Detect, Remask, Repair: Diffusion Editing for Faithful Summarization of Evolving Contexts

# Diffusion Editing for Faithful Summarization of Evolving Contexts
Source: [https://arxiv.org/html/2606.12807](https://arxiv.org/html/2606.12807)
Hao Zou, Zachary Horvitz, Chandhru Karthick, Zhou Yu, Kathleen McKeown Columbia University New York, NY, USA \{hz2999, zfh2000\}@columbia\.edu

###### Abstract

Summaries of real\-world events can become outdated as contexts evolve and new information arrives\. A common response is to generate a new summary from the updated context, but full regeneration discards the previous draft, can obscure what changed, and may be unnecessary when only a few claims are unsupported\. We studylocalized faithfulness repair: updating outdated spans in an existing summary while preserving supported content\. We proposeDetect–Remask–Repair, a diffusion\-based framework that identifies, remasks, and repairs outdated regions with masked diffusion language models\. To evaluate evolving\-context summarization, we introduceStreamSum, a benchmark of synthetic event timelines\. Experiments on DialogSum and StreamSum show that localized diffusion repair provides a controllable alternative to full rewriting: faithfulness\-steered repair improves early drafts, one\-step repair reduces repair cost to under half a second, with the framework enabling faithfulness–speed–preservation tradeoffs across datasets\. We also find that the framework can provide a post\-hoc correction step that improves faithfulness for autoregressive systems\.

Detect, Remask, Repair: Diffusion Editing for Faithful Summarization of Evolving Contexts

Hao Zou, Zachary Horvitz, Chandhru Karthick, Zhou Yu, Kathleen McKeownColumbia UniversityNew York, NY, USA\{hz2999, zfh2000\}@columbia\.edu

## 1Introduction

![Refer to caption](https://arxiv.org/html/2606.12807v1/figures/figure1_overview.png)Figure 1:Overview ofDetect–Remask–Repair\(DRR\) on a StreamSum example\. In an evolving context, an autoregressive draft generated from earlier updates preserves outdated claims\.DRRdetects unsupported spans, selectively re\-masks them, and utilizes a text diffusion model to infill the masked spans based on later updates, yielding a faithful summary\.Existing work on abstractive summarization generally assumes a static setting, where summaries are generated from fixed source documents\. In contrast, real\-world summarization settings are often dynamic\. In breaking news, financial reports, public\-safety incidents, and meetings, information arrives over time\. Summaries generated from partial contexts may initially be plausible but become stale as later evidence materializes\. Figure[1](https://arxiv.org/html/2606.12807#S1.F1)illustrates this setting: claims in an early summary about lithium\-ion battery fires become unsupported after later updates revise the incident count, affected locations, and death toll\.

A natural solution to evolving contexts is to regenerate an entire summary given each update\. While accurate, full regeneration may alter already\-correct content, obscure changes, and add cost when only a subset of facts change\. This is undesirable when compute is limited, or when summaries are persistent artifacts that may be displayed to users, show updates over time, edited by humans, or consumed by downstream systems\. We therefore studylocalized faithfulness repair: given an existing summary and an updated context, identifying and revising only unsupported spans while otherwise preserving supported content\. Equivalently, this can be viewed as a minimal\-change repair problem: the goal is not to rewrite the best possible summary from scratch, but to make small, fast, and targeted edits needed to restore faithfulness\.

Figure[2](https://arxiv.org/html/2606.12807#S1.F2)shows the tradeoffs studied in this paper\. Because the input is an existing summary, repair methods are evaluated not only by final faithfulness, but also by latency and by how much content they rewrite\. We therefore compare correction strategies along three axes: faithfulness, speed, and preservation of the existing summary\. Depending on the task, a system may prioritize rapid updates, stronger faithfulness, or minimal changes to already\-supported content\.

![Refer to caption](https://arxiv.org/html/2606.12807v1/figures/figure2_operating_profile.png)Figure 2:Performance of correction strategies for early\-draft repair\. Higher is better on all axes\. Faithfulness is measured by AlignScore\(Zha et al\.,[2023](https://arxiv.org/html/2606.12807#bib.bib31)\), speed by inverse repair time in seconds, and preservation by inverse normalized edit distance from the existing summary\. Values are normalized within each dataset for visualization\.To perform localized faithfulness repair, we proposeDetect–Remask–Repair, a diffusion\-based framework for identifying stale spans and repairing them in\-place\. Our method first applies a token\-level detector trained on samples from the diffusion model’s unmasking process to identify incorrect summary tokens; we refer to as\[Mask\]\-Disc\. We then selectively re\-mask those tokens and repair them by generating new tokens using a masked diffusion language model conditioned on the updated context\. This design uses the infilling capability of masked diffusion models: instead of rewriting left\-to\-right, the model regenerates only selected spans while conditioning on both the source and surrounding summary context\.

To evaluate localized faithfulness repair, we also introduceStreamSum, a benchmark for evolving\-context summarization\. StreamSum contains synthetic event timelines seeded by real\-world news, where early summaries become stale after later reports come in, which may overturn preliminary claims, shift responsibility or attribution across actors, or change the status of an event\. Unlike standard static summarization datasets, StreamSum directly evaluates whether systems can accurately repair outdated claims under evolving evidence\.

We carry out experiments on StreamSum, as well as another dataset, DialogSumChen et al\. \([2021](https://arxiv.org/html/2606.12807#bib.bib2)\)which contains summaries of multi\-turn conversations\. Experiments on both datasets show that diffusion repair improves faithfulness ofearly\-context summaries\(i\.e\., summaries generated before the dialog or the event unfolds\)\. Additionally, diffusion repair provides controllable faithfulness–speed–preservation tradeoffs\. On DialogSum, our approach significantly improves AlignScore over both initial early\-context summaries and an autoregressive critique\-and\-revise baselineWadhwa et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib27)\)\. On StreamSum, diffusion repair substantially improves out\-of\-date early\-context summaries\. Beyond early\-context summary repair,Detect–Remask–Repairalso improves full\-generation outputs as a post\-hoc correction layer, increasing AlignScore over full regeneration on both datasets and consistently improving ROUGE\-LLin \([2004](https://arxiv.org/html/2606.12807#bib.bib13)\)across our repair settings\.

Our main contributions are: \(1\) We formulate the task oflocalized faithfulness repairfor evolving\-context summarization, where summaries generated from partial context must be updated as new evidence arrives\. \(2\) We proposeDetect–Remask–Repair, a token\-level detect–remask–repair framework that combines outdated span detection with masked diffusion repair\. The procedure makes the correction process inspectable: users can see which spans were selected, and which examples were routed for repair\. \(3\) As part of our method we propose\[Mask\]\-Disc, a lightweight token\-level detector used for both span selection and budgeted repair routing\. \(4\) We introduceStreamSum, an evolving\-event summarization benchmark constructed from real\-world news seeds using agentic timeline synthesis, AlignScore\-based support filtering, and verifier/pruner agents for revision clarity and diversity\. \(5\) We demonstrate that diffusion repair improves early\-context summaries under a controllable faithfulness–speed–preservation tradeoffs and can further improve full\-generation outputs as a post\-hoc faithfulness corrector\.

## 2Related Work

#### Faithfulness and refinement in summarization\.

Factual inconsistency has been widely studied in abstractive summarization, where generated summaries may include unsupported entities, numbers, relations, or eventsTang et al\. \([2024b](https://arxiv.org/html/2606.12807#bib.bib26)\); Wan et al\. \([2023](https://arxiv.org/html/2606.12807#bib.bib28)\); Maynez et al\. \([2020](https://arxiv.org/html/2606.12807#bib.bib15)\); Pagnoni et al\. \([2021](https://arxiv.org/html/2606.12807#bib.bib20)\); Kryscinski et al\. \([2020](https://arxiv.org/html/2606.12807#bib.bib11)\); Fabbri et al\. \([2021](https://arxiv.org/html/2606.12807#bib.bib3),[2022](https://arxiv.org/html/2606.12807#bib.bib4)\); Wan et al\. \([2025](https://arxiv.org/html/2606.12807#bib.bib29)\)\. Prior work addresses this problem through factuality metrics and detectorsKryściński et al\. \([2019](https://arxiv.org/html/2606.12807#bib.bib10)\); Goyal and Durrett \([2020](https://arxiv.org/html/2606.12807#bib.bib5)\); Scialom et al\. \([2021](https://arxiv.org/html/2606.12807#bib.bib22)\); Zha et al\. \([2023](https://arxiv.org/html/2606.12807#bib.bib31)\); Tang et al\. \([2024a](https://arxiv.org/html/2606.12807#bib.bib25)\), as well as training objectives, reranking, post\-editing, decoding constraints, and critique\-based refinementNan et al\. \([2021](https://arxiv.org/html/2606.12807#bib.bib18)\); Wan et al\. \([2023](https://arxiv.org/html/2606.12807#bib.bib28)\); Madaan et al\. \([2023](https://arxiv.org/html/2606.12807#bib.bib14)\); Wadhwa et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib27)\)\. Our work is closely related to Detect–Critique–Refine \(DCR\)Wadhwa et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib27)\), which decomposes factual correction into detecting problematic sentences, generating natural\-language feedback, and refining the output with an autoregressive model\. Rather than producing sentence\-level critiques and rewriting autoregressively, we use a masked diffusion model to identify and directly edit incorrect tokens\.

#### Diffusion language models and remasking\.

Masked diffusion language models generate text by iteratively unmasking tokens, offering an alternative to left\-to\-right autoregressive generationLi et al\. \([2022](https://arxiv.org/html/2606.12807#bib.bib12)\); Sahoo et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib21)\); Nie et al\. \([2025](https://arxiv.org/html/2606.12807#bib.bib19)\); Zou et al\. \([2023](https://arxiv.org/html/2606.12807#bib.bib33)\)\. Recent work explores remasking to improve diffusion generation:Wang et al\. \([2025](https://arxiv.org/html/2606.12807#bib.bib30)\)introduces inference\-time remasking for pretrained masked diffusion models, whileHuang et al\. \([2025](https://arxiv.org/html/2606.12807#bib.bib8)\)trains diffusion models to identify and remask incorrect tokens using randomly masked or randomly replaced text\. In contrast, we perform targeted remasking using\[Mask\]\-Disc, a classifier trained on model\-refilled summary corruptions which contain fluent but potentially unsupported tokens\.

#### Summarizing evolving contexts\.

Prior work on update summarization, timeline summarization, and incremental summarization studies how to select or generate summaries from evolving document streamsHwang et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib9)\); Aslam et al\. \([2013](https://arxiv.org/html/2606.12807#bib.bib1)\); Hu et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib7)\); McCreadie et al\. \([2014](https://arxiv.org/html/2606.12807#bib.bib16)\); Habernal et al\. \([2016](https://arxiv.org/html/2606.12807#bib.bib6)\)\. In contrast, StreamSum evaluates a complementary setting: repairing an existing summary after later evidence arrives\.

## 3Methods

![Refer to caption](https://arxiv.org/html/2606.12807v1/x1.png)Figure 3:Overview ofDetect–Remask–Repair\. During training \(left\), we mask the source and reference summary, then use a masked diffusion model to refill selected summary positions, producing corrupted summaries that supervise both\[Mask\]\-Discand the one\-step repair model\. At inference \(right\),\[Mask\]\-Discscores draft tokens, high\-staleness positions are re\-masked, and a masked diffusion model infills them using the updated context while preserving supported spans; an aggregate risk score can also route summaries to repair or skip\.We proposeDetect–Remask–Repair, a framework for localized summary repair under evolving contexts\. Given a draft summary produced from partial context, the goal is to identify unsupported spans under the updated context and revise only those regions\. Figure[3](https://arxiv.org/html/2606.12807#S3.F3)summarizes the training and inference pipeline\.

### 3\.1Problem Setup

We usexearlyx\_\{\\mathrm\{early\}\}for the context available when an initial summary is produced andxfullx\_\{\\mathrm\{full\}\}for the updated context used for repair\. Although our experiments use an early/full split, the same formulation can be applied to intermediate updates by treating the currently available accumulated context asxfullx\_\{\\mathrm\{full\}\}\.

In our formulation, we start with an initial summary, based on the initial contextxearlyx\_\{\\mathrm\{early\}\}\. For example, an autoregressive summarizer can produce this initial summaryyearly∼pAR​\(y∣xearly\)y\_\{\\mathrm\{early\}\}\\sim p\_\{\\mathrm\{AR\}\}\(y\\mid x\_\{\\mathrm\{early\}\}\)\. As new evidence arrives, some claims inyearlyy\_\{\\mathrm\{early\}\}may become unsupported with respect toxfullx\_\{\\mathrm\{full\}\}\. The goal of evolving\-context summary repair is to produce a repaired summaryyrep=R​\(xfull,yearly\)y\_\{\\mathrm\{rep\}\}=R\(x\_\{\\mathrm\{full\}\},y\_\{\\mathrm\{early\}\}\)that is faithful to the updated context while preserving supported content from the draft\. Unlike full regeneration, which rewrites a new summary fromxfullx\_\{\\mathrm\{full\}\}, localized repair aims to update only stale spans, such as revised numbers, attributions, or events, while copying supported surrounding phrasing whenever possible\.

### 3\.2Detect–Remask–Repair

Detect–Remask–Repairdecomposes localized repair into three steps\. First, a token\-level detectorDθD\_\{\\theta\}assigns a staleness score to each draft token index,ii:

si=Dθ​\(xfull,yearly,i\),s\_\{i\}=D\_\{\\theta\}\(x\_\{\\mathrm\{full\}\},y\_\{\\mathrm\{early\}\},i\),
where highersis\_\{i\}indicates that tokeniiis more likely to be stale or unsupported\. Second, the system selects a set of high\-staleness positionsMkM\_\{k\}and converts them into mask tokens:

yimask=\{\[MASK\],i∈Mk,yi,otherwise\.y^\{\\mathrm\{mask\}\}\_\{i\}=\\begin\{cases\}\\texttt\{\[MASK\]\},&i\\in M\_\{k\},\\\\ y\_\{i\},&\\text\{otherwise\}\.\\end\{cases\}Optionally, selected tokens are expanded into short spans to enable larger edits\. Third, a repair modelGϕG\_\{\\phi\}fills the masked positions conditioned on the updated context:

yrep=Gϕ​\(xfull,ymask\)\.y\_\{\\mathrm\{rep\}\}=G\_\{\\phi\}\(x\_\{\\mathrm\{full\}\},y^\{\\mathrm\{mask\}\}\)\.All unselected tokens remain fixed\. The full inference procedure is given in Appendix[A\.1](https://arxiv.org/html/2606.12807#A1.SS1)\.

#### Budgeted repair\.

The same detector also supports sample\-level staleness calculation\. For each draft summary, we compute a faithfulness score by averaging the top\-kkstaleness scores:

ρ​\(y\)=1k​∑i∈TopK⁡\(s,k\)si\.\\rho\(y\)=\\frac\{1\}\{k\}\\sum\_\{i\\in\\operatorname\{TopK\}\(s,k\)\}s\_\{i\}\.A budgeted repair policy repairs only the topp%p\\%highest\-risk summaries and skips the rest\. This letsDetect–Remask–Repairavoid over\-editing already faithful summaries and allocate repair computation to examples most likely to benefit\.

### 3\.3Training\[Mask\]\-Discwith Diffusion Unmasking

Masked diffusion models are only trained to iteratively predictmasked tokens\. As a result, we train\[Mask\]\-Disc, a separate lightweight discriminator to predict summary tokens that are incorrect and should be remasked\.

\[Mask\]\-Discis implemented as a lightweight token\-level classifier on top of masked diffusion language model representations, using a linear classification head to predict whether each visible summary token is faithful or stale\. Because human token\-level hallucination labels are expensive, we construct synthetic supervision from diffusion\-style corruptions\. Starting from a faithful context\-summary pair\(x,y0\)\(x,y\_\{0\}\), we sample a noise level and apply it to both the source context and reference summary, producing a masked contextx~\\tilde\{x\}and a partially masked summary\. We then run a masked diffusion language model onx~\\tilde\{x\}concatenated with the partially masked summary, but only refill selected masked summary positions\. This yields a corrupted summaryyty\_\{t\}containing gold tokens, model\-filled tokens, and remaining masks\. Because the refiller conditions on a partially masked context, the filled tokens can be fluent and plausible while still unsupported by the original source\.

We label each visible token by comparing it with the reference:

zi=𝕀​\[yt,i=y0,i\],mi=𝕀​\[yt,i≠\[MASK\]\]\.z\_\{i\}=\\mathbb\{I\}\[y\_\{t,i\}=y\_\{0,i\}\],\\qquad m\_\{i\}=\\mathbb\{I\}\[y\_\{t,i\}\\neq\\texttt\{\[MASK\]\}\]\.Remaining mask tokens are ignored\. The detector predicts

pi=pθ​\(zi=1∣x,yt,i\)p\_\{i\}=p\_\{\\theta\}\(z\_\{i\}=1\\mid x,y\_\{t\},i\)and minimizes token\-level cross entropy:

ℒdisc=−∑i=1Lmi​\[zi​log⁡pi\+\(1−zi\)​log⁡\(1−pi\)\]\.\\mathcal\{L\}\_\{\\mathrm\{disc\}\}=\-\\sum\_\{i=1\}^\{L\}m\_\{i\}\\left\[z\_\{i\}\\log p\_\{i\}\+\(1\-z\_\{i\}\)\\log\(1\-p\_\{i\}\)\\right\]\.We define the token staleness score assi=1−pis\_\{i\}=1\-p\_\{i\}, so higher values indicate tokens more likely to require repair\. At inference time, we use the detector output as the staleness score for re\-masking\.

### 3\.4Repair Models

The repair module receives the updated contextxfullx\_\{\\mathrm\{full\}\}and selectively masked summaryymasky^\{\\mathrm\{mask\}\}, then fills the masked spans while leaving unmasked tokens fixed\.\[Mask\]\-Discdetermines*which*spans are re\-masked and, when budgeted routing is enabled,*which*examples are repaired\. The repair model determines*how*the selected masks are filled\. We instantiate this module with three variants\.

#### Iterative faithfulness\-steered repair\.

Our main iterative variant uses LLaDANie et al\. \([2025](https://arxiv.org/html/2606.12807#bib.bib19)\)to fill the spans selected by\[Mask\]\-Discthrough masked diffusion decoding\. Only the selected spans are allowed to change\. To improve faithfulness, we add FK\-SteeringSinghal et al\. \([2025](https://arxiv.org/html/2606.12807#bib.bib24)\)with BS\-Fact, implemented as BERTScoreZhang et al\. \([2020](https://arxiv.org/html/2606.12807#bib.bib32)\)precision againstxfullx\_\{\\mathrm\{full\}\}, as the source\-grounded reward\. During decoding, particles are scored by comparing estimated clean summaries against the full context and resampled toward higher\-reward trajectories\. Thus,\[Mask\]\-Discsupplies edit locations and routing, while FK\-Steering guides the denoising process after masks are selected\. Further details are in Appendix[A\.3](https://arxiv.org/html/2606.12807#A1.SS3)\.

#### Fast one\-step repair\.

The one\-step variants use the same\[Mask\]\-Discdetection, re\-masking, and optional routing procedure, but replace iterative masked diffusion decoding with a single repair pass\. Given the updated context and selectively masked summary, the one\-step model predicts repaired tokens:

qψ​\(y0∣xfull,ymask\)\.q\_\{\\psi\}\(y\_\{0\}\\mid x\_\{\\mathrm\{full\}\},y^\{\\mathrm\{mask\}\}\)\.The model is trained on the same corrupted\-to\-clean pairs used for repair supervision, with loss applied to selected repair positions\. This compresses localized repair to one model evaluation and yields a low\-latency repair variant\. The full objective is given in Appendix[A\.4](https://arxiv.org/html/2606.12807#A1.SS4)\.

#### DCR\-distilled repair\.

We also build on Detect–Critique–Refine \(DCR\)Wadhwa et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib27)\), a strong autoregressive refinement framework that detects factual errors, generates natural\-language feedback, and revises the summary\. Instead of using DCR directly at inference time, we use its refinements as teacher targets for one\-step masked repair\. This distills a strong autoregressive refiner into a faster masked repair model while preserving the same explicit re\-masking step used byDetect–Remask–Repair\.

## 4Experiments

We evaluate whetherDetect–Remask–Repaircan repair summaries when later evidence changes the support for earlier claims\. Our experiments are designed to answer four questions: \(1\) can localized diffusion repair improve early\-context summaries, \(2\) How does the proposed method compare with autoregressive critique\-and\-revise refinement, \(3\) What quality–efficiency tradeoffs arise from iterative and one\-step repair, and \(4\) Can diffusion repair also serve as a post\-hoc corrector for full\-context summaries? Dataset statistics, timing, and hardware details are provided in Appendix[C](https://arxiv.org/html/2606.12807#A3)and Appendix[D](https://arxiv.org/html/2606.12807#A4)\.

### 4\.1Datasets and Evaluation Settings

#### DialogSum\.

We evaluate on DialogSumChen et al\. \([2021](https://arxiv.org/html/2606.12807#bib.bib2)\), a dialogue summarization benchmark\. Dialogues provide a natural setting for evolving\-context repair: in meetings, interviews, or customer\-support conversations, a system may produce an interim summary before the conversation has ended, and later turns can add details, clarify earlier statements, or change the interpretation of prior claims\. To simulate this setting, we split each dialogue into an early context and a full context\. A LLaMA\-3\-8BMeta AI \([2024](https://arxiv.org/html/2606.12807#bib.bib17)\)autoregressive summarizer first generates a draft summary from the early context\. Repair methods then receive the draft and the full context, and are asked to correct unsupported or incomplete claims while preserving supported content\. This provides a controlled setting for testing localized repair when only part of the source was initially available\.

#### StreamSum\.

We also evaluate onStreamSum, our evolving\-event summarization benchmark\. Each example contains an early context, a full updated context and a gold full\-context summary\. StreamSum covers six broad event domains—politics, business, disasters, international affairs, sports, and science/technology—and targets update patterns where later evidence changes the support for earlier claims, including numeric revisions, attribution changes, status changes, timeline changes, location changes, and outcome reversals\. To construct StreamSum, we use an agentic synthesis pipeline seeded by real\-world articles retrieved from NewsDataHub111[https://www\.newsdatahub\.com/](https://www.newsdatahub.com/)through API calls\. Claude\-based synthesis agents generate revision schemas, instantiate multi\-update event timelines, and write full\-context summaries, while MiniMax\-based verifier and pruning agents filter examples for realism, revision clarity, and diversity\. The pipeline also uses AlignScore prefix trajectories as automatic support signals, retaining examples where the gold summary is weakly supported by early updates but well supported by the full context\. We use the train and validation splits to train\[Mask\]\-Discand one\-step repair models, and evaluate on the held\-out test split\. Additional construction details, prompts, dataset statistics, and representative examples are provided in Appendix[I](https://arxiv.org/html/2606.12807#A9)\.

#### Repair settings\.

We evaluate two settings\. Inearly\-context summary repair,xearlyx\_\{\\mathrm\{early\}\}is the first half of the source context andxfullx\_\{\\mathrm\{full\}\}is the complete source context; systems repair drafts generated fromxearlyx\_\{\\mathrm\{early\}\}usingxfullx\_\{\\mathrm\{full\}\}as evidence\. Inpost\-hoc full\-generation repair, systems receive a summary already generated fromxfullx\_\{\\mathrm\{full\}\}and attempt to correct remaining unsupported spans\. The first setting tests summary generation under evolving evidence, while the second tests whether the sameDetect–Remask–Repairframework can serve as a post\-hoc faithfulness correction layer\.

### 4\.2Baselines and Systems

We compare against three main baselines\.AR Draftis generated by LLaMA\-3\-8B from the early context and measures the quality of summaries before later evidence is available\.Full Regenerationprompts LLaMA\-3\-8B to fully generate an entirely new summary directly from the full context\.DCRis an autoregressive detect–critique–refine baselineWadhwa et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib27)\), which detects errors, generates natural\-language feedback, and refines the summary using a fine\-tuned autoregressive model\.

We evaluate three repair models withinDetect–Remask–Repair:LLaDA\-Steering,LLaDA\-1step, andLLaDA\-1step Distill\. All three use\[Mask\]\-Discfor span selection and optional example\-level routing; they differ in the repair model used to fill the selected masks\. LLaDA\-Steering uses iterative masked diffusion repair with FK\-Steering and a BS\-Fact source\-support reward computed againstxfullx\_\{\\mathrm\{full\}\}\. The one\-step variants use a single\-pass masked repair model, with the distilled version trained from DCR teacher refinements\.

### 4\.3Evaluation Metrics

We report ROUGE\-LLin \([2004](https://arxiv.org/html/2606.12807#bib.bib13)\)and BLEURTSellam et al\. \([2020](https://arxiv.org/html/2606.12807#bib.bib23)\)for summary quality, and AlignScoreZha et al\. \([2023](https://arxiv.org/html/2606.12807#bib.bib31)\)as our primary automatic faithfulness metric\. Unless otherwise stated, AlignScore is always computed against the full context, including for AR drafts generated from early context\. To measure preservation, we report normalized token edit distance from the input draft; lower values indicate fewer changes to the existing summary\. Efficiency is measured by the number of function evaluations \(NFE\) and average wall\-clock time per example\. We assess statistical significance for AlignScore using paired bootstrap resampling over examples with 10,000 resamples\. Unless otherwise noted, repair time reports additional refinement cost\. Metrics details are provided in Appendix[E](https://arxiv.org/html/2606.12807#A5)\.

### 4\.4Implementation Details

We provide full hyperparameters in Appendix[B](https://arxiv.org/html/2606.12807#A2)\. Unless otherwise specified, iterative repair uses 32 denoising steps and 4 particles\. For DialogSum, we use a conservative budgeted repair by default, repairing the highest\-risk 25% of examples; for StreamSum, we use the full repair budget\. One\-step repair models are trained with LoRA adapters on diffusion\-style corrupted summaries\.

## 5Results and Analysis

We evaluateDetect–Remask–Repairalong three axes: faithfulness, speed, and preservation\. Early\-context summary repair refers to repairing a summary generated from the first half of the input using the full context as evidence\.

### 5\.1Local Repair of Early Drafts

For the main table, we report the default repair budget suggested by the budgeted analysis in Table[5](https://arxiv.org/html/2606.12807#A6.T5): conservative repair routing for DialogSum, where many first\-half drafts are already faithful to the full dialogue, and full\-budget repair for StreamSum\.

Table 1:Main results for early\-draft repair\. Rows shaded in gray are our diffusion repair variants\.†\\daggerindicates significantly better AlignScore than DCR, and‡\\ddaggerindicates significantly better AlignScore than the AR draft, under paired bootstrap resampling \(p<0\.05p<0\.05\)\. Time for repair methods reports additional repair/refinement cost averaged over all examples, including examples that are routed to skip repair\.Table[1](https://arxiv.org/html/2606.12807#S5.T1)shows two different repair regimes\. On DialogSum, the AR draft generated from the first half of the dialogue is already relatively faithful to the full dialogue, with an AlignScore of 0\.8513\. LLaDA\-Steering with conservative repair routing improves AlignScore to 0\.8790, a statistically significant improvement over both the AR draft and DCR under paired bootstrap resampling\. This result suggests that the detector can allocate repair computation to examples that benefit from local correction while avoiding unnecessary edits to already faithful summaries\.

StreamSum exhibits a different pattern\. The AR draft generated from the first half of the evolving event context has much lower AlignScore against the full context, 0\.5432\. LLaDA\-Steering significantly improves AlignScore to 0\.6895 under paired bootstrap resampling\. The one\-step variants provide a lower\-latency alternative: LLaDA\-1step reaches 0\.6345 and LLaDA\-1step Distill reaches 0\.6377, while requiring only 0\.31–0\.43 seconds of total repair time\. DCR remains the strongest local repair baseline on StreamSum, reaching 0\.7174, but requires substantially more refinement computation: 9\.46 seconds compared with 7\.69 seconds for LLaDA\-Steering and 0\.31–0\.43 seconds for the one\-step variants\. Full regeneration from the complete context achieves the highest score on StreamSum with inference time 7\.73 seconds\.

Table 2:Preservation under early\-draft repair, measured as normalized token edit distance from the AR draft; lower is better\.#### Preservation\.

Table[2](https://arxiv.org/html/2606.12807#S5.T2)reports normalized token edit distance from the early\-context summary\. Full regeneration rewrites substantially more of the draft than repair methods on both datasets\. LLaDA\-Steering achieves the best DialogSum AlignScore while changing only 9\.1% of the draft; on StreamSum, diffusion repair provides faster targeted corrections that preserve more text, while full regeneration attains higher faithfulness through larger edits\. These results illustrate the faithfulness–speed–preservation tradeoff in Figure[2](https://arxiv.org/html/2606.12807#S1.F2)\.

Table 3:Post\-hoc repair of full\-generation outputs\.‡\\ddaggerindicates significantly better AlignScore than the full\-generation input under paired bootstrap resampling \(p<0\.05p<0\.05\)\. NFE and time for diffusion repair report additional repair cost averaged over all examples\.

### 5\.2Effect of Budgeted Repair

\[Mask\]\-Disccan also be used as a sample\-level router: examples are ranked by aggregate token staleness, and only the topp%p\\%highest\-risk examples are repaired\. This creates a controllable budgeted\-repair setting, where the system can trade off faithfulness, preservation, and computation\. Appendix[F](https://arxiv.org/html/2606.12807#A6)reports the full Top\-25/50/75/All results and timing breakdowns\.

The trends differ across datasets\. On DialogSum, conservative routing is most effective: LLaDA\-Steering reaches 0\.8790 AlignScore when repairing only the top 25% highest\-risk examples, but drops to 0\.8584 when repairing all examples\. This suggests that many first\-half dialogue summaries are already compatible with the full dialogue, so unnecessary edits can hurt faithfulness\. On StreamSum, repair improves as the budget increases: LLaDA\-Steering rises from 0\.5856 at Top\-25 to 0\.6895 when all examples are repaired\. This matches the benchmark design, where later event updates more often make early summaries outdated\. Together, these results show that\[Mask\]\-Discprovides both token\-level edit targets and a coarse sample\-level control signal for balancing preservation, faithfulness, and computation\.

### 5\.3Post\-hoc Repair of Full Generations

Table[3](https://arxiv.org/html/2606.12807#S5.T3)evaluates whether localized repair remains useful after full\-context generation\. Unlike early\-context repair, the input summary has already been generated from the full context, so repair targets remaining unsupported spans rather than newly available evidence\.

On DialogSum, LLaDA\-Steering improves full\-generation AlignScore from 0\.8258 to 0\.8355, a statistically significant gain under paired bootstrap resampling\. The difference from DCR’s AlignScore of 0\.8378 is not statistically significant\. Gains are modest because full\-generation outputs are already strong, but the result shows that token\-level detection and masked diffusion repair can still find useful local corrections after full\-summary generation\.

The post\-hoc results also clarify the roles of our repair variants\. Faithfulness\-steered repair explicitly optimizes a source\-grounded reward during editing, and is the most reliable post\-hoc corrector\. The one\-step variants are much faster and consistently improve ROUGE\-L over full generation, but they are less reliable for AlignScore in this setting\.

### 5\.4Efficient and Inspectable Repair

The results show a tradeoff between faithfulness, computation, and the amount of text that must be rewritten\. DCR is a strong autoregressive refinement baseline, but it relies on natural\-language feedback and autoregressive rewriting\.

In contrast,Detect–Remask–Repairoperates with in\-place token\-level operations:\[Mask\]\-Discscores each summary token, selected spans are re\-masked, and the repair model fills only those masks\. This makes repair more inspectable: users can see which spans were selected, how many examples were routed for repair, and how changing the repair budget affects both faithfulness and cost\. Representative repair examples with edited spans highlighted are shown in Appendix[H](https://arxiv.org/html/2606.12807#A8)\.

This also distinguishes\[Mask\]\-Discfrom existing faithfulness metrics\. Metrics such as AlignScore are useful for evaluation, but they are not designed to provide token\-level edit targets or to run as a lightweight routing module inside a repair pipeline\.\[Mask\]\-Discis trained as a simple token\-level classifier on diffusion\-style corruptions, so at inference time it can cheaply assign staleness scores to summary tokens and aggregate them into a sample\-level repair score\.

The one\-step variants provide the strongest speed advantage, reducing repair to a single model evaluation\. Iterative faithfulness\-steered repair is slower, but provides stronger faithfulness control\. Together, these results suggest a practical workflow: use\[Mask\]\-Discto estimate repair risk, choose a repair budget based on the desired cost–preservation tradeoff, and apply either fast one\-step repair or iterative steering depending on resource constraints\.

Overall, these results position localized diffusion repair as a promising editing tool\. When existing summaries are mostly reusable, as in DialogSum,Detect–Remask–Repaircan improve faithfulness while preserving supported text and limiting repair cost\. When later updates usually require rewriting, as in StreamSum, full regeneration remains the strongest faithfulness baseline\. In such cases, localized repair is still useful when transparent edit locations, preservation, or low\-latency correction are part of the task specification\.

## 6Conclusion

We introducedDetect–Remask–Repair, a diffusion\-based framework for localized faithfulness repair\. When context changes, instead of regenerating an entire summary,Detect–Remask–Repairestimates which text to repair, identifies stale tokens, selectively re\-masks them, and repairs the masked spans with masked diffusion language models\. We also introduced StreamSum, an evolving\-context summarization benchmark where early summaries become stale as later updates arrive\. Our results show that localized diffusion repair can provide an efficient, controllable, and interpretable approach that can repair factual inconsistencies and complement autoregressive methods\.

## Limitations

Our evaluation relies on automatic metrics such as AlignScore, ROUGE\-L, and BLEURT, which are useful but imperfect proxies for faithfulness and summary quality\. StreamSum is synthetic, although seeded by real\-world news, and future work should evaluate localized repair on naturally occurring evolving\-summary data\. Finally, while\[Mask\]\-Discprovides token\-level edit candidates, the current system uses fixed edit budgets; adaptive or uncertainty\-aware edit selection may improve robustness\.

## Ethical Considerations

This work studies faithfulness repair for summaries in evolving contexts\. Improving factual consistency can reduce the risk that users rely on stale or unsupported claims, but automatic repair systems should not be treated as a substitute for human verification in high\-stakes domains\. Our method exposes selected edit locations and repairs them with a masked diffusion model, but it may still miss unsupported claims or introduce new errors\. Any deployment of summarization repair systems should include transparency about model\-generated edits, careful monitoring for factual errors, and human oversight where summaries may affect real\-world decisions\.

## References

- Aslam et al\. \(2013\)Javed Aslam, Fernando Diaz, Matthew Ekstrand\-Abueg, Virgiliu Pavlu, and Tetsuya Sakai\. 2013\.Overview of the trec 2013 temporal summarization track\.In*Proceedings of the Twenty\-Second Text REtrieval Conference \(TREC 2013\)*\. NIST\.
- Chen et al\. \(2021\)Yulong Chen, Yang Liu, Liang Chen, and Yue Zhang\. 2021\.[DialogSum: A real\-life scenario dialogue summarization dataset](https://doi.org/10.18653/v1/2021.findings-acl.449)\.In*Findings of the Association for Computational Linguistics: ACL\-IJCNLP 2021*, pages 5062–5074, Online\. Association for Computational Linguistics\.
- Fabbri et al\. \(2021\)Alexander R\. Fabbri, Wojciech Kryściński, Bryan McCann, Caiming Xiong, Richard Socher, and Dragomir Radev\. 2021\.[Summeval: Re\-evaluating summarization evaluation](https://arxiv.org/abs/2007.12626)\.*Preprint*, arXiv:2007\.12626\.
- Fabbri et al\. \(2022\)Alexander R\. Fabbri, Chien\-Sheng Wu, Wenhao Liu, and Caiming Xiong\. 2022\.[Qafacteval: Improved qa\-based factual consistency evaluation for summarization](https://arxiv.org/abs/2112.08542)\.*Preprint*, arXiv:2112\.08542\.
- Goyal and Durrett \(2020\)Tanya Goyal and Greg Durrett\. 2020\.Evaluating factuality in generation with dependency\-level entailment\.In*Findings of the Association for Computational Linguistics: EMNLP 2020*\.
- Habernal et al\. \(2016\)Ivan Habernal, Steffen Eger, and Iryna Gurevych\. 2016\.[Sequential clustering and contextual importance measures for incremental update summarization](https://aclanthology.org/C16-1102/)\.In*Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers*, pages 1042–1053\.
- Hu et al\. \(2024\)Qisheng Hu, Geonsik Moon, and Hwee Tou Ng\. 2024\.[From moments to milestones: Incremental timeline summarization leveraging large language models](https://doi.org/10.18653/v1/2024.acl-long.390)\.In*Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pages 7232–7246, Bangkok, Thailand\. Association for Computational Linguistics\.
- Huang et al\. \(2025\)Zemin Huang, Yuhang Wang, Zhiyang Chen, and Guo\-Jun Qi\. 2025\.Don’t settle too early: Self\-reflective remasking for diffusion language models\.*arXiv preprint arXiv:2509\.23653*\.
- Hwang et al\. \(2024\)EunJeong Hwang, Yichao Zhou, James Bradley Wendt, Beliz Gunel, Nguyen Vo, Jing Xie, and Sandeep Tata\. 2024\.[Enhancing incremental summarization with structured representations](https://doi.org/10.18653/v1/2024.findings-emnlp.220)\.In*Findings of the Association for Computational Linguistics: EMNLP 2024*, pages 3830–3842, Miami, Florida, USA\. Association for Computational Linguistics\.
- Kryściński et al\. \(2019\)Wojciech Kryściński, Bryan McCann, Caiming Xiong, and Richard Socher\. 2019\.Evaluating the factual consistency of abstractive text summarization\.*arXiv preprint arXiv:1910\.12840*\.
- Kryscinski et al\. \(2020\)Wojciech Kryscinski, Bryan McCann, Caiming Xiong, and Richard Socher\. 2020\.[Evaluating the factual consistency of abstractive text summarization](https://doi.org/10.18653/v1/2020.emnlp-main.750)\.In*Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing \(EMNLP\)*, pages 9332–9346, Online\. Association for Computational Linguistics\.
- Li et al\. \(2022\)Xiang Lisa Li, John Thickstun, Ishaan Gulrajani, Percy Liang, and Tatsunori Hashimoto\. 2022\.[Diffusion\-lm improves controllable text generation](https://api.semanticscholar.org/CorpusID:249192356)\.*ArXiv*, abs/2205\.14217\.
- Lin \(2004\)Chin\-Yew Lin\. 2004\.[ROUGE: A package for automatic evaluation of summaries](https://aclanthology.org/W04-1013/)\.In*Text Summarization Branches Out*, pages 74–81, Barcelona, Spain\. Association for Computational Linguistics\.
- Madaan et al\. \(2023\)Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, Shashank Gupta, Bodhisattwa Prasad Majumder, Katherine Hermann, Sean Welleck, Amir Yazdanbakhsh, and Peter Clark\. 2023\.[Self\-refine: Iterative refinement with self\-feedback](https://arxiv.org/abs/2303.17651)\.*Preprint*, arXiv:2303\.17651\.
- Maynez et al\. \(2020\)Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald\. 2020\.[On faithfulness and factuality in abstractive summarization](https://doi.org/10.18653/v1/2020.acl-main.173)\.In*Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics*, pages 1906–1919, Online\. Association for Computational Linguistics\.
- McCreadie et al\. \(2014\)Richard McCreadie, Craig Macdonald, and Iadh Ounis\. 2014\.[Incremental update summarization: Adaptive sentence selection based on prevalence and novelty](https://www.dcs.gla.ac.uk/~richardm/papers/mccreadie2014_IUS.pdf)\.In*Proceedings of the 23rd ACM International Conference on Conference on Information and Knowledge Management \(CIKM\)*, pages 301–310\.
- Meta AI \(2024\)Meta AI\. 2024\.[Introducing meta llama 3: The most capable openly available llm to date](https://ai.meta.com/blog/meta-llama-3)\.
- Nan et al\. \(2021\)Feng Nan, Ramesh Nallapati, Zhiguo Wang, Cicero Nogueira dos Santos, Henghui Zhu, Dejiao Zhang, Kathleen McKeown, and Bing Xiang\. 2021\.[Entity\-level factual consistency of abstractive text summarization](https://arxiv.org/abs/2102.09130)\.*Preprint*, arXiv:2102\.09130\.
- Nie et al\. \(2025\)Shen Nie, Fengqi Zhu, Zebin You, Xiaolu Zhang, Jingyang Ou, Jun Hu, Jun Zhou, Yankai Lin, Ji\-Rong Wen, and Chongxuan Li\. 2025\.Large language diffusion models\.*arXiv preprint arXiv:2502\.09992*\.
- Pagnoni et al\. \(2021\)Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov\. 2021\.[Understanding factuality in abstractive summarization with frank: A benchmark for factuality metrics](https://arxiv.org/abs/2104.13346)\.*Preprint*, arXiv:2104\.13346\.
- Sahoo et al\. \(2024\)Subham Sekhar Sahoo, Marianne Arriola, Yair Schiff, Aaron Gokaslan, Edgar Marroquin, Justin T Chiu, Alexander Rush, and Volodymyr Kuleshov\. 2024\.[Simple and effective masked diffusion language models](https://api.semanticscholar.org/CorpusID:270380319)\.*ArXiv*, abs/2406\.07524\.
- Scialom et al\. \(2021\)Thomas Scialom, Paul\-Alexis Dray, Patrick Gallinari, Sylvain Lamprier, Benjamin Piwowarski, Jacopo Staiano, and Alex Wang\. 2021\.[Questeval: Summarization asks for fact\-based evaluation](https://arxiv.org/abs/2103.12693)\.*Preprint*, arXiv:2103\.12693\.
- Sellam et al\. \(2020\)Thibault Sellam, Dipanjan Das, and Ankur Parikh\. 2020\.[BLEURT: Learning robust metrics for text generation](https://doi.org/10.18653/v1/2020.acl-main.704)\.In*Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics*, pages 7881–7892, Online\. Association for Computational Linguistics\.
- Singhal et al\. \(2025\)Raghav Singhal, Zachary Horvitz, Ryan Teehan, Mengye Ren, Zhou Yu, Kathleen McKeown, and R V Ranganath\. 2025\.[A general framework for inference\-time scaling and steering of diffusion models](https://api.semanticscholar.org/CorpusID:275470889)\.*ArXiv*, abs/2501\.06848\.
- Tang et al\. \(2024a\)Liyan Tang, Philippe Laban, and Greg Durrett\. 2024a\.[Minicheck: Efficient fact\-checking of llms on grounding documents](https://arxiv.org/pdf/2404.10774)\.In*Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing*\. Association for Computational Linguistics\.
- Tang et al\. \(2024b\)Liyan Tang, Igor Shalyminov, Amy Wing mei Wong, Jon Burnsky, Jake W\. Vincent, Yu’an Yang, Siffi Singh, Song Feng, Hwanjun Song, Hang Su, Lijia Sun, Yi Zhang, Saab Mansour, and Kathleen McKeown\. 2024b\.[Tofueval: Evaluating hallucinations of llms on topic\-focused dialogue summarization](https://arxiv.org/abs/2402.13249)\.*Preprint*, arXiv:2402\.13249\.
- Wadhwa et al\. \(2024\)Manya Wadhwa, Xinyu Zhao, Junyi Jessy Li, and Greg Durrett\. 2024\.[Learning to refine with fine\-grained natural language feedback](https://api.semanticscholar.org/CorpusID:270878552)\.*ArXiv*, abs/2407\.02397\.
- Wan et al\. \(2023\)David Wan, Mengwen Liu, Kathleen McKeown, Markus Dreyer, and Mohit Bansal\. 2023\.[Faithfulness\-aware decoding strategies for abstractive summarization](https://arxiv.org/abs/2303.03278)\.*Preprint*, arXiv:2303\.03278\.
- Wan et al\. \(2025\)David Wan, Jesse Vig, Mohit Bansal, and Shafiq Joty\. 2025\.[On positional bias of faithfulness for long\-form summarization](https://arxiv.org/abs/2410.23609)\.*Preprint*, arXiv:2410\.23609\.
- Wang et al\. \(2025\)Guanghan Wang, Yair Schiff, Subham Sahoo, and Volodymyr Kuleshov\. 2025\.Remasking discrete diffusion models with inference\-time scaling\.*arXiv preprint arXiv:2503\.00307*\.
- Zha et al\. \(2023\)Yuheng Zha, Yichi Yang, Ruichen Li, and Zhiting Hu\. 2023\.[AlignScore: Evaluating factual consistency with a unified alignment function](https://aclanthology.org/2023.acl-long.634)\.In*Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\)*, pages 11328–11348, Toronto, Canada\. Association for Computational Linguistics\.
- Zhang et al\. \(2020\)Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q\. Weinberger, and Yoav Artzi\. 2020\.[Bertscore: Evaluating text generation with bert](https://arxiv.org/abs/1904.09675)\.*Preprint*, arXiv:1904\.09675\.
- Zou et al\. \(2023\)Hao Zou, Zae Myung Kim, and Dongyeop Kang\. 2023\.[A survey of diffusion models in natural language processing](https://arxiv.org/abs/2305.14671)\.*Preprint*, arXiv:2305\.14671\.

## Appendix AAdditional Method Details

### A\.1Inference Algorithm

Algorithm[1](https://arxiv.org/html/2606.12807#alg1)gives the full inference procedure forDetect–Remask–Repair\. At each round,\[Mask\]\-Discscores draft tokens, the highest\-staleness positions are selected and re\-masked, and the repair model infills the selected spans\. The edit schedule𝒦=\(k1,…,kR\)\\mathcal\{K\}=\(k\_\{1\},\\ldots,k\_\{R\}\)controls how many tokens are edited at each round\.

Algorithm 1Detect–Remask–RepairInference1:Updated context

xx, draft summary

y\(0\)y^\{\(0\)\}, detector

DθD\_\{\\theta\}, repair model

GϕG\_\{\\phi\}, edit schedule

𝒦=\(k1,…,kR\)\\mathcal\{K\}=\(k\_\{1\},\\ldots,k\_\{R\}\)
2:Repaired summary

y^\\hat\{y\}
3:

y←y\(0\)y\\leftarrow y^\{\(0\)\}
4:for

r=1,…,Rr=1,\\ldots,Rdo

5:

s←Dθ​\(x,y\)s\\leftarrow D\_\{\\theta\}\(x,y\)⊳\\trianglerightDetect suspicious tokens

6:

Mr←SelectEdits​\(s,kr\)M\_\{r\}\\leftarrow\\textsc\{SelectEdits\}\(s,k\_\{r\}\)
7:if

Mr=∅M\_\{r\}=\\emptysetthen

8:break

9:endif

10:

ymask←ReMask​\(y,Mr\)y^\{\\mathrm\{mask\}\}\\leftarrow\\textsc\{ReMask\}\(y,M\_\{r\}\)⊳\\trianglerightRe\-mask selected spans

11:

y~←Gϕ​\(x,ymask\)\\tilde\{y\}\\leftarrow G\_\{\\phi\}\(x,y^\{\\mathrm\{mask\}\}\)⊳\\trianglerightRepair masked spans

12:

y←Merge​\(ymask,y~,Mr\)y\\leftarrow\\textsc\{Merge\}\(y^\{\\mathrm\{mask\}\},\\tilde\{y\},M\_\{r\}\)
13:endfor

14:return

yy

### A\.2Diffusion\-Style Corruption Details

To train\[Mask\]\-Discand the repair models, we construct corrupted summaries that resemble intermediate states of masked diffusion decoding\. Given a reference summaryy0y\_\{0\}, we sample a corruption level and mask a subset of summary positionsAtA\_\{t\}\. A masked diffusion language model refills a subsetBt⊆AtB\_\{t\}\\subseteq A\_\{t\}, leaving the remaining positions masked\. The resulting sequenceyty\_\{t\}contains three token types: original gold tokens, model\-filled tokens, and unfilled\[MASK\]tokens\.

Visible tokens are labeled by exact match with the reference:

zi=𝕀​\[yt,i=y0,i\],mi=𝕀​\[yt,i≠\[MASK\]\]\.z\_\{i\}=\\mathbb\{I\}\[y\_\{t,i\}=y\_\{0,i\}\],\\qquad m\_\{i\}=\\mathbb\{I\}\[y\_\{t,i\}\\neq\\texttt\{\[MASK\]\}\]\.Tokens that remain masked are excluded from the detector loss\. In practice, we sample multiple corruption levels to expose the detector to easy and difficult states, ranging from lightly corrupted summaries to highly masked summaries\. This creates fluent negative examples without requiring human token\-level annotations\.

### A\.3Faithfulness\-Steered Diffusion Repair

Letc=\(xfull,ymask\)c=\(x\_\{\\mathrm\{full\}\},y^\{\\mathrm\{mask\}\}\)denote the repair condition\. A discrete masked diffusion repair process defines

pϕ​\(y0:T∣c\)=pprior​\(yT\)​∏t=T1pϕ​\(yt−1∣yt,c\),p\_\{\\phi\}\(y\_\{0:T\}\\mid c\)=p\_\{\\mathrm\{prior\}\}\(y\_\{T\}\)\\prod\_\{t=T\}^\{1\}p\_\{\\phi\}\(y\_\{t\-1\}\\mid y\_\{t\},c\),whereyTy\_\{T\}contains the selected masks andy0y\_\{0\}is the repaired summary\. During repair, unmasked draft tokens are fixed and only selected masked positions are updated\.

To encourage faithful repairs, we adapt FK\-SteeringSinghal et al\. \([2025](https://arxiv.org/html/2606.12807#bib.bib24)\)\. Letr​\(y,xfull\)r\(y,x\_\{\\mathrm\{full\}\}\)be a source\-grounded faithfulness reward\. We target an exponentially tilted distribution:

ptarget​\(y0∣xfull\)∝pϕ​\(y0∣c\)​exp⁡\(λ​r​\(y0,xfull\)\)\.p\_\{\\mathrm\{target\}\}\(y\_\{0\}\\mid x\_\{\\mathrm\{full\}\}\)\\propto p\_\{\\phi\}\(y\_\{0\}\\mid c\)\\exp\(\\lambda r\(y\_\{0\},x\_\{\\mathrm\{full\}\}\)\)\.Since direct sampling from this distribution is intractable, we approximate it with particle\-based decoding\. At selected denoising steps, each particle forms an estimated clean summaryy^0\\hat\{y\}\_\{0\}by filling remaining masks, evaluatesr​\(y^0,xfull\)r\(\\hat\{y\}\_\{0\},x\_\{\\mathrm\{full\}\}\), and resamples particles according to the resulting potentials\. This shifts probability mass toward denoising trajectories that yield more source\-supported repairs\.

### A\.4One\-Step and DCR\-Distilled Repair Objectives

The one\-step repair model predicts clean summary tokens from a corrupted summary in a single forward pass:

qψ​\(y0∣xfull,yt\)\.q\_\{\\psi\}\(y\_\{0\}\\mid x\_\{\\mathrm\{full\}\},y\_\{t\}\)\.We train it with token\-level cross entropy over supervised repair positions:

ℒ1​s​t​e​p=−∑i=1Lai​log⁡qψ​\(y0,i∣xfull,yt,i\),\\mathcal\{L\}\_\{\\mathrm\{1step\}\}=\-\\sum\_\{i=1\}^\{L\}a\_\{i\}\\log q\_\{\\psi\}\(y\_\{0,i\}\\mid x\_\{\\mathrm\{full\}\},y\_\{t\},i\),whereaia\_\{i\}indicates the positions included in the repair loss\. In our main setting,ai=1a\_\{i\}=1for masked positions and0otherwise, so the model learns to fill selected repair spans while preserving visible tokens\.

For DCR\-distilled repair, we use teacher refinementsyDCRy\_\{\\mathrm\{DCR\}\}from Detect–Critique–RefineWadhwa et al\. \([2024](https://arxiv.org/html/2606.12807#bib.bib27)\)\. The one\-step repair model is trained to imitate these targets:

ℒdistill=−∑iai​log⁡qψ​\(yDCR,i∣xfull,yt,i\)\.\\mathcal\{L\}\_\{\\mathrm\{distill\}\}=\-\\sum\_\{i\}a\_\{i\}\\log q\_\{\\psi\}\(y\_\{\\mathrm\{DCR\},i\}\\mid x\_\{\\mathrm\{full\}\},y\_\{t\},i\)\.This provides a fast repair model trained from a stronger autoregressive refinement teacher\.

## Appendix BExperimental Details

#### Data construction for\[Mask\]\-Disc\.

For\[Mask\]\-Disctraining, we generate diffusion\-style corrupted summaries from reference context–summary pairs\. We sample a noise level, mask both the source context and the reference summary, and run LLaDA on the masked context concatenated with the masked summary\. Only masked summary positions are editable during refill; masked source tokens remain part of the noisy conditioning context\. This produces partial summaries containing gold tokens, model\-refilled tokens, and remaining masks\. Visible summary tokens are labeled as correct or incorrect by exact match against the reference summary, while remaining masks are ignored\.

#### Corruption hyperparameters\.

We sample denoising steps from\{8,16,32\}\\\{8,16,32\\\}, use a maximum step cap of 64, and sample fill fractions from\{0\.25,0\.5,0\.75\}\\\{0\.25,0\.5,0\.75\\\}\. We require each corrupted example to contain at least one visible token and one remaining mask\. The detector is trained as a token\-level classifier and selected by validation F1 on incorrect\-token detection\.

#### Repair hyperparameters\.

Unless otherwise specified, iterative repair uses 32 denoising steps, 4 particles, and steering weightλ=6\.0\\lambda=6\.0\. For DialogSum, the default token edit budget is 8 for DISC\-gated repair\. For StreamSum, the default token edit budget is 64 because summaries are longer and evolving\-event updates often require correcting multiple related facts\. One\-step repair models are trained with LoRA adapters, supervising masked summary positions unless otherwise stated\.

#### Budgeted DISC\-gating\.

For budgeted repair, we rank examples by the mean of their top\-kk\[Mask\]\-Discsuspiciousness scores\. We then repair only the topp%p\\%highest\-risk examples and keep the remaining drafts unchanged\. DialogSum usesk=8k=8, while StreamSum usesk=64k=64, matching the token\-level edit budgets used by the corresponding repair pipelines\.

#### Timing\.

For early\-draft repair, reported end\-to\-end time includes the initial AR draft generation plus the additional repair or refinement time\. For DISC\-gated settings, time is averaged over all examples, including skipped examples\. For post\-hoc full\-generation repair, total pipeline time is the full\-generation time plus the additional correction time\.

## Appendix CDataset Statistics

Table 4:Dataset statistics\. Lengths are average token counts using simple tokenization\. For StreamSum, Early Len and Full Len correspond to the early and full evolving contexts; DialogSum uses the full dialogue\.Table[4](https://arxiv.org/html/2606.12807#A3.T4)reports dataset sizes and average lengths\.

## Appendix DTiming and Hardware Details

All timing experiments were run on NVIDIA RTX A6000 GPUs with 48GB memory\. The machine used NVIDIA driver version 595\.71\.05 and CUDA 13\.2\. Timing reports wall\-clock seconds per example averaged over the evaluated split\. For repair methods, unless otherwise stated, reported time is additional repair/refinement cost and includes examples skipped by budgeted routing\. For autoregressive LLaMA\-based generation and refinement baselines, we use KV caching during decoding\. Masked diffusion repair does not use autoregressive KV caching\.

## Appendix EEvaluation Details

#### Metrics\.

We evaluate summary quality with ROUGE\-L and BLEURT, and evaluate faithfulness with AlignScore\. ROUGE\-L measures lexical overlap with the reference summary, while BLEURT measures learned semantic similarity to the reference\. AlignScore measures whether a generated summary is supported by the corresponding source context, and is used as our primary automatic faithfulness metric\. We follow the official setup and instructions for all metrics\. For ROUGE\-L, we use the ROUGE package from Google Research\.222[https://github\.com/google\-research/google\-research/tree/master/rouge](https://github.com/google-research/google-research/tree/master/rouge)For BS and BS\-Fact, we use the default English model, RoBERTa\-Large\. All reported model scores are from single runs\.

#### Preservation metric\.

We measure preservation using normalized token edit distance from the AR draft\. Given tokenized draftyary\_\{\\mathrm\{ar\}\}and outputy^\\hat\{y\}, we compute

EditDist​\(yar,y^\)=Lev​\(yar,y^\)max⁡\(\|yar\|,\|y^\|\)\.\\mathrm\{EditDist\}\(y\_\{\\mathrm\{ar\}\},\\hat\{y\}\)=\\frac\{\\mathrm\{Lev\}\(y\_\{\\mathrm\{ar\}\},\\hat\{y\}\)\}\{\\max\(\|y\_\{\\mathrm\{ar\}\}\|,\|\\hat\{y\}\|\)\}\.Lower values indicate stronger preservation of the original draft\.

#### Statistical significance\.

We assess statistical significance for AlignScore using paired bootstrap resampling over test examples with 10,000 resamples\. Since systems are evaluated on the same examples, we compare paired per\-example AlignScore differences\. For each bootstrap resample, we sample examples with replacement and compute the mean difference between two systems\. A difference is considered significant when the 95% bootstrap confidence interval excludes zero, corresponding top<0\.05p<0\.05\.

#### Efficiency measurement\.

We report the number of function evaluations \(NFE\) and average wall\-clock time per example\. For early\-draft repair, end\-to\-end time includes early\-context AR draft generation plus additional repair or refinement time\. For DISC\-gated repair, time is averaged over all examples, counting skipped examples as zero additional repair cost\. For post\-hoc full\-generation repair, total time includes full\-context generation plus additional correction time\.

## Appendix FBudgeted Repair Results

Table[5](https://arxiv.org/html/2606.12807#A6.T5)reports AlignScore as we vary the repair budget\. We rank examples by aggregate\[Mask\]\-Discstaleness score, repair only the topp%p\\%highest\-risk examples, and leave the remaining examples unchanged\. We reportp∈\{25,50,75,100\}p\\in\\\{25,50,75,100\\\}\. This evaluates whether\[Mask\]\-Disccan serve as a sample\-level routing signal in addition to selecting token\-level edit targets\.

As expected, repair time increases with the budget: for early\-draft LLaDA\-Steering, repair time grows from 0\.89s to 3\.54s on DialogSum and from 1\.89s to 7\.69s on StreamSum asppincreases from Top\-25 to All\. The one\-step variants remain much cheaper across budgets, staying below 0\.18s on DialogSum and below 0\.44s on StreamSum\. Full timing breakdowns are provided in Table[7](https://arxiv.org/html/2606.12807#A7.T7)\.

Table 5:Budgeted repair\-routing ablation\. We repair only the topp%p\\%highest\-risk examples ranked by\[Mask\]\-Discand report AlignScore\. Top\-pprepair leaves the remaining examples unchanged\.
## Appendix GAdditional Analysis of\[Mask\]\-DiscRisk Scores

Section[5](https://arxiv.org/html/2606.12807#S5)uses\[Mask\]\-Discas a sample\-level repair router\. For each summary, we compute a risk score by averaging the top\-kktoken suspiciousness scores:

S​\(x,y\)=1k​∑i∈TopK​\(s,k\)si\.S\(x,y\)=\\frac\{1\}\{k\}\\sum\_\{i\\in\\mathrm\{TopK\}\(s,k\)\}s\_\{i\}\.We setkkto match the token\-level edit budget used by the corresponding repair pipeline\. For DialogSum,k=8k=8, reflecting its shorter dialogue summaries and conservative edit budget\. For StreamSum,k=64k=64, reflecting longer summaries and the need to revise multiple related facts in evolving\-event updates\. We use the samekkconsistently for ranking summaries in the budgeted repair ablation\.

#### Correlation with AlignScore\.

To evaluate whether\[Mask\]\-Discrisk is meaningful as a sample\-level signal, we correlateS​\(x,y\)S\(x,y\)with per\-example AlignScore computed against the full context\. A negative correlation indicates that summaries assigned higher risk by\[Mask\]\-Disctend to be less supported by the full context\.

Table 6:Correlation between sample\-level\[Mask\]\-Discrisk and per\-example AlignScore\. The risk signal is strongest in the early\-draft setting, where summaries are more likely to contain stale or unsupported content\.The early\-draft correlations are negative and statistically significant on both datasets, supporting\[Mask\]\-Discas a sample\-level router for budgeted repair\. In the post\-hoc full\-generation setting, the relationship is weaker because summaries are already generated from the full context and contain fewer stale spans\. This is especially visible on StreamSum full\-generation outputs, where the mean risk score is low and the correlation is not statistically significant\.

![Refer to caption](https://arxiv.org/html/2606.12807v1/figures/disc_risk_correlation/dialogsum_early_ar_disc_risk_vs_alignscore.png)

![Refer to caption](https://arxiv.org/html/2606.12807v1/figures/disc_risk_correlation/streamsum_early_ar_disc_risk_vs_alignscore.png)

Figure 4:Sample\-level\[Mask\]\-Discrisk versus AlignScore for early\-context AR drafts\. Higher risk generally corresponds to lower faithfulness, supporting budgeted repair routing\.![Refer to caption](https://arxiv.org/html/2606.12807v1/figures/disc_risk_correlation/dialogsum_fullgen_disc_risk_vs_alignscore.png)

![Refer to caption](https://arxiv.org/html/2606.12807v1/figures/disc_risk_correlation/streamsum_fullgen_disc_risk_vs_alignscore.png)

Figure 5:Sample\-level\[Mask\]\-Discrisk versus AlignScore for full\-generation summaries\. The relationship is weaker in this setting because full\-generation outputs are already conditioned on the complete context\.
#### Risk\-score correlation\.

We also examine whether the sample\-level\[Mask\]\-Discrisk score correlates with AlignScore\. The correlations are negative and statistically significant for early\-draft repair, but small in magnitude: Spearmanρ=−0\.239\\rho=\-0\.239on DialogSum andρ=−0\.299\\rho=\-0\.299on StreamSum\. This suggests that\[Mask\]\-Discprovides a coarse risk signal in the expected direction, but should not be interpreted as a standalone faithfulness metric\. Our main evidence for budgeted routing is therefore the downstream repair behavior in Table[5](https://arxiv.org/html/2606.12807#A6.T5)\.

#### Budgeted Repair Timing

Table[7](https://arxiv.org/html/2606.12807#A7.T7)reports repair time under different budgeted routing thresholds\. Timing reports additional repair/refinement cost averaged over all examples, including examples skipped by the router\.

Table 7:Additional repair time under budgeted routing thresholds\. Times are seconds per example averaged over all examples, including skipped examples\.

## Appendix HRepresentative Repair Examples

To illustrate the inspectability ofDetect–Remask–Repair, Table[8](https://arxiv.org/html/2606.12807#A8.T8)shows representative repairs\. We truncate the source context for space and highlight stale spans selected for repair in red and repaired content in green\. The model operates over tokenizer\-level units; for readability, we display word\-level spans\. These examples are intended to show the kinds of factual updates made by the repair model, rather than to provide a full qualitative evaluation\.

Table 8:Representative inspectable repair examples\. We show truncated context evidence, stale spans selected from the AR draft, the corresponding re\-masked input, and the repaired output\. Selected stale spans are shown in red and repaired content in green\.
## Appendix IStreamSum Construction Details

StreamSum is designed to evaluate summarization repair under evolving evidence\. Each example contains an early context, a full updated context, an early\-context draft summary, and a gold summary of the full context\. The goal is to create examples where a summary generated from the early context is plausible at the time it is written, but later becomes incomplete or partially unsupported after subsequent updates\. The benchmark is constructed from real\-world news seeds retrieved from NewsDataHub333[https://www\.newsdatahub\.com/](https://www.newsdatahub.com/)through API calls, then expanded into controlled multi\-update event timelines using stage\-specific synthesis and validation agents\. Representative examples are shown in Table[11](https://arxiv.org/html/2606.12807#A9.T11)\.

#### Pipeline overview\.

We construct StreamSum with an agentic synthesis pipeline seeded by real\-world news\. The pipeline has eight stages: \(1\) schema brainstorming, \(2\) seed\-article collection from NewsDataHub, \(3\) timeline construction, \(4\) timeline hardening, \(5\) summary construction, \(6\) AlignScore\-based support filtering, \(7\) verifier\-agent validation, and \(8\) diversity\-aware pruning\.

#### Models and tools\.

Generation\-heavy stages use Claude Opus 4\.6 for schema brainstorming, timeline construction, timeline hardening, and full\-context summary construction\. Verification and pruning stages use MiniMax\-M2\.5 to evaluate candidate timelines for realism, entity consistency, revision clarity, and diversity\. The pipeline also calls AlignScore as an automatic support\-scoring tool: for each candidate, we compute support between the gold summary and each timeline prefix, producing a support trajectory from early updates to the full context\. Verifier agents use these trajectories, along with the timeline text, to identify examples where later evidence is necessary for a faithful final summary\.

#### Seed articles\.

We retrieve real\-world seed articles from NewsDataHub using domain\-balanced retrieval policies\. The retrieval stage prioritizes articles with event\-evolution potential, such as preliminary numbers, unresolved outcomes, investigations, attribution uncertainty, named entities, concrete dates, and later correction or update cues\. We deduplicate articles by URL, title, and content hash, and filter low\-value sources such as advertisements, recipes, coupons, and photo\-gallery pages\. The seed articles provide realistic entities, event types, and reporting style, while the synthesis pipeline controls the revision structure needed for evaluation\.

#### Revision schemas\.

A schema\-generation agent proposes candidate event schemas\. Each schema specifies an event family, revision type, scenario description, dominant revision point, why the first half is misleading, the late evidence needed, and expected updates\. We use six event families: politics, business, disaster, international, sports, and science\. We use revision types including outcome flips, numeric updates, attribution updates, status changes, timeline changes, and location changes\. Schemas are required to contain one dominant factual revision rather than multiple unrelated changes\.

#### Timeline construction and hardening\.

Given a schema and seed article, a timeline\-construction agent produces a coherent sequence of 4–5 news\-style updates\. Early updates contain plausible but incomplete information, while later updates introduce evidence that changes the correct final interpretation\. A hardening agent then improves realism, removes redundancy, strengthens the revision arc, and checks that all updates remain anchored to the same central event\. The resulting full context is the concatenation of all updates, while the early context consists of the first part of the timeline\.

#### Gold summary construction\.

A summary\-construction agent writes a concise full\-context summary\. The prompt requires the summary to reflect the final state of the event, include the dominant revision, and avoid claims that are only supported by early updates\. The summary is intended to be well supported by the full context but not fully supported by the early context\.

#### AlignScore prefix filtering\.

We compute AlignScore between the gold summary and each timeline prefix:t1t\_\{1\},t1\+t2t\_\{1\}\{\+\}t\_\{2\}, …, and the full timeline\. This produces a support trajectory for each candidate\. We retain examples where the final summary has strong support under the full context but weaker support under early prefixes\. Concretely, candidates are preferred when the full\-context AlignScore is high, the first\-prefix AlignScore is low, and the full\-minus\-early support gap is large\. This filtering encourages examples where later updates are necessary for a faithful summary\.

#### Verifier and pruning agents\.

A verifier agent evaluates each candidate using both the textual timeline and the AlignScore prefix trajectory\. It scores temporal coherence, entity consistency, dominant revision quality, early\-context misleadingness, final\-context support, and benchmark usefulness\. Finally, a pruning agent selects a diverse subset, balancing event families, revision types, difficulty levels, and near\-duplicate patterns\.

#### Dataset splits\.

The final StreamSum splits contain 2,000 training examples, 500 validation examples, and 500 test examples\. We use the train and validation splits to train\[Mask\]\-Discand one\-step repair models, and reserve the test split for final evaluation\.

Table 9:StreamSum construction pipeline\. The benchmark is built with real\-world news seeds, agentic timeline synthesis, automatic support\-dynamics filtering, and verifier/pruner agents\.
### I\.1StreamSum Agent Prompts

#### Prompts\.

We use separate prompts for schema brainstorming, seed instantiation, timeline construction, timeline hardening, summary construction, verification, and pruning\. Table[10](https://arxiv.org/html/2606.12807#A9.T10)summarizes the role of each prompt, and full prompt files will be released with the benchmark construction code\.

Table 10:Summary of StreamSum synthesis prompts\. Full prompts are released with the benchmark construction code\.
#### Full prompts\.

For space, Table[10](https://arxiv.org/html/2606.12807#A9.T10)summarizes the prompt roles\. The complete prompt files for the brainstormer, retriever/instantiator, timeline constructor, hardener, summary constructor, verifier, and pruner agents are released with the StreamSum construction code\.

### I\.2Example StreamSum Instances

Table[11](https://arxiv.org/html/2606.12807#A9.T11)shows representative StreamSum examples across domains and revision types\. Each example contains an early claim that is plausible from the early context and a later correction that changes the support needed for the final summary\. These examples illustrate the benchmark’s focus on evolving evidence rather than static summarization\.

Table 11:Representative StreamSum examples\. Each row shows an early claim and a later correction/final state, illustrating the dataset’s controlled revision patterns across domains\.