{"id":"24f603f3-212b-45a8-bdba-fc7e247738eb","arxiv_id":"2608.07006","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"For visual RAG with diffusion language models, retrieving more pages helps coverage but hurts accuracy when all pages are fed in; a first-step entropy filter restores the gain.","lead":"This paper shows that giving a diffusion language model more retrieved document images can hurt answer accuracy, even though the right page is more likely to be in the pile. It introduces a training-free filter that inspects the model's first-step uncertainty and admits only helpful images, improving visual question answering on five benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ECF's theoretical guarantee depends on Assumption F.3's unverified equal-residual condition; the zero-threshold admission rule is not certified on natural retrieval contexts.","rationale":"The reader's weakest_assumption correctly identifies Assumption F.3's matched-residual condition as the insecure point in the theoretical argument. My stress-test reaches the same conclusion: Proposition 3 is the only formal link between ECF's positive entropy gain and lower source-conflict risk, and it requires conditions that are not verified and are unlikely to hold exactly with a white-image blank. This concern is load-bearing for the paper's mechanism claim, but it is not fatal to the empirical claim, because the paper's accuracy improvements are presented as independent evidence and the appendix honestly labels the gain as an empirically validated surrogate outside the matched basin. The verdict therefore remains CONDITIONAL, which is what the reader already assigned; no change is needed. I also considered the lack of confidence intervals as a competing concern. That is real, but it attacks the strength of the empirical effect rather than the theoretical explanation, and the effect sizes in Table 3 are large enough that sampling noise is unlikely to overturn the qualitative direction. The Assumption F.3 concern is more central to the paper's distinctive contribution: the claim that a first-step entropy signal can certify safe evidence admission. A concrete controlled test with known source tokens can settle whether the residual equality and epsilon <= 1/2 conditions hold and whether the zero threshold is justified.","tokens_in":21330,"tokens_out":7270,"duration_ms":85629,"concrete_test":"Run a controlled study in the style of Appendix G, but with a diverse set of chart/document templates that mimic natural retrieval cases. For each template, with known correct and competing source tokens, compute for both the real context {c1, ci} and the matched-blank context {c1, blank(ci)}: the block entropy H(q,E), the restricted source-label entropy A_R h2(epsilon) by renormalizing over the known source tokens, and the residual C = H(q,E) - A_R h2(epsilon); also record epsilon. Then test (1) whether |C_{E_real} - C_{E_blank}| <= delta and epsilon <= 1/2 hold for a large fraction of templates, and (2) whether the sign of G_i predicts the sign of (epsilon_i - epsilon_blank) and whether final-answer accuracy is higher when G_i > 0 than when G_i <= 0.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central explanatory claim is that blank-controlled entropy gain G_i tracks source-conflict risk, so admitting candidates with G_i > 0 preserves retrieval coverage while limiting harmful visual exposure. This rests on Proposition 3, which is derived under Assumption F.3: real and matched-blank contexts must have identical source-token pairs, the same conflict set R, the same weights alpha_m, equal residual entropy C_{E_real} = C_{E_blank}, and competing-source mass at most 1/2. The equal-residual condition is asserted as an explicit local condition, but it is never measured on natural data. The blank is an all-white image; nothing rules out a systematic difference in how the DLM treats a white image versus a real page in the first-step answer block, including calibration differences or image-token attention effects. Table 10 compares admission rates and accuracy among blank types, but it does not test residual equality. The appendix itself concedes that outside the matched basin, the gain 'remains an empirically validated surrogate rather than a universal correctness certificate.' Consequently, the theoretical bridge from positive G_i to lower hybrid mass and lower sampling risk is not established for natural retrieval. The empirical accuracy gains may be real, but the mechanism story and the operational zero threshold are not certified by Proposition 3.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies visual retrieval-augmented generation with diffusion language models (DLMs). It reports that expanding the retrieved candidate pool increases answer-page recall but, when all retrieved pages are passed to the generator, often decreases answer accuracy across three DLMs and five visual QA benchmarks. The authors attribute this mismatch to source-coherence loss in parallel denoising, where a position-wise factorized proposal can assemble hybrid answers supported by neither the correct nor the conflicting page. Based on evidence that this interference is visible in the first-step answer-block distribution, they propose ECF, a training-free evidence-admission framework combining layout-based visual slicing with a blank-controlled entropy gain and rank-prior selection. ECF preserves the top-ranked candidate and admits at most one companion. Experiments report average accuracy gains of 2.62 percentage points over the strongest fixed top-k input and 2.37 percentage points over the best competing training-free method with LLaDA2.0-Uni.","tokens_in":21604,"tokens_out":6839,"duration_ms":70375,"significance":"The empirical phenomenon is useful for the visual RAG community: for diffusion-language-model RAG, retrieval breadth should be paired with selective evidence admission, and the first-step answer-block distribution is a plausible cheap signal for pre-decoding evidence assessment. Strengths include controlled second-image interventions across three backends, component ablations, released code, and a formal appendix stating assumptions and proofs. The main caveat is that the theoretical certificate for the entropy-gain admission rule depends on an unverified matched-context assumption, and the numerical claims rest on single-run accuracy without uncertainty estimates. The paper therefore contributes a valuable empirical hypothesis and heuristic, but the mechanism story and the operational zero threshold are not yet fully certified.","major_comments":[{"comment":"The operational interpretation of G_i depends crucially on Assumption F.3's equal-residual condition C_Ereal = C_Eblank = C together with epsilon in [0,1/2]. The appendix itself concedes, after Eq. (40), that outside this 'matched basin' the gain 'remains an empirically validated surrogate rather than a universal correctness certificate,' and no experiment in the paper tests residual equality on natural retrieval contexts. Since Section 4.2 uses the zero threshold G_i > 0 as the admission rule and the introduction motivates it by 'lower competing-source risk,' the theoretical bridge from positive gain to lower hybrid mass and lower sampling risk is not established for the natural contexts where ECF is deployed. Please either measure the residual C on natural data (for example, by estimating the restricted source entropy from renormalized top-two source tokens and comparing H(q,E) - A_R h_2(epsilon_E) for real and blank contexts) or explicitly downgrade Proposition 3's role to a heuristic motivation and soften the corresponding claims.","section":"Appendix F, Assumption F.3 and Proposition 3"},{"comment":"The controlled source-conflict validation is circular as support for the latent-source mechanism. Table 1 reports 'restricted hybrid source mass' obtained by renormalizing model probability over the correct and conflicting source tokens at conflict positions, which is precisely the restricted process posited in Assumptions F.1 and F.2. Consequently the parallel-vs-sequential difference in Table 1 follows substantially from the construction and cannot independently confirm that natural multi-page retrieval produces this two-source structure. The natural-retrieval evidence in Figure 2 and the direct Recall-vs-accuracy comparison do support the empirical mismatch, but the causal story attributed to source-coherence loss should be presented as an interpretation, with the controlled study as an illustration rather than as a test of the theory.","section":"Section 3.4, Table 1 and Appendix G"},{"comment":"All answer-accuracy tables report a single evaluation run over one fixed 2,500-example subset, with no confidence intervals, bootstrap errors, or multiple decoding runs. The headline gains (2.62 percentage points over the strongest fixed top-k input and 2.37 percentage points over the best training-free alternative) are therefore unaccompanied by any measure of variability, and it is unclear whether the ECF-vs-baseline differences are stable. Given that Table 3 shows nonmonotonic behavior (Dream-VL top-2 outperforms top-1 while LLaDA-V degrades sharply), the aggregate claims need per-configuration uncertainty estimates and, ideally, a paired significance test across the 15 model-dataset settings.","section":"Tables 2, 3, 5 and 12"}],"minor_comments":[{"comment":"The caption's statement that 'accuracy decreases as more pages are provided' is only true in macro-average; Table 3 shows Dream-VL top-2 accuracy is higher than top-1 (32.41 vs. 30.27). Please show per-backend fixed-k curves or qualify the caption to acknowledge backend-dependent nonmonotonicity.","section":"Figure 1 and Table 3"},{"comment":"The decay rate lambda is fixed at 0.5 with no sensitivity analysis. Since H(q,E), and therefore G_i and the admission decision, depend on lambda, a small sensitivity table over lambda values would help establish that the reported gains are not an artifact of this choice.","section":"Section 4.1, Eq. (5)"},{"comment":"The evaluation subsets are described as randomly sampled 2,500 question-answer pairs. Please state the sampling seed and confirm that all methods and all backends are evaluated on the identical subsets, since any subset mismatch would weaken the paired comparisons in Tables 2, 3, and 12.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"I do not see grounds for rejection. The direct empirical finding is useful, and the method is simple and clearly described. The main fixes are uncertainty quantification for the headline numbers and either validating Assumption F.3's residual equality on natural contexts or explicitly labeling the Proposition 3 alignment as a heuristic. The appendix's own limitation statement should be moved into the main text if the theoretical claim is retained."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuinely useful empirical paper. The availability-accuracy mismatch is real, measured across three DLMs and five benchmarks, with controlled interventions showing that conflict, not mere image count, drives the degradation. The ECF admission rule, evaluated in the actual decoding context via a blank-controlled entropy gain, is a sensible, training-free, cheap procedure, and it beats fixed top-k and several stronger-looking baselines. The paper also publishes code.\n\nThe new things: the demonstration that parallel denoising loses source coherence, the early observability in the first-step answer block, and the decoupling of candidate-pool size from admitted context size. The Appendix F analysis is honest about being a restricted source process; Theorem 1 is real for that process, and the controlled validation matches it (hybrid mass grows with conflict width, parallel exceeds sequential).\n\nSoft spots, in proportion. The biggest is that Proposition 3's guarantee relies on Assumption F.3's equal-residual-entropy condition between real and blank contexts, which is never measured on natural data. The paper itself concedes this: outside the matched basin, Gi is an empirically validated surrogate rather than a certified correctness certificate. So the zero threshold for admission is an empirical choice, not a proven one. That is fine for a practical method, but the abstract and early framing lean on the theoretical alignment more than the appendix ultimately supports. Second, the main tables report single-run accuracy with no confidence intervals; given the headline gains are around 2.6 percentage points, we need run-to-run variance. The appendix bootstraps the controlled study but not the main results. Third, the blank control is an all-white image; Table 10 tests it against patch shuffle and unrelated real pages, which is good, but that still does not establish that the white blank preserves the residual entropy term.\n\nThe empirical claim — more evidence hurts without selective admission, and conflict is the mechanism — is solid and independent of the theory. Even if the theoretical certificate is weaker than presented, ECF remains an empirically validated heuristic with consistent gains, and the paper is honest enough to include the limitation.\n\nWho this is for: people building DLM-RAG pipelines and anyone working on evidence selection for non-autoregressive generators. It deserves serious review; the referee should push for confidence intervals and for a direct test of residual equality, but the core finding is worth publishing.\n\nRecommendation: engage with it, send to review.","headline":"Solid empirical study showing retrieval breadth hurts DLM-RAG accuracy without selective admission, but the theoretical certificate for ECF only holds under an unverified matched-residual assumption.","tokens_in":22110,"tokens_out":1683,"would_cite":true,"duration_ms":17943,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"In visual RAG with diffusion language models, retrieving more pages improves answer-page recall but lowers answer accuracy when all evidence is admitted, because parallel denoising can splice incompatible sources; a training-free entropy…","keywords":["diffusion language models","visual retrieval-augmented generation","evidence admission","semantic conflict","parallel denoising","answer-block entropy","training-free selection","visual question answering"],"falsifier":"Take a benchmark where the top-ranked page is wrong and a conflicting page is correct, run ECF's blank-controlled gain on each candidate, and check accuracy when the rule admits the lower-ranked page; Proposition 3 assumes the correct source is weakly dominant in the restricted first-step distribution, so if positive gain systematically selects the wrong page and lowers final accuracy, the central confidence–risk alignment is refuted.","tokens_in":21103,"feed_emoji":"🎯","tokens_out":8652,"duration_ms":77607,"temperature":0.7,"pith_summary":"The paper asks whether giving a visual question-answering diffusion language model more retrieved pages helps. It finds that more pages improve the chance that an answer-bearing page is available, but feeding all pages to the model often makes answers less accurate. The cause is semantic conflict: when two pages support different answers, the model's parallel position-by-position decoding can combine tokens from both pages into an answer neither page supports. Because this interference shows up in the first decoding step, the model's own uncertainty can be used to screen evidence before any answer is committed. The proposed training-free filter, ECF, keeps broad retrieval but admits only evidence that reduces first-step answer-block entropy, improving average accuracy by 2.62 percentage points over fixed top-k input.","feed_headline":"More pages, worse answers: diffusion RAG needs selective evidence","feed_subtitle":"Blank-controlled entropy filter admits only evidence that stabilizes the first decoding step, gaining 2.62 accuracy points.","key_machinery":"The load-bearing mechanism is source-coherence loss in parallel denoising, formalized through a restricted two-source process in which a shared latent variable chooses the correct or competing source while the exact-marginal factorized proposal predicts each position independently; the result is hybrid mass $1-(1-\\epsilon)^r-\\epsilon^r$ for conflict width $r$, zero at $r=1$ and strictly increasing with $r$. The admission rule rests on the blank-controlled entropy gain $G_i = H(q,\\{c_1,\\mathrm{blank}(c_i)\\}) - H(q,\\{c_1,c_i\\})$, a geometry-matched counterfactual that isolates the candidate's semantic contribution to the weighted first-step answer-block entropy; under matched local conditions, positive gain is equivalent to lower competing-source mass, lower hybrid mass, and lower parallel exact-match risk. Multi-granularity evidence units (full pages plus layout-derived regions) reduce within-candidate noise, and retrieval rank supplies an asymmetric prior: the second-ranked candidate is the default companion, while lower-ranked candidates must improve both the gain and the absolute joint confidence to be admitted.","core_discovery":"The central discovery is an availability–accuracy mismatch: in visual RAG with diffusion language models, expanding the retrieved candidate pool raises answer-page recall while unconditional admission of all retrieved pages lowers answer accuracy, with semantic conflict between pages as the primary driver. Under parallel denoising, a factorized position-wise proposal loses the shared latent source choice; when two pages support different coherent answers across $r$ positions, the proposal assigns probability $1-(1-\\epsilon)^r-\\epsilon^r$ to hybrid answers supported by neither page, and this unsupported mass grows with conflict width. The mismatch is observable before decoding in the first-step answer-block distribution, which makes evidence assessment possible without a full denoising run. The paper then constructs ECF, which builds multi-granularity evidence units from pages and layout-derived regions, compares each candidate against a geometry-matched blank in the target DLM's first-step block entropy, and uses retrieval rank as a prior to admit at most one companion to the top-ranked evidence. Across three multimodal DLMs and five visual QA benchmarks, ECF improves answer accuracy by 2.62 percentage points on average over the strongest fixed top-$k$ input and stays stable as candidate pools grow.","pith_inferences":["The same source-coherence argument should apply to text-only diffusion RAG, where conflicting passages play the role of conflicting pages; the mechanism in Theorem 1 does not depend on the visual modality, only on parallel decoding over a shared unresolved state.","The rank-prior policy always keeps the top-ranked page, so a dataset where the top-ranked page is the conflicting one and the correct page is second would stress the rule; a natural extension would allow rejecting or replacing $c_1$ on the same entropy evidence.","The choice of blank control is a design decision with measurable consequences: the appendix shows a white blank outperforms patch shuffling and unrelated real pages on ChartQA, suggesting that different controls may suit different document types.","The appendix's approximate guarantee suggests a practical threshold test: if the residual-entropy mismatch between real and blank contexts is large, ECF's zero threshold should be replaced by a positive margin before admitting a candidate; one could estimate that mismatch per dataset."],"forward_implications":["Retrieval breadth and evidence admission should be treated as separate decisions: candidate-pool size can grow as a retrieval budget without degrading decoding, because final decoding is capped at the admitted evidence.","Evidence admission can be decided before decoding: the first-step answer-block distribution exposes conflict, so a single forward pass suffices to gate candidates and the probe cost grows only linearly with pool size.","A generator-native signal can outperform semantic or surrogate-based selection: ECF beats training-free alternatives that rerank by frozen encoders, answer uncertainty, or utility surrogates, and the margin holds across three diffusion backends.","Selective admission is stable as pools expand: ECF accuracy stays roughly flat from $k=3$ to $k=5$ even as answer-page recall keeps rising, while fixed top-$k$ accuracy declines."],"supporting_citations":[{"why":"Supplies the visual retrieval and ranking used to form the candidate pools and the fixed top-k baselines.","marker":"Yu et al. 2025"},{"why":"Provides the strongest training-free evidence-selection alternative (UOVES) that ECF must beat on each dataset.","marker":"Luo et al. 2026"},{"why":"Defines LLaDA2.0-Uni, the primary multimodal DLM backend on which ECF's main gains are reported.","marker":"Inclusion AI et al. 2026"},{"why":"Provides LLaDA-V, one of the three diffusion backends used for cross-model validation.","marker":"You et al. 2025"},{"why":"Provides Dream-VL, the third diffusion backend used for cross-model validation.","marker":"Ye et al. 2026"},{"why":"Establishes that irrelevant content can arise both within and across retrieved pages, motivating multi-granularity evidence units.","marker":"Li et al. 2025"},{"why":"Supplies the masked-diffusion formulation underlying the position-wise proposal analyzed in the source-conflict argument.","marker":"Sahoo et al. 2024"}],"fun_headline_variants":["Diffusion visual RAG: selective evidence beats more pages","More retrieval pages reduce accuracy in diffusion RAG","Evidence entropy filter boosts diffusion RAG accuracy","Why expanding evidence hurts diffusion visual RAG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a blanked-out candidate differs from the real candidate only in meaning, and that the correct page is at least as probable as the competing page in the model's first-step distribution; if real retrieved pages violate that match, the entropy gain is no longer certified to track conflict risk.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion visual RAG: selective evidence beats more pages","More retrieval pages reduce accuracy in diffusion RAG","Evidence entropy filter boosts diffusion RAG accuracy","Why expanding evidence hurts diffusion visual RAG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1466,"prompt_tokens":1078,"completion_tokens":388,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":694,"completion_tokens_details":{"reasoning_tokens":329}},"tokens_in":694,"tokens_out":388,"duration_ms":4364,"temperature":1.0,"reasoning_tokens":329,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T16:45:18.356593+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a benchmark where the top-ranked page is wrong and a conflicting page is correct, run ECF's blank-controlled gain on each candidate, and check accuracy when the rule admits the lower-ranked page; Proposition 3 assumes the correct source is weakly dominant in the restricted first-step distribution, so if positive gain systematically selects the wrong page and lowers final accuracy, the central confidence–risk alignment is refuted.","supporting_citations":[],"review_version":1}