{"id":"b68ff687-2155-433b-a5c5-a1e78f440014","arxiv_id":"2607.17884","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ST-Veto improves reasoning in diffusion MLLMs by vetoing temporally unstable tokens and tokens with weak image grounding, swapping in safer near-boundary candidates.","lead":"A decoding-time method for diffusion-based vision-language models that vetoes tokens which look confident but are temporally unstable or poorly grounded in the image, swapping in safer candidates. It reports consistent accuracy gains across four multimodal reasoning benchmarks with no retraining and negligible speed cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Missing negative control: Veto-and-Swap gains could come from the swap action alone, not from the Taylor/grounding signals; the mechanistic claim is underdetermined.","rationale":"The paper is well-designed in many respects: deterministic decoding, rule-based answer extraction, flip analysis, ablations, stagewise analysis, and POS analysis all support the claim that ST-Veto changes token selection and often improves accuracy. Those are real pieces of evidence, and they are why I am not proposing rejection. However, they do not isolate whether the specific Taylor-residual and image-attention signals are responsible for the improvement. Because the intervene step is discrete and small (only a few percent of tokens are swapped), a plausible alternative explanation is that any policy that swaps a small number of top-k function-word tokens for near-boundary content words would produce a similar gain. The paper's own POS data (Table 13) make this alternative concrete: vetoed tokens are dominated by determiners and prepositions, while swapped tokens are more often nouns and verbs. If a randomized veto signal achieved the same accuracy, the central mechanistic claim—that spatio-temporal stability and visual grounding drive the gains—would be false, even if the empirical accuracy numbers were reproduced. The proposed permutation test is the minimal experiment that can separate 'the act of swapping' from 'the information content of the swap signal.' It is cheap, training-free, and directly answers the reader's weakest-assumption concern. I therefore keep the reader's CONDITIONAL verdict: the concern identifies missing evidence rather than a demonstrated break of the central claim, so no verdict change is needed.","tokens_in":19160,"tokens_out":14172,"duration_ms":154076,"concrete_test":"Permutation control: for each benchmark and L/T setting, run ST-Veto but at each denoising step shuffle the Taylor residuals and image-attention scores randomly among currently masked positions, keeping the same marginal distributions and therefore the same veto/safe counts and number of swaps. Compare accuracy against Base, real ST-Veto, and a content-word-prior control (e.g., prefer nouns/verbs near the candidate boundary). If permuted-signal accuracy is statistically indistinguishable from real ST-Veto, the proxy signals are not the cause; if it drops to Base, the signals carry causal information. Report per-seed results with standard deviations for all four benchmarks, and resolve the Table 1 vs Table 3 MMBench discrepancy for MMaDA.","verdict_should_be":"UNCHANGED","load_bearing_attack":"ST-Veto's causal claim is that tokens are improved because Taylor residuals and image-attention masses carry information about correctness and grounding (Eqs. 5–14). The paper tests this only by ablating each signal off (Table 3), never by randomizing the signal while keeping the veto rate and swap policy fixed. This matters because the policy is a small perturbation: only ~2–7% of top-k tokens are swapped (Table 7), and the POS analysis shows vetoed tokens are mostly function words while swapped tokens are content words (Table 13). A control that permutes e_i and m_i across positions—preserving margins, thresholds, and swap counts—would reveal whether the specific information in the proxies is doing work or whether any near-boundary content-word swap would give most of the gain. Without that control, 'unstable/weakly grounded' could be a post-hoc description of tokens that are merely low-attention or momentum outliers. The empirical consistency (16/16 positive cells) is encouraging, but it is compatible with a simpler 'swap in content words early' explanation. Additionally, the absence of code/error bars and a cross-table inconsistency (MMaDA MMBench 128/64 is 36.55 in Table 1 but 35.55 in Table 3) make the reported magnitudes hard to verify.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ST-Veto, a training-free decoding-time method for diffusion-based multimodal LLMs (dMLLMs). At each unmasking step, ST-Veto identifies top-k confident tokens and vetoes those whose confidence trajectory deviates from a second-order Taylor prediction (temporal instability) or whose image-attention mass falls below a MAD-based threshold (weak visual grounding). Vetoed tokens are replaced by safe near-boundary candidates that satisfy both temporal and grounding criteria. Experiments on M3CoT, ScienceQA, MMBench, and V*Bench with LaViDa and MMaDA claim consistent improvements over standard decoding baselines and adapted AR-style reasoning methods, with gains up to about 8.5 percentage points and negligible throughput overhead. Ablations, flip analysis, stagewise swap distributions, and POS analysis are presented to support the mechanism.","tokens_in":19459,"tokens_out":3561,"duration_ms":35936,"significance":"If the empirical claims hold, ST-Veto would be a valuable, practical contribution: a training-free, near-zero-cost decoding policy that improves reasoning in a class of models (dMLLMs) for which AR-style prompting methods do not transfer. The paper provides several internal analyses—ablation of each signal, stagewise swap distribution, instance-level flip asymmetry, POS distribution of vetoed/swapped tokens—that go beyond a single accuracy number. These analyses, together with the clear formulation of the veto-and-swap policy, make the mechanism plausible and the paper worth serious consideration. However, the central empirical claim currently rests on a few percentage-point gains reported without variance, and the mechanistic interpretation is not fully identified against a simpler 'swap-in-content-words' alternative.","major_comments":[{"comment":"Table 1 reports averages over three seeds but no standard deviations, confidence intervals, or significance tests. Many improvements are small in absolute terms (e.g., MMaDA M3CoT 128/64: 26.32→27.52; MMaDA MMBench 256/128: 32.04→36.74 is larger but still needs variance). Without error bars, the reader cannot assess whether the single-digit gains are robust across seeds. Please report per-seed results or variance estimates, and ideally paired significance tests.","section":"§5.1, Table 1"},{"comment":"There is a load-bearing inconsistency: MMaDA MMBench 128/64 is 36.55 in Table 1 but 35.55 in Table 3, and 256/128 is 36.74 vs 35.74. Since Table 3 is the ablation of the main result, the numbers must agree. Clarify which table is correct and correct the typo; this discrepancy calls into question the reliability of the numerical reporting.","section":"§5.3 (Table 3) vs Table 1"},{"comment":"The causal claim is that Taylor residuals and image-attention masses carry information about correctness and grounding, and that this information drives the gains. The paper ablates each signal off, but both ablations still use the veto-and-swap policy with some signal. Given that only 2–7% of top-k tokens are swapped (Table 7) and swapped tokens are predominantly content words (Table 13), the gains could in principle come almost entirely from the swap action itself—replacing function words with near-boundary content words—rather than from the specific information in e_i and m_i. A permutation control is needed: shuffle e_i and m_i across positions while preserving the veto rate, swap counts, and candidate-line thresholds. If permuted signals retain most of the gain, the mechanistic interpretation in Eqs. (5)–(14) is underdetermined.","section":"§3.3, §5.3, §6 (missing negative control)"},{"comment":"The method's premise is that internal confidence and attention are reliable proxies for token correctness and visual grounding. The paper acknowledges in A.9 that these are 'imperfect proxies,' but it never tests the premise directly. For instance, one could measure whether vetoed tokens are indeed less likely to appear in the final correct sequence, or whether high-confidence tokens correlate with answer correctness in these dMLLMs. Without such a check, the veto criteria (Eq. 11–12) are justified only by the end-task accuracy, which is circular if the proxy assumption itself is part of the claimed mechanism.","section":"§A.9, Eq. (11)–(12)"}],"minor_comments":[{"comment":"MAD is used throughout (Eq. 11, 12, 14) but never defined. Define median absolute deviation and specify whether the scaling factor 1.4826 is used.","section":"§3.3"},{"comment":"Panels (c) and (d) contain unreadable glyph sequences (e.g., '/uni00000018/...'), likely a font-encoding artifact. The figure should be regenerated with proper text rendering.","section":"Figure 3"},{"comment":"The 'Attribute recognition' row has missing spacing ('26435.61') and should read '264 35.61'.","section":"Table 14"},{"comment":"Sahoo et al. (2024a) and (2024b) are the same paper and should be merged or distinguished by venue details; currently listed twice with identical titles.","section":"References"},{"comment":"The text says H=4 'provides the best cost-performance trade-off,' but Table 4 shows the third-order variant (H=5) matching or exceeding H=4 in several cells (e.g., MMBench 128/64: 38.11 vs 38.05). The text acknowledges this, but the wording 'best' is slightly strong; consider 'best trade-off' or 'strongest overall.'","section":"§6.2, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an interesting and timely problem, and the method is coherent. My main concern is that the empirical support is not yet at the level that would justify the strong claims of 'consistent outperformance' and 'up to 9%' gains: the absence of variance reporting, the cross-table inconsistency, and the missing randomization control all need to be addressed. I would encourage the authors to add the permutation experiment and full variance reporting; if those come out favorably, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"ST-Veto is a training-free decoding-time intervention for diffusion MLLMs that vetoes and swaps tokens using a second-order Taylor residual on confidence trajectories plus image-attention mass. The combination is new in this setting, and the paper gives a coherent, well-specified account of why it should help: dMLLMs unmask in parallel, so a token that looks confident under a noisy context can become a bad anchor. The ablations, flip analysis, stagewise swap distribution, and POS analysis all hang together and support the intended mechanism. The efficiency claim also holds up: overhead relative to base decoding is negligible, unlike the AR-style baselines that double the number of generation passes.\n\nWhat keeps me from endorsing the headline numbers is reporting, not the idea. Table 1 shows three-seed means but no error bars or per-seed spread, and no code or data are released. Several hyperparameters (attention layers, MAD multiplier, Taylor history) are tuned on the evaluation benchmarks, and the \"up to 9%\" in the abstract is hard to map to the absolute gains in Table 1. There is also a numeric inconsistency: MMaDA MMBench 128/64 is 36.55 in Table 1 but 35.55 in Table 3. Any of this is addressable, but right now the central claim is not independently checkable.\n\nThe missing negative control bothers me more than the missing error bars. The paper claims the Taylor and grounding signals carry information about token correctness. It ablates each off, but never randomizes the signals while preserving veto rate and swap policy. Since only 2–7% of top-k tokens are swapped and swapped tokens are mostly content words, a plausible alternative is that any policy that swaps in stable, visually grounded content words early gives most of the gain. The current analyses don't rule that out. That said, the paper's own limitation section is honest about internal signals being imperfect proxies, and the method is conservative—it only swaps when a safe near-boundary candidate exists. So I don't think the central idea is wrong; I think the mechanistic claim is stronger than the evidence.\n\nThis is a useful paper for people working on diffusion LMs and decoding-time control. It deserves a serious referee round, but it should come back with code, variance estimates, and ideally the permutation control. I would not cite it in my own work until the numbers are reproducible.","headline":"Training-free veto-and-swap for diffusion MLLMs is a plausible new idea, but the main numbers need error bars, code, and a stronger control before I'd trust the mechanistic claim.","tokens_in":61,"tokens_out":1812,"would_cite":false,"duration_ms":25114,"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 training-free decoding policy that vetoes tokens with unstable confidence trajectories or weak visual grounding consistently improves reasoning accuracy in diffusion-based multimodal LLMs, by up to 9 percentage points, without extra train","keywords":["diffusion multimodal LLMs","decoding policy","token veto","Taylor prediction","visual grounding","image attention","training-free","chain-of-thought reasoning"],"falsifier":"Run ST-Veto on a dMLLM where calibration is known to be poor—for example, one whose high-confidence tokens are no more likely correct than low-confidence ones—and compare accuracy against base decoding. If accuracy does not improve or degrades, the confidence-based veto signal is not doing the work. A more direct test: use a benchmark where the model is systematically overconfident about hallucinated objects and check whether the flip ratio becomes symmetric or negative.","tokens_in":19020,"feed_emoji":"🚫","tokens_out":3490,"duration_ms":32577,"temperature":0.7,"pith_summary":"ST-Veto claims that the reasoning failures of diffusion-based multimodal large language models (dMLLMs) can be substantially reduced by a training-free decoding policy: before a token is committed as context for the next denoising step, check whether its confidence trajectory is smooth and whether it is visually grounded. The paper proposes two veto signals—a second-order Taylor residual that flags tokens whose confidence deviates from predicted dynamics, and an image-attention mass that flags weakly grounded tokens—and swaps vetoed tokens with safe near-boundary candidates. Across two dMLLMs and four benchmarks, the method improves accuracy by up to 9 percentage points at no extra generation cost. The result matters because it suggests that better reasoning in diffusion models can come from decoding-time control rather than retraining, and that autoregressive chain-of-thought methods do not transfer to the parallel denoising paradigm.","feed_headline":"Veto policy lifts diffusion MLLM accuracy by up to 9%","feed_subtitle":"Training-free decoding swaps unstable or weakly grounded tokens for safer ones.","key_machinery":"The load-bearing object is the Veto-and-Swap policy. It defines a veto criterion from two signals: the Taylor residual, the deviation of a token's current confidence from a second-order Taylor prediction computed from a four-step history of that same masked position, and the image attention mass, the token's attention toward image tokens averaged over layers L/3, L/2, and 2L/3. A token is vetoed if either signal falls below median minus one MAD (median absolute deviation); a swap candidate is safe only if both signals exceed median plus one MAD and its confidence lies within one MAD of the top-k boundary. The policy then replaces vetoed top-k tokens with the highest-confidence safe candidate","core_discovery":"ST-Veto establishes that in diffusion-based multimodal language models, the standard practice of unmasking the currently most confident tokens at each denoising step is suboptimal: tokens can be confidently wrong because they are locally obvious under a noisy context, yet unstable over time or poorly anchored to the image. The paper shows that a veto-and-swap policy, which rejects tokens with large negative Taylor residuals of confidence and tokens with low image-attention mass, and replaces them with safe candidates that are both temporally stable and visually grounded, consistently improves reasoning accuracy across models and benchmarks. The gains are asymmetric in paired analysis (many m","pith_inferences":["The same veto-and-swap principle could be tested on text-only discrete diffusion LLMs, where only the Taylor residual signal is available; a temporal-only veto might already improve coherence.","The second-order Taylor prediction can be viewed as a null model for smooth confidence evolution; its residual measures local volatility that could be linked to token-level uncertainty in future work.","ST-Veto's conservative design—swapping only when a safe near-boundary candidate exists—bounds its downside, making it a plausible default decoding strategy for production dMLLMs.","Combining ST-Veto with training-time objectives that explicitly smooth confidence trajectories or sharpen image grounding could yield further gains beyond decoding-time control."],"forward_implications":["If correct, training-free decoding policies can improve dMLLM reasoning across benchmarks without retraining or extra generation steps.","Autoregressive chain-of-thought prompting methods are structurally mismatched with diffusion decoding and should not be expected to transfer; diffusion-native control is needed.","The veto signals track token-level competence: vetoed tokens are disproportionately function words, swapped tokens content words, supporting an early-anchor interpretation.","Gains persist at longer generation lengths and extend to additional benchmarks such as MME, POPE, and MMMU.","Temporal stability and visual grounding are complementary signals; combining them outperforms either alone in ablations."],"fun_headline_variants":["ST-Veto lifts diffusion MLLM accuracy by up to 9%","Veto unstable tokens to boost diffusion MLLM reasoning","Training-free token veto improves MLLM accuracy 9%","Swap unstable, weakly grounded tokens for safer MLLM decoding","Diffusion MLLM gains 9% with spatio-temporal token veto"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole veto signal rests on the premise that a token's model confidence and its attention to image tokens are trustworthy indicators of whether the token is correct and visually grounded; the paper itself concedes that both are imperfect proxies that can be misled by non-smooth confidence trajectories and diffuse or semantically misaligned image attention.","fun_headline_variants_meta":{"raw":{"variants":["ST-Veto lifts diffusion MLLM accuracy by up to 9%","Veto unstable tokens to boost diffusion MLLM reasoning","Training-free token veto improves MLLM accuracy 9%","Swap unstable, weakly grounded tokens for safer MLLM decoding","Diffusion MLLM gains 9% with spatio-temporal token veto"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1175,"prompt_tokens":704,"completion_tokens":471,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":380}},"tokens_in":448,"tokens_out":471,"duration_ms":4866,"temperature":1.0,"reasoning_tokens":380,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:43:00.526346+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ST-Veto on a dMLLM where calibration is known to be poor—for example, one whose high-confidence tokens are no more likely correct than low-confidence ones—and compare accuracy against base decoding. If accuracy does not improve or degrades, the confidence-based veto signal is not doing the work. A more direct test: use a benchmark where the model is systematically overconfident about hallucinated objects and check whether the flip ratio becomes symmetric or negative.","supporting_citations":[],"review_version":1}