{"id":"21f97c86-03b9-40b0-bc64-b0d3474fc41e","arxiv_id":"2505.13031","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A unified vision-language model uses a three-stage pipeline with supervised fine-tuning and a new RL algorithm (RGPO) to generate images from instructions that require reasoning, such as math or world knowledge.","lead":"This paper presents MindOmni, a multimodal AI model that both understands and generates images, and it is trained to reason in text before drawing. It introduces a reinforcement learning method, RGPO, that uses image-text similarity as a reward to improve reasoning-based image generation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The RGPO consistency reward is CLIP similarity while WISE's evaluation metric is never specified; if WISE uses embedding-based semantic matching, the headline WISE gains reflect reward overfitting, not reasoning.","rationale":"The paper is a plausible systems contribution: three-stage training, RGPO with a CLIP consistency reward, and extensive ablations including stage contributions and KL coefficients. The WISE improvement from Stage 2 to Stage 3 (0.54 to 0.60) is modest but consistent. However, the central claim that RGPO 'unleashes reasoning generation' is only as strong as the evaluation that measures reasoning. The consistency reward in Sec. 3.4 is explicitly CLIP cosine similarity between the generated image and the ground-truth prompt. WISE is a semantic alignment benchmark whose evaluation function is not described in this manuscript. If WISE uses a CLIP-style embedding match, then the RL objective and the benchmark are the same measurement, and the reported WISE gains (including the 0.71 with higher-quality data) are expected from reward maximization independent of any reasoning chain. The paper's Table 9 isolates the consistency reward's contribution (~3.7 WISE points), which is comparable to the Stage-3 gain in Table 5 (6 points), reinforcing that the reward, not the CoT, drives the improvement. The concrete test of inspecting WISE's official metric and computing the correlation with the reward will determine whether this concern lands. If the metric is independent, the central claim is substantially supported; if not, the conditional verdict should be maintained or downgraded. I therefore agree with the reader's assessment and see no reason to change the verdict.","tokens_in":17174,"tokens_out":10220,"duration_ms":107972,"concrete_test":"Inspect the official WISE scoring code from reference [23] and identify the exact function behind the 'Overall' column in Table 3. Then, using the released MindOmni checkpoint and the WISE prompt set, compute the Spearman correlation between the per-prompt CLIP cosine similarity used in Eq. (3.4) and the per-prompt WISE score. If WISE's metric is CLIP-based or the correlation is high (rho > 0.7), the WISE gains should be attributed to reward overfitting; if WISE uses an independent language/VQA-based judge and the correlation is low, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"MindOmni's central claim rests on WISE results (Table 3: 0.60/0.71; Table 5: Stage 3 adds +0.06). The only image-level reward in RGPO (Sec. 3.4) is the consistency reward: CLIP cosine similarity between the generated image and the reference ground-truth prompt. The paper never specifies how WISE's 'overall' score is computed, nor whether it uses CLIP or a correlated embedding-based metric. If it does, RGPO is directly maximizing a proxy for the benchmark's scoring function; the WISE gains would then be a foreseeable product of reward optimization, not a demonstration of general reasoning generation. The effect size is material: Table 9 shows the consistency reward contributes about 3.7 WISE points (56.1 to 59.8). Since Appendix A.1 only states the resolution used for WISE and the Limitation section does not address this overlap, the gap is unaddressed. This is load-bearing because the paper's novel contribution—reasoning generation via RGPO—is evidenced primarily by WISE.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MindOmni, a unified vision-language model that performs both image understanding and generation, and introduces a three-stage training pipeline. The first stage trains a connector between a VLM (Qwen2.5-VL) and a decoder-only diffusion module (OmniGen); the second stage performs supervised fine-tuning on chain-of-thought (CoT) instruction data; the third stage applies the proposed Reasoning Generation Policy Optimization (RGPO), an RL algorithm with a format reward, a CLIP-based consistency reward, and separate text and image KL regularizers. The authors report state-of-the-art scores on WISE (0.71), GenEval (0.83), and DPG-Bench (83.0), while preserving understanding performance (MMMU 51.6, MMBench 83.2). The central claim is that RL with multimodal feedback, especially the CLIP consistency reward, improves the model's ability to generate images from prompts that require explicit reasoning.","tokens_in":17423,"tokens_out":5873,"duration_ms":54042,"significance":"If the results hold, the paper would make a useful contribution to reasoning-driven image generation by demonstrating that reinforcement learning can be applied to a unified multimodal model with diffusion-based generation. The three-stage pipeline is clearly described, and the ablations (Tables 5, 7, 8, 9) provide useful information about stage contributions, KL coefficients, group sizes, and reward functions. The code release promise and the relatively honest reporting of some negative controls (e.g., 'w/o thinking') are positive aspects. However, the headline WISE result is confounded by the use of additional external data in the starred variant, and the paper never specifies WISE's evaluation metric or analyzes its overlap with the CLIP-based consistency reward. Additionally, the ablation evidence attributes only a small, possibly non-significant gain to the consistency reward itself, which weakens the claim that multimodal feedback is the key driver of RGPO. These issues need to be resolved before the paper's central claims can be fully accepted.","major_comments":[{"comment":"The headline WISE score of 0.71 is reported for MindOmni*, the variant trained with additional higher-quality external data (references [25, 2]) and evaluated with thinking mode. The abstract and introduction state this as the method's result without qualification, while the unstarred MindOmni achieves 0.60. Because the gain from 0.60 to 0.71 is confounded with the change in training data, it cannot be attributed to the proposed RGPO algorithm. The paper should either report the unstarred result in the abstract or explicitly separate the contributions of data quality and RGPO, for example by training the starred variant without Stage 3.","section":"Abstract; Table 3; Sec. 4.2"},{"comment":"The consistency reward in RGPO is defined as the CLIP cosine similarity between the generated image and the ground-truth prompt, but the manuscript never specifies how WISE (reference [23]) computes its score. If WISE relies on CLIP or a similar embedding-based semantic matching, then RGPO is directly optimizing a proxy of the benchmark's scoring function, and the reported WISE gains from RL would reflect reward overfitting rather than a general improvement in reasoning generation. The paper must state WISE's evaluation metric and provide an analysis of the overlap, for example by correlating the CLIP consistency reward with WISE scores or by ablating the reward with an alternative semantic-alignment measure (e.g., a VQA-based reward) to show that the effect is not driven by metric overlap.","section":"Sec. 3.4; Table 3; Appendix A.1"},{"comment":"The ablation in Table 9 shows that the format reward improves WISE from 56.1 to 59.4, while adding the consistency reward only increases it to 59.8, a gain of 0.4 points. The text states that 'Consistency reward plays a key role in steering reasoning generation,' but this is not supported by the numbers. Moreover, the format reward is a text-only reward that checks the presence of <think> and <answer> tags; therefore the ablation does not demonstrate that multimodal feedback (the claimed novelty of RGPO over GRPO) is the main driver of the observed WISE improvement. The authors should present the individual reward contributions more fairly and, if they wish to claim that the consistency reward is important, provide additional evidence such as an evaluation on a held-out reasoning-generation benchmark.","section":"Sec. 4.4; Table 9"}],"minor_comments":[{"comment":"The text says the model's understanding performance after RL is 'nearly identical to the original VLM backbone, with only a 0.1% average difference,' but Table 4 shows differences of +0.5 (MMMU), -0.2 (MMB), and -0.8 (RWQA), which average -0.17 in signed difference and 0.5 in absolute difference; please reconcile this statement with the table.","section":"Sec. 4.2; Table 4"},{"comment":"'MindOmni (w/o thinking)' scores 0.43 on WISE, which is below the OmniGen baseline of 0.44; the paper does not comment on this regression. It would be helpful to explain why the base model without CoT thinking performs worse than the diffusion-only baseline.","section":"Table 3"},{"comment":"The text KL divergence term is written as π_ref/π_θ - log(π_ref/π_θ) - 1, which is a form of KL(π_ref||π_θ), but the notation D^T_KL(π_θ||π_ref) suggests the opposite direction; please clarify the definition, the expectation, and whether the term is per-token or sequence-level.","section":"Eq. (5)"},{"comment":"The limitation section mentions that CoT content is in plain text, but it does not mention the potential reward-evaluation overlap on WISE or the data confound for the starred variant; these are directly relevant to the paper's main claims.","section":"Appendix A.3"},{"comment":"There are several typos and minor errors, including 'Ldiffuison' in Eq. (2), 'responss' in the Fig. 10 caption, and 'Qualitatvie' in Appendix A.4; the manuscript should be proofread.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper's central novelty claim is carried by the WISE results, but as it stands the headline number is confounded by additional external data, and the reward-evaluation overlap risk is not addressed. The ablations actually show that the consistency reward contributes only a small gain relative to the format reward, which further weakens the claim that multimodal feedback is the key enabler. I would ask the authors to clarify WISE's evaluation metric, analyze the overlap, and report unconfounded results in the abstract. These are addressable in a revision, so I do not recommend rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper is a legitimate engineering contribution—unifying Qwen2.5-VL with a diffusion decoder and applying GRPO-style RL to generate explicit chain-of-thought for image generation. The three-stage training is sensible, and the ablations (Tables 5, 7, 8, 9) are genuinely informative: they show SFT matters, RL adds a small but real increment, and the consistency reward drives the WISE gains. The qualitative examples in the figures are impressive, and the model retains understanding performance. The math is standard GRPO with two extra KL terms; no red flags there. Datasets are not released, only the model, so reproducibility rests on the promised code.\n\nBut the headline claims should be read carefully. The 'state-of-the-art' WISE 0.71 and GenEval 0.83 come from MindOmni*, a variant trained with additional higher-quality data (GPT-4o and BLIP-3o outputs, presumably). The main MindOmni gets 0.60 on WISE, which is actually below BAGEL's 0.70. So the SOTA claim rests on the data-boosted model, and the margin over BAGEL is 0.01. That is not a robust claim.\n\nThe bigger issue is the reward-evaluation overlap. The consistency reward in RGPO is CLIP cosine similarity between generated image and ground-truth prompt. WISE is a semantic evaluation for text-to-image generation; the paper never specifies its scoring metric. If WISE uses an embedding-based semantic alignment (CLIP or a similar space), then RGPO is directly optimizing a proxy for the benchmark. Table 9 shows the consistency reward adds about 3.7 WISE points (56.1 to 59.8). That is a material chunk of the RL gain. The paper does not analyze this overlap, and the limitation section only mentions plain-text CoT. This is a load-bearing gap for the central claim that RL improves reasoning generation, not just alignment to a particular scoring function.\n\nNone of this makes the paper worthless. The architecture and training recipe will be useful to the community, and the ablation structure is a good model for future work. But the authors need to (a) report WISE's scoring criterion or run a variant with a different consistency metric, (b) compare MindOmni* against BAGEL and others with matched training data, and (c) show that the RL gains transfer to non-CLIP evaluations.\n\nWho is this for? People building unified multimodal models and anyone thinking about RL for generative tasks. It is not a conceptual breakthrough, but it is a solid, if over-sold, empirical paper. I would send it to peer review—the topic is timely and the experiments are mostly well-designed—but I would not accept it without the reward-overlap analysis and a fairer comparison.","headline":"Solid RL-for-generation pipeline with honest ablations, but the headline reasoning-generation gains are partly a data-boosted variant and the CLIP consistency reward likely overlaps with WISE's scoring, so the central claim needs a careful referee.","tokens_in":798,"tokens_out":895,"would_cite":true,"duration_ms":44168,"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":"MindOmni claims that reinforcement learning with a CLIP consistency reward lets a unified vision-language model generate explicit chains of thought before images, reaching 0.71 on WISE and 0.83 on GenEval.","keywords":["reasoning generation","vision language model","reinforcement learning","chain-of-thought","text-to-image generation","unified multimodal model","CLIP consistency reward","RGPO"],"falsifier":"Score MindOmni's WISE outputs with a human panel or with detection-based checks, such as counting objects or verifying colors and spatial relations, instead of embedding similarity; if the 0.71 overall score falls toward the 0.43 no-thinking baseline while CLIP-based scores stay high, the consistency reward is optimizing the evaluator rather than the reasoning itself.","tokens_in":16996,"feed_emoji":"🧠","tokens_out":7456,"duration_ms":70443,"temperature":0.7,"pith_summary":"MindOmni is a unified multimodal model that both understands and generates images, and the paper argues that a three-stage pipeline can make it produce explicit chains of thought before rendering. The stages are: pretraining a lightweight connector between a vision-language backbone and a decoder-only diffusion module, supervised fine-tuning on coarse-to-fine instruction data that includes reasoning, and a new reinforcement-learning step called Reasoning Generation Policy Optimization (RGPO). The central claim is that RGPO's multimodal consistency reward, CLIP similarity between the generated image and the ground-truth prompt, is what turns ordinary text-to-image generation into reasoning generation for prompts that need arithmetic, world knowledge, or spatial and temporal inference. If the claim holds, it offers a recipe for teaching generative models to reason aloud before drawing without sacrificing their basic generation or understanding abilities.","feed_headline":"Reinforcement learning raises reasoning-aware image scores to 0.71","feed_subtitle":"MindOmni adds a consistency reward so the model thinks in text before rendering, outperforming prior unified models on WISE.","key_machinery":"The central object is RGPO, a group-relative policy optimization algorithm modified for multimodal rollouts. During training, the policy emits a chain of thought in text and then a diffusion-decoded image; the reward combines a format check for the CoT tags with a consistency reward computed as CLIP cosine similarity between the generated image and the ground-truth prompt. Separate KL regularizers for text and image distributions stabilize training and prevent the model from forgetting basic generation and understanding during the RL stage.","core_discovery":"The paper's central claim is that reinforcement learning can be applied to the text-and-image generation policy of a unified model with a reward that looks at the rendered image, not just the text tokens. RGPO samples several chain-of-thought-and-image rollouts per prompt, scores them with a format reward, whether the CoT is wrapped in <think> and <answer> tags, and a consistency reward, CLIP cosine similarity between the generated image and the reference prompt, normalizes these scores into an advantage, and updates the policy with a clipped GRPO-style objective. Two KL regularizers, one on text tokens and one on visual features, keep the policy near its reference model. The paper reports that this raises the model's overall WISE reasoning-generation score to 0.60, or 0.71 with higher-quality data and thinking mode, keeps GenEval at 0.81/0.83, and preserves MMMU understanding at 51.6, close to the original backbone.","pith_inferences":["Because the consistency reward uses CLIP similarity and the paper does not state WISE's evaluation metric, a reader should test whether WISE also scores images by embedding similarity; if it does, part of the measured reasoning gain could be reward overfitting rather than general reasoning.","A testable extension is to replace the CLIP consistency reward with object-detection-based or question-answering-based rewards; RGPO only requires a differentiable reward, so the same pipeline could optimize for more concrete visual properties.","The same RL recipe could be applied to vision editing by defining a consistency reward on the edited region, such as VAE feature similarity to the reference, which the paper shows but does not train with RGPO.","The observed weakness that longer completions do not linearly improve WISE suggests a future direction of allocating reasoning length by problem difficulty, which the paper does not explore."],"forward_implications":["Explicit reasoning-before-generation can be trained with reinforcement learning, not just hand-written CoT templates.","The consistency reward ties language quality to image correctness; as RL proceeds, both consistency reward and output length increase.","Without the CoT SFT stage, applying RGPO alone yields weaker results (WISE 0.49 vs 0.60), so the SFT stage is load-bearing for the RL gain.","Separate KL regularizers let the model improve reasoning without sacrificing basic generation: GenEval stays at 0.83 and understanding benchmarks remain near the backbone.","The method generalizes across reasoning categories such as culture, time, space, biology, physics, and chemistry, with the largest margins over prior models in time and physics reasoning."],"supporting_citations":[{"why":"Supplies the vision-language backbone whose understanding and reasoning abilities MindOmni builds on.","marker":"[1]"},{"why":"Supplies the decoder-only diffusion module and the X2I generation data used as the base model.","marker":"[47]"},{"why":"Provides the GRPO objective that RGPO adapts for multimodal rollouts.","marker":"[33]"},{"why":"Establishes the CoT reinforcement-learning recipe and system-prompt style that inspires Stage 3.","marker":"[14]"},{"why":"Is the WISE benchmark on which the paper claims state-of-the-art reasoning generation at 0.71 overall.","marker":"[23]"},{"why":"Is the GenEval benchmark used to show that basic text-to-image generation is preserved at 0.83.","marker":"[12]"},{"why":"Is the main unified-model baseline outperformed on GenEval and MMMU.","marker":"[5]"},{"why":"Supports applying RL after SFT to improve visual reasoning in multimodal models.","marker":"[57]"}],"fun_headline_variants":["MindOmni applies RL to text-to-image, boosting reasoning scores to 0.71","RL with CLIP consistency reward improves reasoning generation in VLMs","Policy optimization with visual feedback raises WISE score to 0.71","Think-then-render: RL teaches vision language models to reason","RGPO: New RL method makes VLMs reason before generating images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The consistency reward in RGPO is CLIP cosine similarity between the generated image and the ground-truth prompt, and the paper does not specify how WISE computes its scores; if WISE also relies on the same embedding-based semantic matching, the RL signal is directly optimizing the benchmark, and the reported gains could reflect reward overfitting rather than general reasoning.","fun_headline_variants_meta":{"raw":{"variants":["MindOmni applies RL to text-to-image, boosting reasoning scores to 0.71","RL with CLIP consistency reward improves reasoning generation in VLMs","Policy optimization with visual feedback raises WISE score to 0.71","Think-then-render: RL teaches vision language models to reason","RGPO: New RL method makes VLMs reason before generating images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00072,"raw_usage":{"total_tokens":3202,"prompt_tokens":886,"completion_tokens":2316,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":2220}},"tokens_in":502,"tokens_out":2316,"duration_ms":15906,"temperature":1.0,"reasoning_tokens":2220,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:20:49.643676+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Score MindOmni's WISE outputs with a human panel or with detection-based checks, such as counting objects or verifying colors and spatial relations, instead of embedding similarity; if the 0.71 overall score falls toward the 0.43 no-thinking baseline while CLIP-based scores stay high, the consistency reward is optimizing the evaluator rather than the reasoning itself.","supporting_citations":[{"cited_title":"Advances in Neural Information Processing Systems pp","cited_arxiv_id":null,"evidence_quote":"Is the GenEval benchmark used to show that basic text-to-image generation is preserved at 0.83."}],"review_version":1}