Does More Retrieved Evidence Help Visual Retrieval-Augmented Generation with Diffusion Language Models?

# Does More Retrieved Evidence Help Visual Retrieval-Augmented Generation with Diffusion Language Models?
Source: [https://arxiv.org/html/2608.07006](https://arxiv.org/html/2608.07006)
Anonymous Submission

###### Abstract

Visual retrieval\-augmented generation \(RAG\) commonly expands the retrieved evidence set to improve answer\-page coverage, implicitly assuming that all available evidence should be passed to the generator\. We show that this assumption does not hold for diffusion language models \(DLMs\): retrieving more pages increases answer\-page recall, whereas unconditionally passing all retrieved pages to the generator often reduces answer accuracy, primarily because of semantic conflict\. A latent\-source analysis explains this mismatch through source\-coherence loss in parallel denoising, where position\-wise proposals can combine incompatible visual sources into unsupported answers\. We further find that such interference is already visible in the first\-step answer\-block distribution, making it possible to assess evidence before decoding\. To preserve retrieval coverage while limiting harmful visual exposure, we propose the Entropy\-Based Candidate Filter \(ECF\), a training\-free evidence\-admission framework\. To reduce irrelevant content within individual candidates, ECF constructs multi\-granularity evidence units; to identify beneficial additional evidence, it uses blank\-controlled block confidence and retrieval rank to determine whether and which candidate should enter the final context\. 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\-kkinput and, with LLaDA2\.0\-Uni, by 2\.37 percentage points on average over the best competing training\-free result for each dataset\. These results show that broader retrieval benefits visual DLM\-RAG through selective evidence admission rather than unconditional evidence expansion\.

## 1Introduction

Most modern large language models generate text autoregressively and have achieved strong performance across question answering, reasoning and tool\-use tasks\(Xiaoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib31); Maet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib32)\)\. More recently, diffusion language models \(DLMs\) have attracted increasing attention for their bidirectional conditioning and parallel decoding capabilities\(Austinet al\.[2021](https://arxiv.org/html/2608.07006#bib.bib10); Hoogeboomet al\.[2021](https://arxiv.org/html/2608.07006#bib.bib11); Sahooet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib12); Nieet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib13); Gaoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib33)\)\. Recent studies\(Inclusion AIet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib25); Youet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib26); Yeet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib27)\)have further extended DLMs to multimodal generation, enabling visual inputs to guide the denoising process\.

However, many knowledge\-intensive visual questions cannot be answered reliably from model parameters alone, creating a need for external evidence\. Retrieval\-augmented generation \(RAG\) addresses this need by conditioning generation on retrieved content\(Izacard and Grave[2021](https://arxiv.org/html/2608.07006#bib.bib1); Shiet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib2); Asaiet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib3)\), and has been extended to visual tasks\(Yuet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib8); Liet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib9)\)\. Visual RAG is challenging because answer\-bearing evidence is often localized within a visually dense page, while each retrieved page introduces a complete visual source containing additional structured content\. To compensate for imperfect retrieval, existing visual RAG systems, which are predominantly built around autoregressive LLMs, often provide the generator with multiple top\-ranked pages to improve answer\-page coverage\(Yuet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib8); Luoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib16)\)\. Recent work has improved the quality of this context through finer\-grained region retrieval and generation\-aware evidence selection\(Liet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib9); Luoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib16)\)\. Yet their implications for multimodal DLMs remain unclear: when several retrieved images are provided, they jointly condition all unresolved answer positions during parallel denoising\. Whether an individually relevant or useful candidate remains beneficial after entering this shared masked state is not well understood\. This motivates our central question:*does providing more retrieved evidence actually improve visual RAG with diffusion language models?*

Our experiments show that more retrieved evidence does not necessarily help\. To examine how fixed input size affects retrieval coverage and generation quality, we evaluate fixed top\-kkinput withk∈\{1,2,3\}k\\in\\\{1,2,3\\\}using the same VisRAG\-ranked candidate pools, prompts, decoding budgets, and answer scorer across three multimodal DLMs and five visual QA benchmarks\. Figure[1](https://arxiv.org/html/2608.07006#S1.F1)reports macro\-average answer\-page Recall@kkacross the five benchmarks and macro\-average answer accuracy across all 15 model–dataset settings: Recall increases withkk, whereas accuracy decreases as more pages are provided\. Figure[2](https://arxiv.org/html/2608.07006#S1.F2)\(a\) further shows that conflicting pages cause the largest accuracy drop across all three backends, indicating that semantic conflict, rather than the mere presence of multiple images, drives the degradation\. Thus, expanding the candidate pool improves answer\-page availability, but indiscriminately admitting additional candidates can make generation less reliable\.

To explain this mismatch, Section[3\.3](https://arxiv.org/html/2608.07006#S3.SS3)develops a restricted latent\-source analysis of the first unresolved denoising state\. When correct and conflicting pages support different coherent answers, multiple unresolved positions share a latent source choice, whereas a position\-wise factorized proposal does not preserve this coupling\. The analysis shows that this dependence loss assigns probability to cross\-source answers supported by neither page, with the unsupported mass increasing as conflict spans more answer positions\. This identifies source\-coherence loss as a mechanism behind the observed availability–accuracy mismatch\. Crucially, this failure is already visible before final decoding\. On naturally retrieved ChartQA cases, conflicting pages cause the largest drop in first\-step answer\-block accuracy and the largest Jensen–Shannon divergence from the correct\-only distribution in Figure[2](https://arxiv.org/html/2608.07006#S1.F2)\(b\); the same condition also produces the largest final\-answer degradation in Figure[2](https://arxiv.org/html/2608.07006#S1.F2)\(a\)\. This early\-to\-final correspondence motivates the first\-step block distribution as a generator\-native signal for candidate assessment\.

![Refer to caption](https://arxiv.org/html/2608.07006v1/x1.png)Figure 1:Answer\-page Recall@kkand answer accuracy under fixed top\-kkinput\.We therefore propose theEntropy\-BasedCandidateFilter \(ECF\), a training\-free framework that controls visual evidence exposure along two complementary axes\. For dense documents, multi\-granularity evidence construction combines full pages with layout\-derived regions to make localized evidence retrievable\. For generator\-side admission, ECF evaluates each additional candidate in the context in which it would actually be used\. Given the top\-ranked evidencec1c\_\{1\}and a candidatecic\_\{i\}, it compares the target DLM’s first\-step block entropy under\{c1,ci\}\\\{c\_\{1\},c\_\{i\}\\\}and\{c1,blank​\(ci\)\}\\\{c\_\{1\},\\mathrm\{blank\}\(c\_\{i\}\)\\\}, yielding a conditional confidence gain that controls for the structural effect of an additional visual input\. Under the local conditions in Proposition[3](https://arxiv.org/html/2608.07006#Thmtheorem3), a positive gain corresponds to lower competing\-source risk\. Rank\-prior selection then uses these gains to decide whether to expand beyondc1c\_\{1\}and, if so, which candidate to admit\. Across three multimodal DLMs and five visual QA benchmarks, ECF improves over the strongest fixed top\-kkinput by 2\.62 percentage points on average\. With LLaDA2\.0\-Uni, it further outperforms the strongest training\-free alternative by 2\.37 percentage points on average\.

![Refer to caption](https://arxiv.org/html/2608.07006v1/x2.png)Figure 2:Effects of additional visual evidence on ChartQA\. \(a\) Final\-answer accuracy drops relative to correct\-only input under controlled second\-image interventions across three backends; correct\-only accuracies are shown below the backend names\. \(b\) First\-step answer\-block accuracy and Jensen–Shannon divergence from the correct\-only distribution on natural retrieval cases\.Our contributions are:

- •We identify an availability–accuracy mismatch in visual DLM\-RAG: broader retrieval improves answer\-page coverage but can reduce answer accuracy\. We trace this failure to source\-coherence loss in parallel denoising\.
- •We introduce ECF, which assesses additional evidence through a geometry\-matched counterfactual in the target DLM’s first\-step answer block\. Together with multi\-granularity evidence construction, it decouples candidate\-pool size from final\-context size\.
- •Across five visual QA benchmarks and three multimodal DLMs, we show that selective evidence admission provides more reliable gains than fixed top\-kkinput and remains stable as the candidate pool expands\.

## 2Related Work

#### Visual RAG under imperfect evidence\.

Visual RAG retrieves external evidence directly from document images, preserving layout, graphics, and textual content that may be lost in text\-only pipelines\(Yuet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib8); Liet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib9)\)\. Existing work has improved visual evidence quality along two directions: stronger page\-level retrieval and fusion, and finer\-grained retrieval of query\-relevant regions\. RegionRAG develops the latter direction by learning query–region alignment, and shows that irrelevant content can arise both within an individual page and across multiple retrieved pages\(Liet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib9)\)\. Related robustness studies likewise find that irrelevant or poorly placed context can degrade generation\(Yoranet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib5); Yuet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib6); Liuet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib7)\)\. These studies establish evidence granularity and candidate redundancy as important sources of noise, but do not characterize how multiple visual sources interact once they jointly condition diffusion generation\.

#### Diffusion language models and DLM\-RAG\.

Discrete and masked diffusion language models generate through iterative denoising\(Austinet al\.[2021](https://arxiv.org/html/2608.07006#bib.bib10); Hoogeboomet al\.[2021](https://arxiv.org/html/2608.07006#bib.bib11); Sahooet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib12); Nieet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib13); Gaoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib33)\)rather than autoregressive factorization\(Gaoet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib34); Xionget al\.[2024](https://arxiv.org/html/2608.07006#bib.bib35)\), and recent work extends this paradigm to multimodal generation\(Inclusion AIet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib25); Youet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib26); Yeet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib27)\)\. Because multiple positions can be updated from a shared masked state, their behavior depends on denoising order and on dependencies not represented by isolated position\-wise marginals\(Niet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib14); Kanget al\.[2026](https://arxiv.org/html/2608.07006#bib.bib18); Zhanget al\.[2026](https://arxiv.org/html/2608.07006#bib.bib19)\)\. For DLM\-RAG, SPREAD uses query relevance to guide reveal order, while SARDI updates retrieval from tentative predictions\(Yuet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib17); Jüngeret al\.[2026](https://arxiv.org/html/2608.07006#bib.bib15)\)\. These methods study how available evidence is used or refreshed during denoising, rather than which visual candidates should jointly enter the initial masked context or how incompatible sources interact within it\.

#### Adaptive evidence selection\.

Adaptive RAG systems decide when to retrieve and which evidence to retain using retrieval scores, generation uncertainty, self\-reflection, learned verifiers, or surrogate multimodal models\(Asaiet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib3); Jianget al\.[2023](https://arxiv.org/html/2608.07006#bib.bib4); Luoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib16)\)\. UOVES provides the most direct generation\-aware formulation for visual evidence selection: it defines evidence utility as information gain on the model’s output distribution and derives a latent helpfulness objective that can be estimated by a lightweight surrogate in a training\-free manner\(Luoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib16)\)\. This formulation moves beyond semantic relevance, but operationally ranks candidates by their individual helpfulness and retains a prescribed top\-KKsubset\. It does not directly address context\-dependent admission—whether an additional candidate remains beneficial when combined with already available visual evidence, or whether context expansion should be rejected before DLM denoising\. This work addresses this gap by studying evidence admission for DLMs whose unresolved answer positions share a multi\-source visual context\.

## 3Why More Retrieved Evidence Can Hurt Visual DLM\-RAG

### 3\.1Preliminaries: Masked Diffusion Generation

Unlike autoregressive models that decode from left to right, DLMs generate answers through iterative denoising\(Sahooet al\.[2024](https://arxiv.org/html/2608.07006#bib.bib12); Nieet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib13); Gaoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib33)\)\. LetYt=\(Y1t,…,YMt\)Y^\{t\}=\(Y\_\{1\}^\{t\},\\ldots,Y\_\{M\}^\{t\}\)be the answer block at stepttandUtU\_\{t\}its unresolved masked positions\. Generation starts from a fully maskedYTY^\{T\}withUT=\{1,…,M\}U\_\{T\}=\\\{1,\\ldots,M\\\}\. Given questionqq, visual evidenceEE, and stateYtY^\{t\}, one forward pass exposespθ\(m\)​\(y∣Yt,q,E\)p\_\{\\theta\}^\{\(m\)\}\(y\\mid Y^\{t\},q,E\)for everym∈Utm\\in U\_\{t\}\. A schedule selectsSt⊆UtS\_\{t\}\\subseteq U\_\{t\}and updates those positions from the same unresolved state using the position\-wise proposal

Qt​\(YSt∣Yt,q,E\)=∏m∈Stpθ\(m\)​\(Ym∣Yt,q,E\)\.Q\_\{t\}\(Y\_\{S\_\{t\}\}\\mid Y^\{t\},q,E\)=\\prod\_\{m\\in S\_\{t\}\}p\_\{\\theta\}^\{\(m\)\}\(Y\_\{m\}\\mid Y^\{t\},q,E\)\.\(1\)The process repeats until no masks remain\. Although implementations differ in their schedules and update budgets, their first forward pass exposes distributions over the prospective answer block before any answer token is committed\. This multi\-position lookahead enables parallel decoding, but a position\-wise proposal may fail to preserve a source choice shared across unresolved positions\. We analyze this first unresolved state and suppress the fixedYTY^\{T\}below\.

### 3\.2Problem Setting: Availability vs\. Admission

Each example consists of a questionqq, a visual corpus𝒟\\mathcal\{D\}, and one or more reference answers\. A retriever returns a ranked candidate poolCk=\(c1,c2,…,ck\)C\_\{k\}=\(c\_\{1\},c\_\{2\},\\ldots,c\_\{k\}\), from which the generator receives an admitted evidence sequenceE⊆CkE\\subseteq C\_\{k\}\. Fixed top\-kkinput setsE=CkE=C\_\{k\}, conflating candidate availability with evidence admission\. Answer\-page Recall@kkmeasures availability—whetherCkC\_\{k\}contains an annotated answer\-bearing page—whereas answer accuracy measures generation quality under the admitted evidenceEE\. We seek to preserve the coverage benefit of a larger candidate pool while limiting harmful visual exposure during denoising\.

### 3\.3Source\-Coherence Loss in Parallel Denoising

The controlled interventions in Figure[2](https://arxiv.org/html/2608.07006#S1.F2)\(a\) show that the degradation is driven primarily by competing visual semantics rather than the presence of a second visual slot\. We therefore model the retrieved pages as competing sources that support different coherent answers and formalize why this conflict is especially damaging when several positions are predicted from the same unresolved denoising state\.

###### Definition 1\(Conflict width and source coherence\)\.

Let the correct and conflicting pages support answersa=\(a1,…,aM\)a=\(a\_\{1\},\\ldots,a\_\{M\}\)andd=\(d1,…,dM\)d=\(d\_\{1\},\\ldots,d\_\{M\}\)\. The conflict set isR=\{m:am≠dm\}R=\\\{m:a\_\{m\}\\neq d\_\{m\}\\\}and its width isr=\|R\|r=\|R\|\. The answersaaandddare*source\-coherent*; the hybrid setℋ\\mathcal\{H\}contains assignments that agree with them outsideRRbut select at least one token from each source insideRR, and hence are supported by neither source\.

We formalize this unresolved source conflict through local two\-source ambiguity and an exact\-marginal factorized proposal on a restricted source\-label process; their precise definitions and scope are given as Assumptions F\.1 and F\.2 in the Technical Supplement\. At each conflict position, this process retains and renormalizes the correct\- and competing\-source tokens\. Hereϵ∈\(0,1\)\\epsilon\\in\(0,1\)denotes competing\-source mass in this restricted process\. The key issue is that positions inRRshare a latent source choice, whereas the position\-wise proposal does not\. Forr=2r=2, the mixture supports only\(a1,a2\)\(a\_\{1\},a\_\{2\}\)and\(d1,d2\)\(d\_\{1\},d\_\{2\}\), but the product assigns mass2​ϵ​\(1−ϵ\)2\\epsilon\(1\-\\epsilon\)to the two cross\-source combinations\.

###### Theorem 1\(Source\-conflict amplification\)\.

Under Assumptions F\.1 and F\.2 in the Technical Supplement, for any0<ϵ<10<\\epsilon<1andr≥1r\\geq 1, the factorized proposal assigns

Q∥​\(a∣E\)\\displaystyle Q\_\{\\parallel\}\(a\\mid E\)=\(1−ϵ\)r,\\displaystyle=\(1\-\\epsilon\)^\{r\},Q∥​\(d∣E\)\\displaystyle Q\_\{\\parallel\}\(d\\mid E\)=ϵr,\\displaystyle=\\epsilon^\{r\},\(2\)Q∥​\(ℋ∣E\)\\displaystyle Q\_\{\\parallel\}\(\\mathcal\{H\}\\mid E\)=1−\(1−ϵ\)r−ϵr\.\\displaystyle=1\-\(1\-\\epsilon\)^\{r\}\-\\epsilon^\{r\}\.Consequently, hybrid mass is zero forr=1r=1and strictly increases with conflict widthrr\.

Thus, once several positions depend on the unresolved source, position\-wise decisions can splice coherent branches and move probability mass to unsupported answers\.

###### Corollary 2\(Excess risk from lost source coupling\)\.

Under Assumptions F\.1 and F\.2 in the Technical Supplement, suppose the coherent branchaais the unique accepted assignment in this restricted process, and define its exact\-match sampling risk asℛ​\(Q\)=1−Q​\(a\)\\mathcal\{R\}\(Q\)=1\-Q\(a\)\. LetQcausalQ\_\{\\mathrm\{causal\}\}be an exact causal factorization of the same mixture, which preserves the selected source after the prefix chooses a branch\. Thenℛcausal=ϵ\\mathcal\{R\}\_\{\\mathrm\{causal\}\}=\\epsilon, and for everyr≥2r\\geq 2,

ℛ∥−ℛcausal=\(1−ϵ\)−\(1−ϵ\)r\>0,r≥2\.\\mathcal\{R\}\_\{\\parallel\}\-\\mathcal\{R\}\_\{\\mathrm\{causal\}\}=\(1\-\\epsilon\)\-\(1\-\\epsilon\)^\{r\}\>0,\\qquad r\\geq 2\.\(3\)Hence factorizing exact marginals incurs excess risk solely from losing their shared source dependence\.

The causal construction is an oracle comparator that isolates this dependence cost, rather than a universal ranking of model families\. The result explains why conflict is more damaging than a semantically empty input and motivates admission before final denoising; proofs are provided in the Technical Supplement\.

Figure[3](https://arxiv.org/html/2608.07006#S3.F3)summarizes the mechanism: fixed top\-kkinput can expose parallel denoising to competing visual sources, and the factorized position\-wise proposal can convert unresolved source competition into unsupported hybrid answers\.

![Refer to caption](https://arxiv.org/html/2608.07006v1/x3.png)Figure 3:Fixed top\-kkinput can improve answer\-page availability while conflicting sources induce unsupported hybrid proposals\.Table 1:Restricted hybrid source mass under controlled visual conflict\. Each cell averages 32 base templates and both page orders; the parallel–sequential effect has the same direction under either order\.
### 3\.4Empirical Validation and Early Observability

#### Controlled source\-conflict validation\.

Table[1](https://arxiv.org/html/2608.07006#S3.T1)evaluates the source\-conflict prediction through a same\-weight LLaDA comparison differing only in proposal construction\. Hybrid mass is zero atr=1r=1, increases with conflict width, and is consistently higher under parallel than source\-conditioned sequential prediction forr≥2r\\geq 2, but remains negligible for native Qwen AR\. This supports source\-coherence loss as the failure channel behind harmful multi\-page input\.

#### Early observability on natural retrieval\.

The same failure is visible before any answer token is committed\. On natural ChartQA cases, a conflicting page reduces first\-step answer\-block accuracy from 65\.0 to 41\.1 and produces the largest divergence from the correct\-only distribution, whereas matched\-blank and duplicated\-correct controls remain substantially closer \(Figure[2](https://arxiv.org/html/2608.07006#S1.F2)\(b\)\)\. This ordering mirrors final\-answer degradation \(Figure[2](https://arxiv.org/html/2608.07006#S1.F2)\(a\)\), showing that conflict is already expressed in the initial answer state and persists through denoising\. The early\-to\-final correspondence motivates the first\-step distribution as a generator\-native signal for evidence admission\.

## 4Entropy\-Based Candidate Filter

The analysis motivates ECF to control visual evidence exposure along two axes: evidence granularity and final\-context composition\. At evidence construction, full pages and layout\-derived regions form a multi\-granularity corpus spanning global and localized evidence\. At admission, ECF evaluates candidates before decoding, using a geometry\-matched blank to isolate their effect on prospective answer\-block confidence\. Given a ranked top\-kkpool, it expands the visual context only when supported by the DLM’s pre\-denoising state\.

![Refer to caption](https://arxiv.org/html/2608.07006v1/x4.png)Figure 4:Overview of ECF’s three core components\. Visual slicing constructs multi\-granularity evidence units from full pages and layout\-derived regions\. Entropy computation measures prospective answer\-block confidence and isolates each candidate’s semantic contribution with a geometry\-matched blank\. Rank\-prior selection combines these signals with retrieval order to preserve the top\-ranked candidate and admit at most one companion\.### 4\.1DLM\-Native Block Confidence

Because candidate influence is visible before final denoising, we evaluate uncertainty in the DLM’s prospective answer state\. For an evidence setEE, define the entropy of answer slotmm:

hm​\(q,E\)=−∑ypθ\(m\)​\(y∣q,E\)​log⁡pθ\(m\)​\(y∣q,E\)\.h\_\{m\}\(q,E\)=\-\\sum\_\{y\}p\_\{\\theta\}^\{\(m\)\}\(y\\mid q,E\)\\log p\_\{\\theta\}^\{\(m\)\}\(y\\mid q,E\)\.\(4\)We aggregate the first answer block as

H​\(q,E\)\\displaystyle H\(q,E\)=∑m=1Mαm​hm​\(q,E\),\\displaystyle=\\sum\_\{m=1\}^\{M\}\\alpha\_\{m\}h\_\{m\}\(q,E\),\(5\)αm\\displaystyle\\alpha\_\{m\}=exp⁡\(−λ​m\)∑j=1Mexp⁡\(−λ​j\)\.\\displaystyle=\\frac\{\\exp\(\-\\lambda m\)\}\{\\sum\_\{j=1\}^\{M\}\\exp\(\-\\lambda j\)\}\.The decay rateλ\\lambdacontrols the emphasis on earlier answer positions; we useλ=0\.5\\lambda=0\.5throughout\. This block statistic uses the multi\-position state implicated by Theorem[1](https://arxiv.org/html/2608.07006#Thmtheorem1), rather than relying on a single initial token that may encode formatting or a standard prefix\.

### 4\.2Blank\-Controlled Candidate Gain

Raw block\-entropy changes mix the candidate’s semantic contribution with the structural effect of adding another visual input\. ECF separates these effects with geometry\-matched real and blank\-reference contexts\.

For a top\-ranked evidence unitc1c\_\{1\}and a candidate unitcic\_\{i\}, construct matched contexts

Eireal=\{c1,ci\},Eiblank=\{c1,blank​\(ci\)\}\.E\_\{i\}^\{\\mathrm\{real\}\}=\\\{c\_\{1\},c\_\{i\}\\\},\\quad E\_\{i\}^\{\\mathrm\{blank\}\}=\\\{c\_\{1\},\\mathrm\{blank\}\(c\_\{i\}\)\\\}\.\(6\)The blank has the same dimensions ascic\_\{i\}and preserves the resulting image\-token layout while removing candidate semantics\. We define the blank\-controlled entropy gain

Gi=H​\(q,Eiblank\)−H​\(q,Eireal\)\.G\_\{i\}=H\(q,E\_\{i\}^\{\\mathrm\{blank\}\}\)\-H\(q,E\_\{i\}^\{\\mathrm\{real\}\}\)\.\(7\)The difference therefore isolates howcic\_\{i\}changes the prospective answer state relative to an equally shaped semantic null\.

#### Alignment with source\-conflict risk\.

The matched\-context decomposition and local correct\-source condition are formalized in Assumption F\.3 of the Technical Supplement\.

###### Proposition 3\(Local confidence–risk alignment\)\.

Under Assumption F\.3 in the Technical Supplement, forr≥2r\\geq 2, positive candidate gain is equivalent to reducing competing\-source mass, hybrid mass, and parallel exact\-match risk:

Gi\>0\\displaystyle G\_\{i\}\>0⟺ϵi<ϵblank\\displaystyle\\Longleftrightarrow\\epsilon\_\{i\}<\\epsilon\_\{\\mathrm\{blank\}\}\(8\)⟺Φr​\(ϵi\)<Φr​\(ϵblank\)\\displaystyle\\Longleftrightarrow\\Phi\_\{r\}\(\\epsilon\_\{i\}\)<\\Phi\_\{r\}\(\\epsilon\_\{\\mathrm\{blank\}\}\)⟺ℛ∥​\(ϵi\)<ℛ∥​\(ϵblank\),\\displaystyle\\Longleftrightarrow\\mathcal\{R\}\_\{\\parallel\}\(\\epsilon\_\{i\}\)<\\mathcal\{R\}\_\{\\parallel\}\(\\epsilon\_\{\\mathrm\{blank\}\}\),whereϵi\\epsilon\_\{i\}andϵblank\\epsilon\_\{\\mathrm\{blank\}\}are the competing\-source masses under the real and blank contexts, respectively,Φr​\(ϵ\)=1−\(1−ϵ\)r−ϵr\\Phi\_\{r\}\(\\epsilon\)=1\-\(1\-\\epsilon\)^\{r\}\-\\epsilon^\{r\}is the hybrid mass, andℛ∥​\(ϵ\)=1−\(1−ϵ\)r\\mathcal\{R\}\_\{\\parallel\}\(\\epsilon\)=1\-\(1\-\\epsilon\)^\{r\}is the parallel exact\-match risk\.

Proposition[3](https://arxiv.org/html/2608.07006#Thmtheorem3)gives the gain a direct operational meaning: relative to an equally shaped blank, positive gain indicates that the candidate reduces competing\-source uncertainty and therefore lowers both hybrid mass and parallel sampling risk\. This is why ECF uses a candidate\-specific entropy difference rather than raw entropy\. The alignment is deliberately local: outside the correct\-source basin, low entropy may reflect a confidently wrong source, so retrieval rank supplies the complementary prior\.

Table 2:Answer accuracy \(%\) under fixed top\-kkinput and training\-free evidence\-selection alternatives with LLaDA2\.0\-Uni\. Fixed top\-2 input unconditionally passes the two highest\-ranked retrieved pages; ECF and the other baselines use a candidate\-pool size ofk=3k=3and admit at most one companion\.Table 3:Answer accuracy \(%\) across DLM backends\. ECF uses candidate\-pool sizek=3k=3\. For each dataset within a backend, the best and second\-best results are shown in bold and underlined, respectively; only the best average is shown in bold\.Table 4:Answer\-page Recall@kkand answer\-page usage \(%\)\.Table 5:ECF with larger candidate pools\.

### 4\.3Rank\-Prior Selection

The gain provides generator\-side evidence for each candidate, while retrieval order supplies the prior for searching the ranked pool\. ECF combines both signals so that lower\-ranked candidates replace the default expansion only with stronger evidence\. It consumes only the ranked candidates, without retriever\-specific scores, score calibration, or learned parameters, and can be paired with any visual retriever that returns an ordered list\.

Algorithm 1Rank\-Prior Selection in ECF0:Question

qq; ranked candidate pool

Ck=\(c1,…,ck\)C\_\{k\}=\(c\_\{1\},\\ldots,c\_\{k\}\); DLM

pθp\_\{\\theta\}; matched\-blank operator

blank​\(⋅\)\\mathrm\{blank\}\(\\cdot\)
0:Admitted evidence set

E⋆E^\{\\star\}
1:Stage 1: Preserve the top\-ranked candidate

2:if

k=1k=1then

3:return

\{c1\}\\\{c\_\{1\}\\\}
4:endif

5:Stage 2: Test the default rank\-2 expansion

6:

H1←H​\(q,\{c1\}\)H\_\{1\}\\leftarrow H\(q,\\\{c\_\{1\}\\\}\)
7:

G2←H​\(q,E2blank\)−H​\(q,E2real\)G\_\{2\}\\leftarrow H\(q,E\_\{2\}^\{\\mathrm\{blank\}\}\)\-H\(q,E\_\{2\}^\{\\mathrm\{real\}\}\)
8:if

G2≤0G\_\{2\}\\leq 0then

9:return

\{c1\}\\\{c\_\{1\}\\\}
10:endif

11:

j←2j\\leftarrow 2
12:Stage 3: Evaluate lower\-ranked replacements

13:for

m=3,…,km=3,\\ldots,kdo

14:

Gm←H​\(q,Emblank\)−H​\(q,Emreal\)G\_\{m\}\\leftarrow H\(q,E\_\{m\}^\{\\mathrm\{blank\}\}\)\-H\(q,E\_\{m\}^\{\\mathrm\{real\}\}\)
15:

H1​m←H​\(q,\{c1,cm\}\)H\_\{1m\}\\leftarrow H\(q,\\\{c\_\{1\},c\_\{m\}\\\}\)
16:if

Gm\>GjG\_\{m\}\>G\_\{j\}and

H1​m<H1H\_\{1m\}<H\_\{1\}then

17:

j←mj\\leftarrow m
18:endif

19:endfor

20:Stage 4: Construct the admitted evidence set

21:return

\{c1,cj\}\\\{c\_\{1\},c\_\{j\}\\\}

The policy always retainsc1c\_\{1\}and admits at most one companion\. Fork≥2k\\geq 2,c2c\_\{2\}serves as the rank\-prior default and is admitted only ifG2\>0G\_\{2\}\>0; otherwise, the policy returnsE⋆=\{c1\}E^\{\\star\}=\\\{c\_\{1\}\\\}\. Conditional onG2\>0G\_\{2\}\>0, the admissible lower\-ranked replacements are

𝒜k=\{m∈\{3,…,k\}:\\displaystyle\\mathcal\{A\}\_\{k\}=\\\{m\\in\\\{3,\\ldots,k\\\}:\{\}Gm\>G2,\\displaystyle G\_\{m\}\>G\_\{2\},\(9\)H\(q,\{c1,cm\}\)<H\(q,\{c1\}\)\}\.\\displaystyle H\(q,\\\{c\_\{1\},c\_\{m\}\\\}\)<H\(q,\\\{c\_\{1\}\\\}\)\\\}\.The selected companion is

j⋆=arg⁡maxj∈\{2\}∪𝒜k⁡Gj,E⋆=\{c1,cj⋆\},j^\{\\star\}=\\arg\\max\_\{j\\in\\\{2\\\}\\cup\\mathcal\{A\}\_\{k\}\}G\_\{j\},\\qquad E^\{\\star\}=\\\{c\_\{1\},c\_\{j^\{\\star\}\}\\\},\(10\)with ties favoring the higher\-ranked candidate\. This construction treatsc2c\_\{2\}as the default expansion while requiring lower\-ranked candidates to improve both relative gain and absolute joint confidence\. Its probe cost grows linearly withkk, but each real or blank probe requires only a single first\-step forward pass rather than a complete denoising trajectory\. Full answer generation is performed only once with at most two admitted images\.

## 5Experiments

### 5\.1Experimental Setup

#### Benchmarks and system setup\.

We evaluate ECF with three multimodal DLM backends: LLaDA2\.0\-Uni\(Inclusion AIet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib25)\), LLaDA\-V\(Youet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib26)\), and Dream\-VL\(Yeet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib27)\)\. For retrieval, a VisRAG\-style retriever\(Yuet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib8)\)ranks visual evidence units for each question\.

We evaluate five visual QA benchmarks: ChartQA\(Masryet al\.[2022](https://arxiv.org/html/2608.07006#bib.bib20)\)and InfoChartQA\(Xieet al\.[2025](https://arxiv.org/html/2608.07006#bib.bib21)\)for charts, plus DocVQA\(Mathewet al\.[2021](https://arxiv.org/html/2608.07006#bib.bib22)\), InfoVQA\(Mathewet al\.[2022](https://arxiv.org/html/2608.07006#bib.bib23)\), and TATDQA\(Zhuet al\.[2022](https://arxiv.org/html/2608.07006#bib.bib24)\)for dense documents\. ECF retrieves full pages and layout\-derived slices from a multi\-granularity corpus\.

#### Evaluation protocol\.

We report answer accuracy \(%\) as the primary metric, using a unified relaxed matching rule for textual and numeric answers across all datasets and models\. For diagnostics, we also report Recall@kkfor answer\-page availability and answer\-page usage for the presence of answer\-bearing evidence in the admitted context\. Further details appear in the Technical Supplement\.

#### Training\-Free Alternatives\.

We compare ECF with five training\-free alternatives under LLaDA2\.0\-Uni\. UOVES selects two pages with a multimodal surrogate; Answer\-UQ selects the candidate with the lowest generated\-answer uncertainty; Frozen Semantic reranks with a frozen vision–text encoder; SPREAD changes reveal order using query relevance; and SARDI retrieves from tentative DLM predictions\(Luoet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib16); Yuet al\.[2026](https://arxiv.org/html/2608.07006#bib.bib17); Jüngeret al\.[2026](https://arxiv.org/html/2608.07006#bib.bib15)\)\. All methods share VisRAG retrieval, the answer budget, and the scorer; fixed top\-kkinputs expand evidence unconditionally\.

### 5\.2Main Results

#### LLaDA2\.0\-Uni\.

Table[2](https://arxiv.org/html/2608.07006#S4.T2)tests whether ECF’s gains merely reflect avoiding unconditional evidence expansion\. ECF improves over matched fixed top\-2 input by 6\.90 percentage points on average\. More importantly, it retains a 2\.37\-percentage\-point average margin over the strongest training\-free alternative on each dataset, even though these methods already select, rerank, or adapt retrieved evidence\. Thus, passing fewer pages alone does not explain the gain, supporting a generator\-native criterion for evidence admission before denoising\.

#### Cross\-DLM validation\.

Table[3](https://arxiv.org/html/2608.07006#S4.T3)evaluates the admission rule across DLM backends\. ECF achieves the best accuracy in 10 of the 15 model–dataset pairs and the highest average for all three backends\. Across the 15 pairs, the same training\-free rule improves accuracy by 2\.62 percentage points on average over the strongest fixed top\-kkinput, without backend\-specific training or calibration\. Its cross\-backend behavior characterizes ECF as a reusable evidence\-admission layer rather than a generator\-specific optimization\.

#### Candidate Availability and Answer\-Page Usage\.

Table[4](https://arxiv.org/html/2608.07006#S4.T4)separates candidate availability from evidence admitted for generation\. Recall@kkgrows with pool size, but Table[3](https://arxiv.org/html/2608.07006#S4.T3)shows that admitting more pages does not correspondingly improve accuracy\. Answer\-page usage measures whether answer\-bearing evidence survives admission, separating retrieval coverage from effective generation context\.

### 5\.3Analysis and Ablations

#### Larger candidate pools\.

Table[5](https://arxiv.org/html/2608.07006#S4.T5)evaluates ECF beyond the maink=3k=3candidate pool\. Unlike fixed top\-kkinput, whose accuracy decreases as more retrieved pages enter the decoding context \(Figure[1](https://arxiv.org/html/2608.07006#S1.F1)\), ECF remains stable as its candidate pool grows: macro accuracy is 30\.11, 30\.19, and 30\.32 fork=3,4,5k=3,4,5, respectively, despite diminishing marginal Recall gains\. Increasingkkexpands only the candidates screened by ECF, while final decoding remains capped at two images\. This decoupling makeskka flexible retrieval budget: small pools suffice when retrieval saturates early, whereas larger pools can exploit informative lower\-ranked evidence from weaker retrievers or more complex datasets without the degradation of fixed top\-kkinput\. ECF thus adapts across retrieval regimes rather than relying on a fixedkk\.

![Refer to caption](https://arxiv.org/html/2608.07006v1/x5.png)Figure 5:Macro\-average answer accuracy across DocVQA, InfoVQA, and TATDQA for the slicing and gating ablation\.
#### Slicing and admission components\.

Figure[5](https://arxiv.org/html/2608.07006#S5.F5)isolates the contributions of the two components in the complete ECF pipeline: layout\-aware slicing and entropy\-based candidate filtering\. Each component independently improves the fixed top\-1 baseline by about four percentage points, while combining them yields a substantially larger gain of 9\.72 percentage points\. These gains reflect their complementary roles: slicing reduces unrelated content within each retrieved image, whereas admission limits interference across images in the final denoising context\.

## 6Conclusion

Retrieving more evidence does not guarantee better visual DLM\-RAG: larger candidate pools improve answer\-page availability, but admitting every candidate can reduce accuracy when position\-wise parallel proposals lose shared visual\-source dependence and assign probability to unsupported answers\. Controlled and natural\-retrieval studies locate this interference in the first\-step answer block, enabling evidence assessment before decoding\. ECF combines multi\-granularity evidence construction with blank\-controlled, rank\-aware admission to reduce within\-candidate noise and limit which candidates reach final denoising\. Across five visual QA benchmarks and three multimodal DLMs, ECF averages a 2\.62\-percentage\-point gain over the strongest fixed top\-kkinput and remains stable as candidate pools grow\. These findings recast pool size as a retrieval budget: broader retrieval helps when evidence is screened before denoising rather than accumulated unconditionally\.

## References

- A\. Asai, Z\. Wu, Y\. Wang, A\. Sil, and H\. Hajishirzi \(2024\)Self\-RAG: learning to retrieve, generate, and critique through self\-reflection\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=hSyW5go0v8)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p2.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px3.p1.1)\.
- J\. Austin, D\. D\. Johnson, J\. Ho, D\. Tarlow, and R\. van den Berg \(2021\)Structured denoising diffusion models in discrete state\-spaces\.InAdvances in Neural Information Processing Systems,Vol\.34,pp\. 17981–17993\.External Links:[Link](https://proceedings.neurips.cc/paper/2021/hash/958c530554f78bcd8e97125b70e6973d-Abstract.html)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Gao, J\. Bai, Y\. Huang, X\. Fu, Q\. Sun, and Y\. Song \(2026\)Unifying deductive and abductive reasoning in knowledge graphs with masked diffusion model\.InProceedings of the ACM Web Conference 2026,pp\. 3600–3611\.Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2608.07006#S3.SS1.p1.11)\.
- Y\. Gao, J\. Bai, T\. Zheng, Q\. Sun, Z\. Zhang, X\. Fu, J\. Li, and Y\. Song \(2025\)Controllable logical hypothesis generation for abductive reasoning in knowledge graphs\.arXiv preprint arXiv:2505\.20948\.Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1)\.
- E\. Hoogeboom, D\. Nielsen, P\. Jaini, P\. Forré, and M\. Welling \(2021\)Argmax flows and multinomial diffusion: learning categorical distributions\.InAdvances in Neural Information Processing Systems,Vol\.34,pp\. 12454–12465\.External Links:[Link](https://proceedings.neurips.cc/paper/2021/hash/67d96d458abdef21792e6d8e590244e7-Abstract.html)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1)\.
- Inclusion AI, T\. Bie, H\. Chen, T\. Chen, Z\. Cheng, L\. Cui, K\. Gan, Z\. Huang, Z\. Lan, H\. Li, J\. Li, T\. Lin, Q\. Qin, H\. Wang, X\. Wang, H\. Wu, Y\. Xin, and J\. Zhao \(2026\)LLaDA2\.0\-Uni: unifying multimodal understanding and generation with diffusion large language model\.External Links:2604\.20796,[Link](https://arxiv.org/abs/2604.20796)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1),[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p1.1)\.
- G\. Izacard and E\. Grave \(2021\)Leveraging passage retrieval with generative models for open domain question answering\.InProceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume,Online,pp\. 874–880\.External Links:[Document](https://dx.doi.org/10.18653/v1/2021.eacl-main.74),[Link](https://aclanthology.org/2021.eacl-main.74/)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p2.1)\.
- Z\. Jiang, F\. Xu, L\. Gao, Z\. Sun, Q\. Liu, J\. Dwivedi\-Yu, Y\. Yang, J\. Callan, and G\. Neubig \(2023\)Active retrieval augmented generation\.InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,Singapore,pp\. 7969–7992\.External Links:[Document](https://dx.doi.org/10.18653/v1/2023.emnlp-main.495),[Link](https://aclanthology.org/2023.emnlp-main.495/)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px3.p1.1)\.
- P\. Jünger, J\. Lovelace, L\. Zhao, D\. Go, and K\. Q\. Weinberger \(2026\)Self\-augmenting retrieval for diffusion language models\.InProceedings of the 43rd International Conference on Machine Learning,External Links:[Link](https://icml.cc/virtual/2026/poster/60996)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1),[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px3.p1.1)\.
- W\. Kang, K\. Galim, S\. Oh, M\. Lee, Y\. Zeng, S\. Zhang, C\. R\. C\. Hooper, Y\. Hu, H\. I\. Koo, N\. I\. Cho, and K\. Lee \(2026\)ParallelBench: understanding the trade\-offs of parallel decoding in diffusion LLMs\.InThe Fourteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=OsZr5T7Cd0)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1)\.
- Y\. Li, Z\. Lu, Z\. Liu, Y\. Sun, C\. Liu, and H\. Xie \(2025\)RegionRAG: region\-level retrieval\-augmented generation for visual document understanding\.External Links:2510\.27261,[Link](https://arxiv.org/abs/2510.27261)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p2.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px1.p1.1)\.
- N\. F\. Liu, K\. Lin, J\. Hewitt, A\. Paranjape, M\. Bevilacqua, F\. Petroni, and P\. Liang \(2024\)Lost in the middle: how language models use long contexts\.Transactions of the Association for Computational Linguistics12,pp\. 157–173\.External Links:[Document](https://dx.doi.org/10.1162/tacl%5Fa%5F00638),[Link](https://doi.org/10.1162/tacl_a_00638)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px1.p1.1)\.
- W\. Luo, Z\. Hu, X\. Wang, Z\. Yu, H\. Zhang, and Z\. Huang \(2026\)Utility\-oriented visual evidence selection for multimodal retrieval\-augmented generation\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),San Diego, California, United States,pp\. 35091–35124\.External Links:[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1620),[Link](https://aclanthology.org/2026.acl-long.1620/)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p2.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px3.p1.1),[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px3.p1.1)\.
- X\. Ma, Q\. Liu, D\. Jiang, G\. Zhang, Z\. Ma, and W\. Chen \(2025\)General\-reasoner: advancing LLM reasoning across all domains\.InAdvances in Neural Information Processing Systems,D\. Belgrave, C\. Zhang, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, and N\. Chen \(Eds\.\),Vol\.38,pp\. 56596–56618\.External Links:[Link](https://proceedings.neurips.cc/paper_files/paper/2025/file/51c52cbb4b3faa891ca0621d13efe1be-Paper-Conference.pdf)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1)\.
- A\. Masry, D\. X\. Long, J\. Q\. Tan, S\. Joty, and E\. Hoque \(2022\)ChartQA: a benchmark for question answering about charts with visual and logical reasoning\.InFindings of the Association for Computational Linguistics: ACL 2022,Dublin, Ireland,pp\. 2263–2279\.External Links:[Document](https://dx.doi.org/10.18653/v1/2022.findings-acl.177),[Link](https://aclanthology.org/2022.findings-acl.177/)Cited by:[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p2.1)\.
- M\. Mathew, V\. Bagal, R\. Tito, D\. Karatzas, E\. Valveny, and C\. V\. Jawahar \(2022\)InfographicVQA\.In2022 IEEE/CVF Winter Conference on Applications of Computer Vision,pp\. 2582–2591\.External Links:[Document](https://dx.doi.org/10.1109/WACV51458.2022.00264),[Link](https://doi.org/10.1109/WACV51458.2022.00264)Cited by:[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p2.1)\.
- M\. Mathew, D\. Karatzas, and C\. V\. Jawahar \(2021\)DocVQA: a dataset for VQA on document images\.In2021 IEEE Winter Conference on Applications of Computer Vision,pp\. 2199–2208\.External Links:[Document](https://dx.doi.org/10.1109/WACV48630.2021.00225),[Link](https://doi.org/10.1109/WACV48630.2021.00225)Cited by:[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p2.1)\.
- Z\. Ni, S\. Wang, Y\. Yue, T\. Yu, W\. Zhao, Y\. Hua, T\. Chen, J\. Song, C\. Yu, B\. Zheng, and G\. Huang \(2026\)The flexibility trap: rethinking the value of arbitrary order in diffusion language models\.InProceedings of the 43rd International Conference on Machine Learning,External Links:[Link](https://icml.cc/virtual/2026/poster/61998)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1)\.
- S\. Nie, F\. Zhu, Z\. You, X\. Zhang, J\. Ou, J\. Hu, J\. Zhou, Y\. Lin, J\. Wen, and C\. Li \(2025\)Large language diffusion models\.External Links:2502\.09992,[Link](https://arxiv.org/abs/2502.09992)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2608.07006#S3.SS1.p1.11)\.
- S\. S\. Sahoo, M\. Arriola, Y\. Schiff, A\. Gokaslan, E\. Marroquin, J\. T\. Chiu, A\. Rush, and V\. Kuleshov \(2024\)Simple and effective masked diffusion language models\.InAdvances in Neural Information Processing Systems,Vol\.37,pp\. 130136–130184\.External Links:[Document](https://dx.doi.org/10.52202/079017-4135),[Link](https://proceedings.neurips.cc/paper_files/paper/2024/hash/eb0b13cc515724ab8015bc978fdde0ad-Abstract-Conference.html)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1),[§3\.1](https://arxiv.org/html/2608.07006#S3.SS1.p1.11)\.
- W\. Shi, S\. Min, M\. Yasunaga, M\. Seo, R\. James, M\. Lewis, L\. Zettlemoyer, and W\. Yih \(2024\)REPLUG: retrieval\-augmented black\-box language models\.InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),Mexico City, Mexico,pp\. 8371–8384\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.naacl-long.463),[Link](https://aclanthology.org/2024.naacl-long.463/)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p2.1)\.
- Q\. Xiao, H\. Shi, Y\. Gao, W\. Hu, H\. Jing, T\. Zheng, B\. Xu, Z\. Zhang, W\. Wang, H\. Li,et al\.\(2026\)SING: synthetic intention graph for scalable active tool discovery in llm agents\.arXiv preprint arXiv:2606\.16591\.Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1)\.
- T\. Xie, M\. Lin, M\. Liu, Y\. Ye, C\. Chen, and S\. Liu \(2025\)InfoChartQA: a benchmark for multimodal question answering on infographic charts\.External Links:2505\.19028,[Link](https://arxiv.org/abs/2505.19028)Cited by:[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p2.1)\.
- J\. Xiong, G\. Liu, L\. Huang, C\. Wu, T\. Wu, Y\. Mu, Y\. Yao, H\. Shen, Z\. Wan, J\. Huang,et al\.\(2024\)Autoregressive models in vision: a survey\.arXiv preprint arXiv:2411\.05902\.Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1)\.
- J\. Ye, S\. Gong, J\. Gao, J\. Fan, S\. Wu, W\. Bi, H\. Bai, L\. Shang, and L\. Kong \(2026\)Dream\-VL & Dream\-VLA: open vision\-language and vision\-language\-action models with diffusion language model backbone\.External Links:2512\.22615,[Link](https://arxiv.org/abs/2512.22615)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1),[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p1.1)\.
- O\. Yoran, T\. Wolfson, O\. Ram, and J\. Berant \(2024\)Making retrieval\-augmented language models robust to irrelevant context\.InThe Twelfth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=ZS4m74kZpH)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px1.p1.1)\.
- Z\. You, S\. Nie, X\. Zhang, J\. Hu, J\. Zhou, Z\. Lu, J\. Wen, and C\. Li \(2025\)LLaDA\-V: large language diffusion models with visual instruction tuning\.External Links:2505\.16933,[Link](https://arxiv.org/abs/2505.16933)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p1.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1),[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p1.1)\.
- C\. Yu, J\. Wang, Y\. Li, H\. Chang, G\. Lan, Q\. Sun, J\. Li, J\. Li, and Z\. Zhang \(2026\)Unlocking the potentials of retrieval\-augmented generation for diffusion language models\.External Links:2601\.11342,[Link](https://arxiv.org/abs/2601.11342)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1),[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px3.p1.1)\.
- S\. Yu, C\. Tang, B\. Xu, J\. Cui, J\. Ran, Y\. Yan, Z\. Liu, S\. Wang, X\. Han, Z\. Liu, and M\. Sun \(2025\)VisRAG: vision\-based retrieval\-augmented generation on multi\-modality documents\.InThe Thirteenth International Conference on Learning Representations,External Links:[Link](https://openreview.net/forum?id=zG459X3Xge)Cited by:[§1](https://arxiv.org/html/2608.07006#S1.p2.1),[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px1.p1.1),[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p1.1)\.
- W\. Yu, H\. Zhang, X\. Pan, P\. Cao, K\. Ma, J\. Li, H\. Wang, and D\. Yu \(2024\)Chain\-of\-note: enhancing robustness in retrieval\-augmented language models\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,Miami, Florida, USA,pp\. 14672–14685\.External Links:[Document](https://dx.doi.org/10.18653/v1/2024.emnlp-main.813),[Link](https://aclanthology.org/2024.emnlp-main.813/)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px1.p1.1)\.
- S\. Zhang, L\. Yu, R\. Brekelmans, L\. Tang, S\. Asif, and G\. Ver Steeg \(2026\)Generation order and parallel decoding in masked diffusion models: an information\-theoretic perspective\.External Links:2602\.00286,[Link](https://arxiv.org/abs/2602.00286)Cited by:[§2](https://arxiv.org/html/2608.07006#S2.SS0.SSS0.Px2.p1.1)\.
- F\. Zhu, W\. Lei, F\. Feng, C\. Wang, H\. Zhang, and T\. Chua \(2022\)Towards complex document understanding by discrete reasoning\.InProceedings of the 30th ACM International Conference on Multimedia,pp\. 4857–4866\.External Links:[Document](https://dx.doi.org/10.1145/3503161.3548422),[Link](https://doi.org/10.1145/3503161.3548422)Cited by:[§5\.1](https://arxiv.org/html/2608.07006#S5.SS1.SSS0.Px1.p2.1)\.