{"id":"99bf46f6-31ff-4b5c-a406-2e4e22f1cf49","arxiv_id":"2607.13188","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CO2Jump couples text and image denoising through cross-modal attention and remasking, achieving best joint accuracy on three concurrent-generation tasks.","lead":"This paper introduces a new sampling method that lets an AI model generate text and images jointly, with each modality revising the other during generation. It reports better joint performance on image editing, maze solving, and nonogram puzzles, while stressing that the theoretical framework is heuristic and relies on closed commercial evaluation models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1 does not implement Eq. (7): image scores use z_text_t, not the sampled z_text_s, so the claimed within-step coupling is absent.","rationale":"The reader's weakest assumption targets exactly the Sec 4.2/4.3 approximation. My stress-test sharpens it into an internal inconsistency: the algorithm's own ordering makes it impossible for the image update to see the same-step text decisions. The cross-attention is computed on H_text_t, before any text tokens are born; mask-aware bias further suppresses masked text positions. Therefore the 'latest text decisions' entering the image score are not z_text_s but stale text beliefs from z_t. This is not a matter of disagreement with a heuristic approximation being slightly lossy; the stated chain-rule decomposition is not what Algorithm 1 executes. The empirical story — CO2Jump beats independent-parallel baselines because of within-step coupling — is unsupported by the described procedure. The proposed oracle test would settle whether the surrogate is faithful or whether the gains come from other components (e.g., remasking, rank normalization, entropy gating). Until that test is run, the conditional verdict is appropriate: the method may still be a useful heuristic, but its central mechanism claim and the attribution of empirical success to that mechanism need revision or explicit verification. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":21929,"tokens_out":9446,"duration_ms":120097,"concrete_test":"Run an oracle-chain-rule variant: after text death/birth in Algorithm 1, perform one extra forward pass at (z_text_s, z_image_t) and use the resulting image confidence for image death/birth. Compare oracle vs CO2Jump on ImgEditBench and JMaze-Test500 at NFE=128; also compute per-step Spearman correlation between CO2Jump's CoupledConf and the oracle's image scores. If oracle outperforms CO2Jump or rank correlation is low, the Sec 4.2 surrogate is unfaithful and the mechanism story requires retraction; if oracle does not improve, the claimed within-step coupling contributes nothing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (7) is the paper's central derivation: image update p(z_image_s | z_t, z_text_s, c) should condition on the text state sampled at the same step. In Algorithm 1, however, CoupledConf (Eq. 12) is computed before the text death/birth, using A^{image->text}_t from H_text_t with mask-aware bias B_t that downweights masked text positions. The text tokens that will be revealed in z_text_s are masked at t, so they are downweighted or excluded from CrossSignal (Eq. 10). The image update therefore conditions on the previous text state z_text_t, not on z_text_s. Given z_t, text and image updates are conditionally independent, which is precisely the uncoupled-parallel failure mode attributed to MMaDA-Parallel. The chain-rule factorization is thus not implemented; a heuristic score replaces it. The empirical gains may be real, but they cannot be attributed to the claimed within-step cross-modal correction. Table 3 ablates rank normalization, gating, and remasking, but never isolates this surrogate, so the central mechanism is untested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SC-CMJP, a framework for joint text/image generation with masked diffusion models, and CO2Jump, a training-free single-pass sampler that uses cross-modal attention-weighted confidence scores and a remasking jump to couple and self-correct the two modalities. The authors evaluate CO2Jump against MDM, ReMDM, and MMaDA-Parallel on three tasks: image editing with understanding, maze solving, and nonogram solving, using newly curated corpora (JEdit-1M, JMaze-200K, JNono-200K). The central claim is that CO2Jump achieves best joint performance and that its gains compound with more denoising steps because one modality's transition rates are functionals of the other modality's confidence within each step. The paper also positions this as a unified Markov Jump Process formulation with a chain-rule factorization that conditions the image update on the same-step text state.","tokens_in":22259,"tokens_out":8767,"duration_ms":97266,"significance":"If the central mechanism were realized as described, the paper would make a substantial contribution: a training-free, single-forward-pass sampler that couples modalities within a step, with self-correction via remasking, plus large-scale benchmarks with algorithmically verifiable ground truth for maze and nonogram tasks. The empirical setup is extensive and compares against representative baselines, and the datasets could be a useful community resource. However, the paper's own derivation is not implemented by the algorithm: the claimed within-step coupling is replaced by an attention-weighted surrogate that depends only on the previous state, and no rate matrix or target-preservation proof is provided for the asserted Markov Jump Process. The empirical results may be real, but they are currently attributed to a mechanism that is not instantiated.","major_comments":[{"comment":"The central chain-rule factorization Eq. (7) says the image update should condition on the text state sampled at the same step, z_text_s. Algorithm 1 does not do this. CoupledConf (Eq. 12) is computed before the text death/birth jumps (lines 10–12 before lines 15–19), and CrossSignal (Eq. 10) uses SelfConf_text extracted from z_t, with the mask-aware bias B_t in Eq. (9) downweighting masked text positions. The text tokens that will be revealed in z_text_s are masked at t and are therefore downweighted or excluded from the image score. Consequently, given z_t, the image update is a deterministic function of z_t (up to Gumbel noise) and is independent of the text update — exactly the uncoupled-parallel failure mode the paper attributes to MMaDA-Parallel in Section 3. The paper acknowledges this approximation in Section 4.2 ('approximating z_text_s with quantities already produced by the si","section":"§4.2–4.3, Eq. (7), Algorithm 1"},{"comment":"The paper frames SC-CMJP as a Coupled Markov Jump Process with modality-specific rate matrices whose intensities depend on the complementary modality's confidence. No such rate matrix is ever defined. Algorithm 1 is a deterministic top-k/bottom-k selection rule with hard quotas N_remask and N_unmask derived from idealized counts; there are no stochastic jump rates, and no theorem shows that the resulting transition kernel preserves the target distribution. The ReMDM posterior Eq. (4) is a valid modified posterior for a fixed sigma_t, but the score-dependent death/birth selection is not shown to correspond to Eq. (4) or to any Markov kernel. The term 'Markov Jump Process' is therefore not substantiated. Either provide a rate matrix and a proof of marginal preservation, or drop the MJP formalism and describe CO2Jump as a heuristic sampler.","section":"§2.2, §4.4"},{"comment":"The death and birth quotas are computed from idealized counts U_t^a = floor(alpha_t L_a) and M_t^a = floor((1-alpha_t)L_a), but the actual number of unmasked tokens at runtime can deviate from these idealized values due to previous floor operations, the score-dependent deaths, and the extra-unmask compensation. The paper does not show that the realized token counts follow the intended alpha_t schedule or that the valid-posterior constraint sigma_t <= (1-alpha_s)/alpha_t is satisfied state-by-state. This matters because the ReMDM-style remasking argument in Eq. (4) relies on the schedule being preserved. Please provide a runtime count analysis or explicitly state that CO2Jump intentionally deviates from the idealized schedule and characterize the deviation.","section":"§4.4, Algorithm 1 lines 16–19"},{"comment":"The headline image-understanding metric is a COCO-style mAP against a 'pseudoscene graph' constructed by Gemini on the model's own generated target image. No validation of this metric against human annotations or an independent oracle is reported, so absolute mAP values are difficult to interpret. In addition, Table 2 reports no confidence intervals; several improvements over the best baseline are small (e.g., Maze in-distribution joint accuracy: 0.469 vs. 0.461, +0.008), and Figure 5's monotonic-scaling claim is based on a single run without error bars. Please report multiple seeds, confidence intervals, and at least a sanity-check validation of the pseudo-grounding metric.","section":"§5.1, Tables 1–2, Figure 5"}],"minor_comments":[{"comment":"Typos: 'geneneration', 'undestanding', 'espectially' should be corrected.","section":"Abstract/Introduction"},{"comment":"The comment 'approx. z_text_s via attention' is misleading; the computed quantity is a function of z_text_t, not z_text_s. The comment should say so explicitly.","section":"Algorithm 1, line 12"},{"comment":"The monotonicity claim is strong for a single run; error bars or multiple seeds would make it convincing.","section":"Figure 5"},{"comment":"The Qwen3-VL-8B and pre-trained Lumina-DiMOO rows use different numbers of forward passes / turns than the fine-tuned samplers. The text should clarify that these are references, not controlled baselines, to avoid apples-to-oranges reading.","section":"Table 1"},{"comment":"The caption says 'coupling-driven unmasking' for events in high-Self/low-Cross space, but the surrounding text describes these as remask events; this looks like a typo for 'remasking'.","section":"Figure 7(b)"}],"recommendation":"major_revision","confidential_remarks":"The empirical study is extensive and the datasets are potentially valuable, but the paper's theoretical framing is not supported by the algorithm as written. The central chain-rule claim is not implemented, and the MJP formalism is not operationalized. I do not see this as a case for outright rejection: the heuristic sampler may still be a useful contribution, and the authors could either provide the missing rate-matrix/preservation analysis or reframe the paper as a training-free heuristic with a more honest description of the coupling. The revision needs to be substantial, hence major_revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on arXiv:2607.13188. The empirical story is more solid than the theoretical one, and the central claim about within-step coupling does not hold up to a close reading of Algorithm 1.\n\nWhat's genuinely new and good: the idea of using cross-modal attention-weighted confidence from a single forward pass to guide unmasking/remasking in a frozen MDM is a useful and practical contribution. The three datasets (JEdit-1M, JMaze-200K, JNono-200K) are a real resource, and evaluating on maze and nonogram against algorithmic ground truth is the right kind of test. The scaling curves with NFE are interesting, though I'd want error bars before trusting the monotonicity claim.\n\nThe main problem is that Algorithm 1 does not implement Eq. (7). The image scores are computed before the text death/birth updates, using cross-attention to H_text_t with a mask-aware bias that downweights masked positions. So the tokens that will be revealed at step s are masked at t and are effectively excluded from the CrossSignal. The image update conditions on z_text_t, not z_text_s. Given z_t, the text and image updates are conditionally independent, which is exactly the failure mode the paper attributes to MMaDA-Parallel. The paper calls this an approximation, but it is not a faithful surrogate — it discards the information that would make the coupling real. The ablation never isolates this surrogate, so the central mechanism is untested.\n\nThe 'Markov Jump Process' framing is also mostly decorative. The sampler is a confidence-ranking heuristic; no coupled rate matrix is derived or shown to preserve a target distribution. That is an overclaim rather than a fatal flaw, but it needs to be corrected.\n\nOn evaluation: the metrics rely on pseudo-grounding and proprietary oracle models, and no error bars or released code/data are provided. These are addressable, but they matter for confidence.\n\nOverall, the paper deserves a serious referee. The empirical results may hold up, but the narrative needs to match the algorithm. If you read it, focus on the gap between Eq. (7) and Algorithm 1; that is where the contribution lives or dies. I would send it to review, but with the expectation that the authors either implement true within-step coupling or substantially soften the claims.","headline":"Useful empirical work and new datasets, but the within-step coupling that is the paper's headline mechanism is not actually in the algorithm: the image branch conditions on the previous text state, not the state just sampled.","tokens_in":22723,"tokens_out":3584,"would_cite":true,"duration_ms":39816,"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":"A single training-free sampler can generate text and image concurrently, letting each modality correct the other in every denoising step, and it beats uncoupled parallel decoders on joint editing and visual-reasoning benchmarks.","keywords":["masked diffusion models","joint multimodal generation","Markov jump processes","remasking","self-correction","cross-modal attention","image editing","visual reasoning"],"falsifier":"Take CO2Jump and shuffle the cross-modal attention weights at every step, or replace the CrossSignal with random text confidences, while keeping all remasking counts and schedules identical; if joint accuracy on maze and nonogram solving and image-editing mAP stay about the same, the cross-modal signal is not the cause of the gains.","tokens_in":21840,"feed_emoji":"🖼️","tokens_out":5419,"duration_ms":47361,"temperature":0.7,"pith_summary":"Text and image generation are usually run sequentially or in parallel branches that only share history. This paper claims that a frozen masked diffusion model can instead generate both modalities concurrently, in a single forward pass per step, by coupling each modality's unmasking decisions to the other's confidence through cross-modal attention, and by remasking tokens the other modality contradicts. The resulting sampler, CO2Jump, is claimed to achieve the best joint performance on image editing plus understanding and on two visual-reasoning tasks (maze and nonogram solving), and to improve monotonically as more denoising steps are used. If true, concurrent generation with self-correction is available without retraining or an auxiliary evaluator, and the benefits of coupling compound across the trajectory.","feed_headline":"Coupled diffusion sampler tops joint text-image tasks in one pass","feed_subtitle":"A training-free sampler lets text and image correct each other at every denoising step; the gains grow with more steps.","key_machinery":"The load-bearing object is the Coupled Confidence score. After one forward pass, every position gets a Self-Confidence from Gumbel-max sampling of the model's belief. Text positions are scored by self-confidence alone, because the text update needs no cross-modal information, whereas image positions are scored by an entropy-gated blend of self-confidence and a CrossSignal: the text self-confidences weighted by image-to-text cross-attention from the same hidden states, with a mask-aware bias. An entropy gate λ = H_image/(H_image+H_text) decides how much the image defers to text. Scores are rank-normalized within each modality so vocabularies are comparable. These scores drive a death-birth ju","core_discovery":"The paper's central claim is that cross-modal coupling can be extracted from a frozen jointly-trained masked diffusion model at inference time, without extra forward passes. It rewrites the joint reverse step using the chain rule: text updates first from the joint state, and the image update should condition on the freshly sampled text. Since that fresh text is unavailable without a second forward pass, the image side instead reads the model's own text self-confidence, propagates it through cross-modal attention, and mixes it with image self-confidence under an entropy-derived gate. A death jump then remasks the lowest-confidence committed tokens and a birth jump reveals the highest-confiden","pith_inferences":["Because the framework is modality-agnostic, the same self-correcting coupling could be carried over to audio+transcript, code+execution trace, or video+narration, where one stream's late commitments could retract the other's errors; the authors note this as future work.","The attention-based surrogate for the missing text state is a testable hypothesis: if it is faithful, then joint NELBO training has implicitly stored cross-modal denoising knowledge in hidden states, suggesting training objectives could be designed to make that knowledge even more accessible at inference.","A finer-grained extension would be per-token or per-region entropy gates instead of the single scalar λ; the paper's gating only modulates the global balance between modalities, so local handling of uncertain versus confident regions is left open.","The monotonic scaling result hints that the coupling benefit may continue well beyond 512 steps or with adaptive schedules; this is an empirical extrapolation not established by the paper."],"forward_implications":["Joint text-image samples can be produced from a frozen model in one forward pass per step, with no auxiliary evaluator, so concurrent generation need not multiply inference cost.","Cross-modal contradictions that previous samplers could not repair — because committed tokens could not be remasked — become correctable during decoding.","The reported monotonic improvement with more denoising steps means users can trade compute for joint quality in a predictable way, without the regressions seen in uncoupled baselines.","On tasks where text and image are logically interlocked, such as maze and nonogram solving, the two modalities converge to the same algorithmic solution, not merely plausible but mismatched outputs.","The chain-rule decomposition provides a principled reason why sharper text grounding should tighten image-side decisions: text decisions are made first within each step, then image scoring is conditioned on them."],"fun_headline_variants":["Self-correcting jumps couple text and image in one pass","Text and image correct each other during joint diffusion","One-pass sampler lets text and image self-correct","Coupling text and image confidence corrects joint generation","Training-free sampler couples text and image per step"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The image branch's cross-modal correction is actually a surrogate — attention-weighted text beliefs from the current forward pass stand in for the text tokens that will only be sampled in this step; if that proxy does not track the real next-step text state, the coupling gives no true conditioning and the reported gains would not be due to cross-modal correction.","fun_headline_variants_meta":{"raw":{"variants":["Self-correcting jumps couple text and image in one pass","Text and image correct each other during joint diffusion","One-pass sampler lets text and image self-correct","Coupling text and image confidence corrects joint generation","Training-free sampler couples text and image per step"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1405,"prompt_tokens":887,"completion_tokens":518,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":456}},"tokens_in":631,"tokens_out":518,"duration_ms":24120,"temperature":1.0,"reasoning_tokens":456,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T05:54:50.002434+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take CO2Jump and shuffle the cross-modal attention weights at every step, or replace the CrossSignal with random text confidences, while keeping all remasking counts and schedules identical; if joint accuracy on maze and nonogram solving and image-editing mAP stay about the same, the cross-modal signal is not the cause of the gains.","supporting_citations":[],"review_version":1}