and data). If the code matches the equations, train without Rr and Ro (only Rp) on GenEval counting; the metric should degrade.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that stage-aware rewards improve T2I alignment. For this to hold, each reward must increase when the corresponding error decreases. Eq. 5 defines Rp = exp(||G(pθ)-G(pτ)||_p), which is monotonically increasing in the discrepancy between the policy's reconstructed image and the teacher's image. Maximizing Rp therefore drives the policy away from the teacher distribution, the opposite of 'process refining.' Eq. 7 defines Rn = (1/Nn) Σ exp(|En(D(Ie,i)) - y*|/τ), which is monotonically increasing in count deviation; an image with the wrong number of objects receives a higher reward than one with the correct count. Since Ro = Rn + Rc + Rs + Rh, the outcome reward is corrupted for counting, and since Rr = Ro(P') - Ro(P), the semantic reasoning reward inherits the same inversion. Additionally, the teacher model τ is never specified: no architecture, training data, or source is given, so Eq. 5 cannot be computed or reproduced. These are not implementation details; they are the mechanism itself. The ablations in Table 4 show process refining improves counting by 7.96%, which is inconsistent with the stated reward direction. Unless the equations are corrected (e.g., exp(-||...||) and exp(-|...|/τ)) and τ is defined, the reported improvements cannot be attributed to the proposed stage-aware guidance.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Visual-CoG, a stage-aware reinforcement learning framework for text-to-image generation. Generation is decomposed into three stages—semantic reasoning, process refining, and outcome evaluation—with a distinct reward for each: Rr in Eq. (4), Rp in Eq. (5), and Ro in Eqs. (6)–(9). The authors train Show-o with this staged reward and report large gains on GenEval, T2I-CompBench, and a newly introduced VisCog-Bench. The paper also introduces VisCog-Bench, a 100-prompt benchmark with unusual position, composition, color, and reasoning subtasks. The central claim is that providing immediate, stage-aware rewards improves alignment on multi-attribute and ambiguous prompts beyond a single final reward.","tokens_in":12888,"tokens_out":6335,"duration_ms":61686,"significance":"If the method worked as described, the claimed improvements are substantial and the staged-reward idea is a plausible and interesting direction for autoregressive text-to-image alignment. The paper also contributes a benchmark aimed at unusual and reasoning-demanding prompts, which could be useful to the community. However, the central reward equations contain two sign inversions that make the optimization objective as written the opposite of the stated goal, and the teacher model in the process-refining stage is never specified. These issues are not cosmetic: they undermine the mechanism that the paper's results are attributed to. The strengths are the clear problem framing and the breadth of evaluation; the weaknesses are the mathematical errors and the underspecified teacher model, which together prevent the reported results from being reproduced or interpreted as supporting the proposed method.","major_comments":[{"comment":"The process-refining reward is defined as Rp = exp(||G(pθ) - G(pτ)||_p). This is monotonically increasing in the discrepancy between the policy reconstruction and the teacher reconstruction. Maximizing Rp therefore encourages the policy to move away from the teacher distribution, which is the opposite of the stated 'process refining' objective. As written, this reward cannot explain the counting improvement attributed to Rp in Table 4. Please clarify whether a negative exponent (or a minimization objective) was intended, and define the teacher model τ and its preferred distribution pτ—the architecture, training data, and source are not given anywhere in the paper, so Eq. (5) is not computable or reproducible as written.","section":"§3.2, Eq. (5)"},{"comment":"The counting reward is defined as Rn = (1/N_n) Σ_i exp(|En(D(I_{e,i})) - y*_{in}|/τ). This function increases with the absolute counting error: an image with the wrong number of objects receives a higher reward than one with the correct count. The accompanying text states that the design 'penalizes larger deviations more severely,' but the formula rewards them. This inversion corrupts the outcome reward Ro in Eq. (9) and, through Eq. (4), also distorts the semantic reasoning reward Rr. If the intended formula is exp(-|...|/τ), the equations must be corrected and all experiments dependent on them must be rerun before the reported counting results can be attributed to this reward.","section":"§3.2, Eq. (7)"},{"comment":"The semantic reasoning reward is a difference of the outcome reward Ro for the reasoning prompt and the original prompt. Because Ro is itself the reward being optimized, Rr measures whether the reasoning prompt produces a better final image under the same proxy, not whether the model performed any form of reasoning. The VisCog-Bench reasoning evaluation in Table 5 uses Qwen2.5-VL as an automatic judge, and the benchmark has only 100 prompts with no reported confidence intervals or human-judgment agreement. The claim that the method improves 'reasoning' is therefore not yet independently established; a reasoning-specific reward or a more controlled evaluation with inter-annotator agreement would be needed.","section":"§3.2, Eq. (4) and §4.4"}],"minor_comments":[{"comment":"The sentence introducing Eq. (10) says 'the loss is defined in Eq. 2: R = Rr + Rp + Ro,' but Eq. (2) is the policy-gradient loss L(θ), not a reward. Please separate the definition of the scalar reward R from the policy-gradient objective.","section":"§3.2, Eq. (10)"},{"comment":"In Eq. (5), the symbol p is used both for the norm order and for the policy distributions pθ and pτ, which is confusing. Please use a different letter for the norm, e.g., ||·||_q.","section":"§3.2, Eq. (5)"},{"comment":"The claimed per-reward improvements (6.99% for position from Rr, 7.96% for counting from Rp, and 9.53% for color from Ro) are not directly readable from the table because each ablation row combines multiple rewards. Please specify the exact row comparisons used to compute these numbers.","section":"Table 4"},{"comment":"VisCog-Bench contains only 100 prompts, and the human evaluation in Table 5 is reported without confidence intervals or statistical significance tests. Please provide variance estimates or significance testing.","section":"§4.1 and §4.4"},{"comment":"The checklist states that all code and data will be released, but no code or data appendix is included in the submission. Please clarify the availability of the promised artifacts at submission time.","section":"Reproduction Checklist"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the two sign inversions in Eqs. (5) and (7) and the unspecified teacher model are serious enough that the current manuscript is not reproducible as written. I believe these issues are fixable in principle, but the authors must correct the equations, fully specify the teacher distribution, and rerun all affected experiments. Given the very large reported gains, I would also ask the authors to verify that the GenEval-style templates used to create training prompts do not leak into the evaluation benchmarks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. The stage-aware reward decomposition (semantic reasoning, process refining, outcome evaluation) is a sensible idea, and the VisCog-Bench, though small, targets a real evaluation gap. But the central equations as written invert the optimization: Eq. 5 rewards divergence from the teacher model, and Eq. 7 rewards counting error. That is not a detail; it is the mechanism.\n\nOn the positive side, the paper correctly identifies a limitation of final-only rewards in RL for T2I, and the three-stage credit assignment is a plausible remedy. The ablations in Table 4 show each reward component helps, which is interesting if the implementation actually uses the correct signs. The benchmark of 100 prompts for unusual position/composition/color and reasoning is a useful idea, though small.\n\nNow the soft spots, in rough order of severity.\n\nFirst, Eq. 5 defines Rp = exp(||G(pθ)-G(pτ)||_p). Maximizing that pushes the policy away from the teacher distribution, the opposite of \"process refining.\" Eq. 7 defines Rn as the average of exp(|count - target|/τ), which increases with counting error. The text says both should penalize deviation. Unless the exponential signs are typos (exp(-...)), the policy is being rewarded for doing worse. I think they are likely typos, but as presented the method is not coherent.\n\nSecond, the teacher model τ that defines the preferred distribution pτ is never described: no architecture, training data, or source. Eq. 5 cannot be computed or reproduced. That is a load-bearing gap.\n\nThird, the semantic reasoning reward Rr = Ro(P') - Ro(P) uses the same outcome evaluator that appears in the final reward. That makes the \"reasoning\" signal dependent on the same metric the policy is trained to exploit; it is not an independent measure of reasoning quality. This is a legitimate concern, though less damning than the sign errors.\n\nFourth, the empirical section has no error bars, no seeds, and omits T2I-R1, a closely related CoT+RL baseline cited in the paper. Also, the reproduction checklist at the end asserts \"yes\" to a code appendix and statistical tests that do not exist in the manuscript. That is not a great look.\n\nBottom line: the idea is worth a second look, but the paper as written cannot be relied on. If the sign errors are just typos and the teacher model is specified, the method could be salvaged with substantial revision. As is, I would tell the authors to fix the equations, define τ, add the missing baseline, and release code.\n\nIf you are the editor, I would send it to peer review rather than desk-reject, because the fixes are identifiable and the benchmark has value; but the reviewers should be instructed to check the equations carefully. I would not cite this version.","headline":"Stage-aware reward decomposition is a good idea, but the two central reward equations are sign-inverted as written and the teacher model is unspecified; major revision before the results can be trusted.","tokens_in":13562,"tokens_out":3805,"would_cite":false,"duration_ms":34304,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A stage-aware reward chain—semantic reasoning, process refining, and outcome evaluation—improves text-to-image alignment on multi-attribute and ambiguous prompts, raising GenEval overall from 68.29 to 83.86.","keywords":["text-to-image generation","reinforcement learning","stage-aware rewards","chain of guidance","semantic reasoning","autoregressive unified models","mask token prediction","visual cognition benchmark"],"falsifier":"Train Visual-CoG with the teacher model replaced by the policy itself, so $p_\\tau = p_\\theta$ in Eq. 5: the authors state it would be meaningless for the policy to recover its own masked output, so the process-refining reward should stop contributing and GenEval should drop toward the $R_r+R_o$ ablation's 76.23. If performance stays near 83.86, the process-refining reward is not doing the claimed work. A simpler check: compute Eq. 5 as printed, without a negative sign, and verify whether high-reward samples are closer to or farther from the teacher distribution.","tokens_in":12387,"feed_emoji":"🎨","tokens_out":10089,"duration_ms":89657,"temperature":0.7,"pith_summary":"Visual-CoG is a training strategy for autoregressive text-to-image models that replaces a single final reward with three stage-aware rewards: semantic reasoning, process refining, and outcome evaluation. The paper argues that final-only reward signals, as used by prior CoT and RL text-to-image works, make it hard to tell which stage of generation helped and therefore lead to suboptimal policies. By rewarding each stage immediately, Visual-CoG lets the model allocate credit to early interpretation and intermediate refinement, not just the finished picture. On the GenEval benchmark, the method lifts the Show-o baseline from 68.29 to 83.86 overall, and on the authors' new VisCog-Bench it raises overall alignment from 58.07 to 77.50. If the result holds, stage-aware rewards are a practical way to make text-to-image models follow multi-attribute and ambiguous prompts.","feed_headline":"Stage-by-stage rewards lift text-to-image score by 15 points","feed_subtitle":"A three-stage visual chain of guidance raises GenEval overall alignment from 68.3 to 83.9.","key_machinery":"The machinery is a stage-aware reward chain, $R = R_r + R_p + R_o$, plugged into a GRPO-style clipped policy gradient. $R_r$ is a contrastive semantic-reasoning reward: generate two images from the original prompt and from a reasoned prompt under a fixed seed, then take the difference of their outcome scores. $R_p$ is a process-refining reward defined as an exponential norm between the policy's reconstructed masked patch distribution and a teacher model's preferred distribution $p_\\tau$. $R_o$ is a rule-based outcome reward combining counting, color, and spatial checks, via an open-vocabulary detector and a vision-language classifier, with a holistic HPS aesthetic score. This per-stage reward design is what allows credit assignment to early reasoning and intermediate refinement steps rather than only to the final rendered image.","core_discovery":"The central claim is that decomposing text-to-image synthesis into a chain of guidance and giving an immediate reward at each link improves alignment more than optimizing only the final image. The three stages are: (1) semantic reasoning, where a language model rewrites the prompt and the reward is the outcome-score difference between images generated from the reasoned prompt and the original prompt; (2) process refining, where a masked patch reconstruction step is rewarded for matching a teacher model's preferred distribution; and (3) outcome evaluation, where rule-based spatial, counting, and color checks plus an aesthetic preference score score the final image. The rewards are summed into $R = R_r + R_p + R_o$ and optimized with a clipped policy-gradient objective. In the paper's experiments, this raises Show-o from 68.29 to 83.86 on GenEval, reaches 78.92 on T2I-CompBench Color and 43.71 on Spatial, and lifts VisCog-Bench overall from 58.07 to 77.50.","pith_inferences":["An implication not drawn in the paper: the semantic-reasoning reward is a contrast between two images scored by the same outcome evaluator, so if that evaluator has systematic biases, the reasoning stage may learn to exploit the scorer instead of resolving genuine ambiguity.","The paper leaves unresolved how the teacher distribution $p_\\tau$ is obtained; without a specified teacher, the process-refining reward is not reproducible from Eq. 5 as written, since the exponential norm with no negative sign rewards disagreement rather than closeness.","Because VisCog-Bench's unusual-position, unusual-composition, and unusual-color subtasks each contain only 20 prompts, the reported 19-point overall gain on that benchmark likely carries nontrivial sampling uncertainty; extending those subtasks would test whether the effect is stable."],"forward_implications":["On GenEval, stage-aware rewards take Show-o from 68.29 to 83.86 overall, with the position subtask jumping from 31.03 to 79.00 and counting from 66.25 to 80.94.","On T2I-CompBench, Visual-CoG sets the paper's best numbers of 78.92 on color and 43.71 on spatial among the compared models.","On VisCog-Bench, overall alignment rises from 58.07 to 77.50, with the unusual-position subtask improving from 28.75 to 77.50.","Ablations attribute distinct roles to each reward: $R_r$ lifts position by 6.99 points, $R_p$ lifts counting by 7.96 points, and $R_o$ lifts color by 9.53 points; removing any stage lowers overall from 83.86."],"supporting_citations":[{"why":"Supplies the base autoregressive model, Show-o, and the baseline scores that Visual-CoG improves.","marker":"(Xie et al. 2024)"},{"why":"Provides the GenEval benchmark and the unusual-position, unusual-composition, and unusual-color prompt subtasks used in VisCog-Bench.","marker":"(Ghosh, Hajishirzi, and Schmidt 2023)"},{"why":"Provides the T2I-CompBench benchmark for compositional attribute evaluation.","marker":"(Huang et al. 2023)"},{"why":"Supplies the group-relative policy optimization objective used in Eq. 2.","marker":"(Shao et al. 2024)"},{"why":"Motivates the rule-based outcome reward design in the outcome evaluation stage.","marker":"(Guo et al. 2025a)"},{"why":"Provides the HPS human-preference score used as the holistic component $R_h$ of the outcome reward.","marker":"(Wu et al. 2023)"},{"why":"Provides the open-vocabulary detector used for spatial, counting, and color consistency checks.","marker":"(Liu et al. 2024)"},{"why":"Establishes the masked-token prediction formulation underlying the process-refining stage.","marker":"(Chang et al. 2022)"},{"why":"Serves as the automatic evaluator for the reasoning subtask in VisCog-Bench.","marker":"(Bai et al. 2025)"},{"why":"Generates the training prompts and the reasoning-task prompts used for VisCog-Bench.","marker":"(Yang et al. 2025a)"}],"fun_headline_variants":["Stage-wise rewards sharpen text-to-image alignment","Visual-CoG: reward every step, not just the final image","Chain-of-guidance boosts T2I scores by up to 19%","Three-stage guidance beats final-only rewards in T2I","Rewarding each stage lifts GenEval by 15 points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The process-refining reward assumes that the teacher model $\\tau$ provides a preferred distribution that is genuinely better to imitate, yet the paper never specifies $\\tau$'s architecture, training data, or origin; if $\\tau$ is not a better generative distribution, or if the sign in Eq. 5 is wrong as printed, the intermediate guidance is undefined or actively pushes the policy to diverge.","fun_headline_variants_meta":{"raw":{"variants":["Stage-wise rewards sharpen text-to-image alignment","Visual-CoG: reward every step, not just the final image","Chain-of-guidance boosts T2I scores by up to 19%","Three-stage guidance beats final-only rewards in T2I","Rewarding each stage lifts GenEval by 15 points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000291,"raw_usage":{"total_tokens":1713,"prompt_tokens":974,"completion_tokens":739,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":653}},"tokens_in":590,"tokens_out":739,"duration_ms":6536,"temperature":1.0,"reasoning_tokens":653,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:58:10.125105+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train Visual-CoG with the teacher model replaced by the policy itself, so $p_\\tau = p_\\theta$ in Eq. 5: the authors state it would be meaningless for the policy to recover its own masked output, so the process-refining reward should stop contributing and GenEval should drop toward the $R_r+R_o$ ablation's 76.23. If performance stays near 83.86, the process-refining reward is not doing the claimed work. A simpler check: compute Eq. 5 as printed, without a negative sign, and verify whether high-reward samples are closer to or farther from the teacher distribution.","supporting_citations":[],"review_version":2}