ReDeck: Step-Level Render-Grounded Refinement for Document-to-Slide Generation

# Introduction
Source: [https://arxiv.org/html/2609.00194](https://arxiv.org/html/2609.00194)
August 2026

ReDeck: Step\-Level Render\-Grounded Refinement for Document\-to\-Slide Generation

Muzhao Tian4,∗†Zezi Zeng3,∗Yifan Yang1,‡Xin Gao4Yan Li2Zisu Huang4 Xiaohua Wang4Changze Lv4Mingxi Cheng1Bei Liu1Kai Qiu1Qi Dai1 Dong Chen1Yue Dong1Xiaoqing Zheng4,‡Ji Li1Chong Luo1

1Microsoft Corporation2Shanghai Jiao Tong University3Xi’an Jiaotong University4Fudan University

Document\-to\-slide generation is challenging because slides are dense editable artifacts that require both faithful content selection and precise spatial layout\. Recent slide agents adopt iterative reflection, but typically follow a monolithic “one version, one feedback” loop: a slide or deck is rewritten, rendered afterward, and critiqued only at the turn boundary\. This delayed feedback makes local failures such as overflow, overlap, clipping, and off\-canvas placement difficult to attribute and repair\. We proposeReDeck, a step\-level render\-grounded refinement framework that decomposes slide revision into atomic edit actions and returns renderer\-derived observations after each step, turning refinement into “one edit, one observation\.” To balance local repair with global quality,ReDeckuses multi\-granular feedback: step\-level render feedback for spatial errors, a turn\-level adaptive critic for semantic and design guidance, and a submission\-level gate for hard layout validation\. We further introduceDeckQuiz, a benchmark that decouples content fidelity, spatial correctness, and design quality\. Across GPT\-5\.4, Claude\-4\.6, and Gemini\-3\.1,ReDeckconsistently outperforms existing slide\-generation agents, and ablations confirm that feedback timing and granularity are critical for reliable slide refinement\.Project Page:[https://aka\.ms/ReDeck](https://aka.ms/ReDeck)Code:[https://github\.com/microsoft/ReDeck](https://github.com/microsoft/ReDeck)Date:August 2026License:[CC BY 4\.0](https://creativecommons.org/licenses/by/4.0/)∗Equal contribution\.†Work done during an internship at Microsoft\.‡Corresponding authors\.

Presentation slides are a central medium for communicating complex documents, including scientific papers, technical reports, and business analyses\. Unlike plain text summaries, slides are dense visual artifacts: they must preserve source semantics, organize information into a coherent narrative, and arrange text, figures, charts, icons, and visual emphasis within a bounded two\-dimensional canvas\. Document\-to\-slide generation is therefore a coupled semantic\-spatial problem, where content fidelity, layout validity, and visual readability must be optimized jointly\.

Recent document\-to\-slide systems have largely followed a generate\-then\-render paradigm\[[9](https://arxiv.org/html/2609.00194#bib.bib58),[61](https://arxiv.org/html/2609.00194#bib.bib4),[10](https://arxiv.org/html/2609.00194#bib.bib5),[28](https://arxiv.org/html/2609.00194#bib.bib14),[62](https://arxiv.org/html/2609.00194#bib.bib13)\]\. Early methods attempt to produce a complete deck in one pass, for example by extracting an outline and rendering slides with predefined layouts\[[28](https://arxiv.org/html/2609.00194#bib.bib14)\]\. Since one\-shot generation often fails to satisfy the many constraints of a polished deck, recent systems naturally introduce reflection: the agent renders a candidate deck, receives feedback from a critic, and revises the deck in the next turn, following the broader practice of LLM self\-correction\[[49](https://arxiv.org/html/2609.00194#bib.bib37),[58](https://arxiv.org/html/2609.00194#bib.bib36),[33](https://arxiv.org/html/2609.00194#bib.bib6),[40](https://arxiv.org/html/2609.00194#bib.bib7),[64](https://arxiv.org/html/2609.00194#bib.bib11)\]\. However, this standard refinement loop is still*monolithic*: the agent commits a large slide\- or deck\-level rewrite, observes the rendered result only afterward, and then receives a holistic critique\. In other words, existing agents mostly follow a “one version, one feedback” pattern\.

This pattern is poorly matched to slide refinement\. Unlike natural images, slides are editable artifacts with many small, information\-bearing elements\. A single flawed slide may contain high\-level issues such as missing contributions, incorrect claims, weak narrative flow, or poor source grounding, while also containing low\-level issues such as text overflow, element overlap, low contrast, unused space, clipping, and misalignment\. Precisely describing all these issues in one global critique requires difficult alignment between visual regions, source content, and editable code\. Even when the critique is correct, repairing many entangled issues at once is hard: the agent must decide which defect to fix first, infer which edit caused each failure, and avoid breaking already\-correct regions\. As illustrated in Fig\.[1](https://arxiv.org/html/2609.00194#S1.F1)left, after a monolithic rewrite, the deck may contain multiple spatial failures, but their causes are no longer attributable to individual edits\.

This motivates a different refinement principle: feedback should be delivered at the granularity where the corresponding problem can be most reliably observed and fixed\. For local layout errors, waiting until the end of a turn is unnecessarily delayed\. If adding bullets immediately creates overflow, resizing a block immediately causes overlap, or moving an image immediately pushes content off the canvas, the agent should observe that rendered consequence before taking the next action\. Such step\-level feedback reduces the critic’s burden because each observation only needs to explain the consequence of the latest edit, and it reduces the editor’s burden because the agent can fix one problem while the cause is still clear\. At the same time, step\-level layout feedback alone cannot judge whether the whole deck is faithful, complete, coherent, and well designed\. Slide refinement therefore requires*multi\-granular feedback*: fine\-grained render\-grounded observations for local spatial repair, and coarser turn\-level reflection for global semantic and design direction\.

![Refer to caption](https://arxiv.org/html/2609.00194v1/Fig1.png)Figure 1:Turn\-level reflection vs\.ReDeck’s step\-level render\-grounded refinement\.Left: conventional monolithic refinement commits large slide rewrites and receives feedback only after rendering the result, making overflow, overlap, low contrast, and off\-canvas failures difficult to attribute to individual edits\.Right:ReDeckdecomposes refinement into atomic edits and returns renderer\-derived observations after each step, so the agent can fix local layout errors as soon as they appear while receiving global turn\-level guidance\.We introduceReDeck, a*step\-level render\-grounded refinement*framework for document\-to\-slide generation\. Instead of revising slides through monolithic rewrites,ReDeckconstrains the agent to a compositional action space of atomic edits, such as adding an image, inserting bullets, resizing a block, moving an element, or adjusting visual attributes\. After each action, the rendering environment returns structured observations about the current visual state, including overflow, overlap, clipping, off\-canvas placement, contrast issues, and key element geometry\. This changes the refinement loop from “one version, one feedback” to “one edit, one observation”, enabling the agent to ground its next edit in the actual rendered consequence of the previous one, as shown in Fig\.[1](https://arxiv.org/html/2609.00194#S1.F1)right\.

To coordinate local repair with global presentation quality,ReDeckimplements multi\-granular feedback with separated responsibilities\. At the step level, deterministic render feedback reports immediate spatial facts but does not act as a global judge\. At the turn level, an adaptive deck critic maintains a persistent issue list for higher\-level concerns, including narrative flow, content completeness, factual correctness, source fidelity, and visual design\. At the submission level, a validation gate ensures that the total hard\-violation count does not increase relative to the session baseline before a turn is committed\. This design lets each feedback source operate where it is most reliable: renderer observations handle local layout failures, the deck critic handles semantic and design direction, and the submission gate ensures that temporary invalid states are not preserved as final progress\.

To evaluate this framework, we introduceDeckQuiz, a benchmark that separately measures content fidelity, spatial correctness, and design quality\. This decomposition allows us to attribute improvements to the corresponding refinement components rather than relying only on aggregated final\-deck scores\. Across GPT\-5\.4, Claude\-4\.6, and Gemini\-3\.1,ReDeckconsistently outperforms existing slide\-generation agents under a matched per\-task LLM\-call\-count cap\. Ablation studies show that step\-level render grounding substantially improves spatial correctness and stabilizes refinement, while the turn\-level adaptive critic provides complementary gains in semantic and design quality\. Shared\-start controls, external\-benchmark evaluation, a 45\-task blinded human study, and PPTX transfer test the attribution, evaluator dependence, and format scope of the proposed refinement principle\.

Contributions\.1\) Step\-level render\-grounded refinement\.We formulate slide refinement as a sequence of atomic edit actions, each followed by renderer\-derived observations, enabling local layout failures to be detected and repaired at the moment they occur\.2\) Multi\-granular feedback for slide agents\.We combine step\-level deterministic render feedback, turn\-level adaptive reflection, and submission\-level validation, assigning different feedback sources to the levels where they are most reliable and useful\.3\)DeckQuizfor component\-level evaluation\.We introduce a benchmark that decouples content fidelity, spatial correctness, and design quality, enabling fine\-grained attribution of improvements in document\-to\-slide generation\.

## Related Work

Automated Slide Generation\.Early document\-to\-slide work focused on content selection and layout from structured inputs\[[9](https://arxiv.org/html/2609.00194#bib.bib58),[42](https://arxiv.org/html/2609.00194#bib.bib2),[1](https://arxiv.org/html/2609.00194#bib.bib3)\]\. More recent systems frame slide creation as an agentic task:AutoPresent\[[10](https://arxiv.org/html/2609.00194#bib.bib5)\],PPTAgent\[[61](https://arxiv.org/html/2609.00194#bib.bib4)\],SlideGen\[[28](https://arxiv.org/html/2609.00194#bib.bib14)\], andAuto\-Slides\[[56](https://arxiv.org/html/2609.00194#bib.bib17)\]decompose the pipeline across multiple specialized agents, whileSlideTailor\[[59](https://arxiv.org/html/2609.00194#bib.bib16)\]andArcDeck\[[36](https://arxiv.org/html/2609.00194#bib.bib15)\]further address personalization and narrative structure\. Despite these advances, most systems evaluate and refine at a coarse turn\-level granularity\.DeepPresenter\[[62](https://arxiv.org/html/2609.00194#bib.bib13)\]is the most related, as it introduces rendered inspection within the refinement loop, but it does not validate layout integrity after each individual edit, allowing spatial errors to accumulate silently within a turn\. Beyond slides, code\-from\-screenshot work in HTML/web settings\[[41](https://arxiv.org/html/2609.00194#bib.bib53),[26](https://arxiv.org/html/2609.00194#bib.bib54)\]addresses a closely related symbolic\-to\-rendered nonlinearity but operates on single\-page artefacts without a multi\-turn refinement contract, and the broader LLM\-driven layout\-generation literature\[[14](https://arxiv.org/html/2609.00194#bib.bib56),[19](https://arxiv.org/html/2609.00194#bib.bib57),[8](https://arxiv.org/html/2609.00194#bib.bib55)\]predicts element placements in one shot without a render\-grounded repair loop\.

Feedback and Self\-Correction in LLM Agents\.Iterative refinement through feedback is a central paradigm for improving model performance, building on the reasoning\-and\-acting trajectory established by chain\-of\-thought prompting andReAct\[[49](https://arxiv.org/html/2609.00194#bib.bib37),[58](https://arxiv.org/html/2609.00194#bib.bib36)\]and exemplified by frameworks such asSelf\-Refine\[[33](https://arxiv.org/html/2609.00194#bib.bib6)\]andReflexion\[[40](https://arxiv.org/html/2609.00194#bib.bib7)\]which utilize verbal self\-correction\. Grounding feedback in external evidence has been further developed in works likeCritic\[[11](https://arxiv.org/html/2609.00194#bib.bib8)\]and related self\-correction studies\[[18](https://arxiv.org/html/2609.00194#bib.bib9)\]; execution\-feedback variants ground the loop in compiler or test signals\[[4](https://arxiv.org/html/2609.00194#bib.bib39),[45](https://arxiv.org/html/2609.00194#bib.bib40),[17](https://arxiv.org/html/2609.00194#bib.bib41)\], while multi\-agent SOP frameworks\[[15](https://arxiv.org/html/2609.00194#bib.bib48),[52](https://arxiv.org/html/2609.00194#bib.bib49)\]and software\-engineering agent benchmarks\[[21](https://arxiv.org/html/2609.00194#bib.bib42),[48](https://arxiv.org/html/2609.00194#bib.bib43)\]extend the same principle to richer environments\. While environments likeWebArena\[[64](https://arxiv.org/html/2609.00194#bib.bib11)\],AgentBench\[[31](https://arxiv.org/html/2609.00194#bib.bib12)\], andSWE\-agent\[[54](https://arxiv.org/html/2609.00194#bib.bib10)\]demonstrate the necessity of real\-time observations, and GUI / vision agents that ground actions in rendered screenshots\[[16](https://arxiv.org/html/2609.00194#bib.bib51),[60](https://arxiv.org/html/2609.00194#bib.bib52)\]make the same point in the visual modality, slide agents often lack a specific rendered\-state delta to capture immediate layout regressions\. Existing systems predominantly rely on holistic or scalar reflection\[[10](https://arxiv.org/html/2609.00194#bib.bib5),[36](https://arxiv.org/html/2609.00194#bib.bib15),[62](https://arxiv.org/html/2609.00194#bib.bib13)\]\. In contrast,ReDeckdecouples feedback into step\-level objective facts and turn\-level semantic guidance to stabilize long\-horizon refinement\.

Benchmarks for Document Design and Layout\.Accurate evaluation of generated presentations requires moving beyond generic text similarity\[[12](https://arxiv.org/html/2609.00194#bib.bib25),[10](https://arxiv.org/html/2609.00194#bib.bib5),[61](https://arxiv.org/html/2609.00194#bib.bib4),[43](https://arxiv.org/html/2609.00194#bib.bib59)\]toward metrics that preserve source fidelity and canvas integrity\. Recent benchmarks such asPresentBench\[[3](https://arxiv.org/html/2609.00194#bib.bib1)\],SlidesGen\-Bench\[[55](https://arxiv.org/html/2609.00194#bib.bib18)\],ArcDeck\[[36](https://arxiv.org/html/2609.00194#bib.bib15)\],AeSlides\[[37](https://arxiv.org/html/2609.00194#bib.bib19)\], andDECKBench\[[20](https://arxiv.org/html/2609.00194#bib.bib20)\], alongside cross\-format suites likeBizGenEval\[[27](https://arxiv.org/html/2609.00194#bib.bib24)\], have introduced verifiable checks for layout dimensions, whitespace, visual imbalance, and reference\-aligned content fidelity\. These developments align with broader research on multi\-constraint instruction following\[[13](https://arxiv.org/html/2609.00194#bib.bib21)\]and query\-specific rubrics\[[32](https://arxiv.org/html/2609.00194#bib.bib23)\], emphasizing that evaluation signals should be instance\-specific\. Rooted in classical principles of perceptual organization and multimedia learning\[[50](https://arxiv.org/html/2609.00194#bib.bib26),[34](https://arxiv.org/html/2609.00194#bib.bib27),[44](https://arxiv.org/html/2609.00194#bib.bib28),[29](https://arxiv.org/html/2609.00194#bib.bib29)\], our proposedDeckQuizaddresses the variance and bias inherent inLLM\-as\-a\-judgemetrics\[[63](https://arxiv.org/html/2609.00194#bib.bib44),[5](https://arxiv.org/html/2609.00194#bib.bib45),[23](https://arxiv.org/html/2609.00194#bib.bib46),[38](https://arxiv.org/html/2609.00194#bib.bib22),[47](https://arxiv.org/html/2609.00194#bib.bib35)\]by separating hard spatial\-violation criteria from subjective design principles\.

## Method

In this section, we presentReDeck, a step\-level render\-grounded refinement framework designed to resolve the perception\-action gap in slide generation caused by highly nonlinear spatial rendering\. By splitting supervision into multi\-granular feedback, consisting of an inner step\-level render\-feedback channel and an outer turn\-level critic \(Figure[2](https://arxiv.org/html/2609.00194#S3.F2)\),ReDeckreduces the feedback delay fromO⁡\(turn\)O\(\\text\{turn\}\)toO⁡\(edit\)O\(\\text\{edit\}\)\. This dual\-scale architecture efficiently grounds individual edits in rendered reality while relying on the critic for comprehensive design direction\. The remainder of this section is organized as follows: we first formalize the refinement environment and define a compositional action space to prevent untraceable layout errors \(§[3\.1](https://arxiv.org/html/2609.00194#S3.SS1)\); we then detail the core dual\-level feedback mechanisms \(§[3\.2](https://arxiv.org/html/2609.00194#S3.SS2)and[3\.3](https://arxiv.org/html/2609.00194#S3.SS3)\); and finally, we describe the completeReDeckloop that systematically integrates both scales under strict validation gates \(§[3\.4](https://arxiv.org/html/2609.00194#S3.SS4)\)\.

![Refer to caption](https://arxiv.org/html/2609.00194v1/Fig2.png)Figure 2:ReDeckpipeline\.Each refinement turn nests two feedback scales over a partially observable decision process\(S,A,Ω\)\(S,A,\\Omega\)\.Turn lane:Adaptive Deck Critic uses a schedulerπ𝒫\\pi\_\{\\mathcal\{P\}\}to draw a probe subset from library𝒫\\mathcal\{P\}and updates the persistent issue listLkL\_\{k\}\.Edit lane:after every atomic actionata\_\{t\}from the compositional action space, the rendering environment returns render\-state observationot=\(Δviol​\(dt,d0\),At\)o\_\{t\}=\(\\Delta\_\{\\mathrm\{viol\}\}\(d\_\{t\},d\_\{0\}\),A\_\{t\}\)to the next reasoning step\. Step\-level render feedback is observation, not verdict; the only hard gate is the submission boundaryρ\\rho, comparing the rendered violations of the submitted deck against session\-start baselined0d\_\{0\}\.### Problem Modeling and Compositional Action Design

The iterative generation of slide decks involves transitioning between a*semantic reasoning space*and a*rendered spatial space*\. We model this process within a single turn as a sequence of discrete steps\. Letttdenote the inner step index\. At steptt, the dynamic state of the rendering environment is represented as the tuple:st=\(dt,rt,Ct\)s\_\{t\}=\(d\_\{t\},r\_\{t\},C\_\{t\}\), wheredtd\_\{t\}is the editable deck representation \(e\.g\., HTML/CSS or structured presentation objects\),rtr\_\{t\}is the spatial geometry produced by the active rendering stack \(e\.g\., browser DOM boxes for HTML or shape bounds for PPTX\), andCtC\_\{t\}contains compiler or export diagnostics \(e\.g\., syntax errors or asset loading failures\)\. Because the agent only editsdtd\_\{t\}, it cannot directly perceive the physical layoutrtr\_\{t\}or diagnosticsCtC\_\{t\}without rendering\. To avoid overwhelming the agent’s context window with high\-dimensional spatial coordinates and raw logs, the environment filtersrtr\_\{t\}andCtC\_\{t\}to produce a compact, text\-based observationot∈Ωo\_\{t\}\\in\\Omega\. This step\-level observationoto\_\{t\}acts as a filtered report that summarizes only the immediate technical failures and critical element coordinates resulting from the latest edit\.

At the start of each refinement turn, the agent is also provided with a persistent issue listLL\. This list contains high\-level design or content defects \(e\.g\., overlapping text or missing citations\) and remains static throughout the inner steps\. At each steptt, the agent uses the active observationoto\_\{t\}alongside this static issue listLLto select its next actionat∈Aa\_\{t\}\\in A\.

To handle the highly nonlinear mapping from the source codedtd\_\{t\}to the rendered geometryrtr\_\{t\}\[[41](https://arxiv.org/html/2609.00194#bib.bib53),[8](https://arxiv.org/html/2609.00194#bib.bib55)\],ReDeckconstrains the agent to a compositional action spaceAAof atomic edits\. Rather than rewriting an entire slide in one step, the agent performs discrete modifications across three orthogonal categories:*structural actions*for global element arrangements \(e\.g\., adding or removing shapes, reflowing layout\),*content actions*for updating semantic information \(e\.g\., editing text, replacing figures\), and*attribute actions*for precise geometric adjustments \(e\.g\., resizing, repositioning\)\.111At the tool level, these logical action types are implemented through a general\-purposeapply\_editstool that performs targeted search\-and\-replace operations on the slide source code, along withverify\_layout,rollback, andsubmitcontrol tools\. The agent is prompted to perform one logical edit per tool call; the three\-way taxonomy describes the*intent*of each edit rather than distinct tool endpoints\.This atomic design isolates the spatial consequences of each edit, ensuring that any layout violation reported inoto\_\{t\}can be directly attributed to a single, identifiable action\. The action set is completed byrollback, which allows the agent to revert unsuccessful edits and use theedit→\\toobserve→\\torevert→\\tore\-attemptloop as a deliberate local search reminiscent of tree\-search reasoning\[[57](https://arxiv.org/html/2609.00194#bib.bib38)\], andsubmit, which concludes the current turn\. This formulation provides the structural foundation for the dual\-level feedback mechanisms described in the following sections\.

### Inner Loop: Step\-Level Render Feedback

To help the agent see the direct results of its actions,ReDeckevaluates each atomic actionata\_\{t\}immediately, drawing the step\-level supervision idea from process\-reward modelling for reasoning chains\[[30](https://arxiv.org/html/2609.00194#bib.bib34),[46](https://arxiv.org/html/2609.00194#bib.bib50)\]and porting it from*reasoning steps*to*rendering steps*\. Instead of waiting until the end of a turn, the environment compiles the deck after every single edit and returns a step\-level observationot∈Ωo\_\{t\}\\in\\Omegato the next step\. This per\-step loop creates a continuous cycle of editing, rendering, and observing, which helps detect layout errors before they accumulate\. The step\-level observation is defined as:

ot=\(Δviol​\(dt,d0\),At\),Δviol​\(dt,d0\):=\|𝒱⁡\(dt\)\|−\|𝒱⁡\(d0\)\|o\_\{t\}=\\bigl\(\\Delta\_\{\\text\{viol\}\}\(d\_\{t\},\\ d\_\{0\}\),\\ A\_\{t\}\\bigr\),\\hskip 9\.24994pt\\Delta\_\{\\text\{viol\}\}\(d\_\{t\},\\ d\_\{0\}\):=\|\\mathcal\{V\}\(d\_\{t\}\)\|\-\|\\mathcal\{V\}\(d\_\{0\}\)\|where the violation deltaΔviol\\Delta\_\{\\text\{viol\}\}measures the change in compiler failures and DOM\-level errors \(such as text overflow, overlap, and clipping\) compared to the initial baselined0d\_\{0\}\. To help the agent locate these errors, the environment provides this delta along with specific details, including the IDs and bounding boxes of the failing elements\. The layout anchorAtA\_\{t\}records the exact positions\(x,y\)\(x,y\), dimensions\(w×h\)\(w\\times h\), and text previews of the main elements, providing a clear numerical grid for planning while filtering out visual noise like page numbers\. We explicitly exclude screenshots and subjective design scores at this step; step\-level screenshots often cause the agent to hallucinate, while subjective scores are too inconsistent to guide individual edits\.

During the editing process, this step\-level feedback acts as a soft hint rather than a strict constraint\. This soft policy is necessary because fixing a layout often requires going through temporary error states, such as a textbox temporarily overflowing before the agent can shrink its font\. However, to prevent the agent from submitting a slide with unfixed errors,ReDeckapplies a hard check at the end of the turn when the agent calls thesubmitaction\. This is controlled by a simple validation rule:

ρ⁡\(st,at,st\+1\)=\{\[Δviol\(dt\+1,d0\)≤0\]at=𝚜𝚞𝚋𝚖𝚒𝚝,1otherwise\.\\rho\(s\_\{t\},a\_\{t\},s\_\{t\+1\}\)=\\begin\{cases\}\\mathbb\{1\}\\\!\\left\[\\Delta\_\{\\text\{viol\}\}\(d\_\{t\+1\},\\ d\_\{0\}\)\\leq 0\\right\]&a\_\{t\}=\\mathtt\{submit\},\\\\ 1&\\text\{otherwise\.\}\\end\{cases\}Since this rule uses the same baselined0d\_\{0\}as the step\-level observations, the submission check is highly predictable\. Ifρ=0\\rho=0, the submission is rejected, and control returns to the agent with the current observation, forcing it to reduce the violation count to at most the baseline level before the turn can end\.

### Outer Loop: Turn\-Level Adaptive Deck Critic

While step\-level feedback focuses on individual edits, the agent still needs a broad, deck\-wide view of which parts of the presentation are unsatisfactory\. The Adaptive Deck Critic provides this view by evaluating both the slide source codeddand its rendered imagesrrat the start of each turn\. This process updates a persistent issue listLLdefined as:

L←Critic⁡\(d,r,L,π𝒫\)L\\leftarrow\\mathrm\{Critic\}\(d,\\ r,\\ L;\\ \\pi\_\{\\mathcal\{P\}\}\)whereπ𝒫\\pi\_\{\\mathcal\{P\}\}is the scheduling policy\. Rather than giving free\-form advice, the listLLconsists of structured entries\. Each entry contains the issue category, the affected slide ID, a pointer to the evidence, and a status marker\. The critic monitors five key categories: narrative flow, visual layout, completeness, correctness, and source fidelity\. On each turn, the critic either adds new issues toLLor updates the status of existing ones to represent whether they are still present, resolved, or regressed\. This structured approach ensures that resolving an issue requires a formal verdict from the critic rather than a simple self\-report by the agent\.

Probe library\.To detect these diverse problems, the critic uses a library𝒫\\mathcal\{P\}of 237 single\-purpose probes\. These probes are based on published slide benchmarks, classic visual design principles, and multimedia learning guidelines\. Each probe is implemented in one of three ways: geometric checks \(such as finding text overlaps or font size errors\), source\-retrieval checks \(using BM25 index lookups to verify factual claims\), or rubric\-based multimodal evaluations\. This structure ensures that LLM critiques enter the loop as formal, typed entries rather than unstructured conversation\. To maintain feedback stability, the schedulerπ𝒫\\pi\_\{\\mathcal\{P\}\}dynamically activates a small budget of three to five probes per turn\. The scheduler ranks probes based on which slides were recently edited, previous regression history, and the need to recheck resolved issues\. Running all 237 checks on every turn would create high judge variance, causing the critic to find inconsistent, minor issues that lead to conflicting edits and prevent the layout from converging\.

### The FullReDeckFramework

At the initiation of each refinement turnkk, the adaptive deck critic evaluates the current deck sourcedk,0d\_\{k,0\}to update the persistent issue listLkL\_\{k\}, which remains static throughout the turn\. The agent then executes an inner sequence of atomic edits in a thought–action–observation trajectory\[[58](https://arxiv.org/html/2609.00194#bib.bib36)\], consuming bothLkL\_\{k\}and the active step\-level observationoto\_\{t\}to select each actionat∈Aa\_\{t\}\\in A, until it invokes thesubmitaction to trigger the validation gateρ\\rho\. Upon a successful submission, the final compiled sourcedk,Tkd\_\{k,T\_\{k\}\}is persisted as the baseline for the next turn, settingdk\+1,0=dk,Tkd\_\{k\+1,0\}=d\_\{k,T\_\{k\}\}; a failed submission returns control to the agent for further corrective adjustments\.

To prevent optimization drift, the two scales operate under strictly decoupled write permissions: the turn\-level critic holds exclusive authority to append new issues, bind localization evidence, or verify the resolution of existing entries inLkL\_\{k\}, whereas the agent has read\-only access to these fields and the high\-frequency step\-level channel serves strictly as an instantaneous spatial report without modifyingLkL\_\{k\}\. This asymmetric information barrier prevents the agent from self\-reporting success, keepingLkL\_\{k\}an objective, environment\-enforced target\. The dual\-scale synergy is essential for convergence: without the critic, the agent over\-optimizes local layouts at the expense of deck\-level coherence; without step\-level spatial feedback, it edits the underlying source code blindly\. We empirically evaluate this dual\-scale topology in §[4](https://arxiv.org/html/2609.00194#S4)\.

## Experiments

### Setup

Task and benchmarks\.We target*scientific presentation generation*from full papers, the dominant evaluation setting in prior slide\-agent work\[[28](https://arxiv.org/html/2609.00194#bib.bib14),[59](https://arxiv.org/html/2609.00194#bib.bib16),[62](https://arxiv.org/html/2609.00194#bib.bib13),[61](https://arxiv.org/html/2609.00194#bib.bib4)\]; cross\-domain transfer across the fivePresentBenchdomains is in Sec\.[4\.5](https://arxiv.org/html/2609.00194#S4.SS5)\.DeckQuiz\(ours\) bundles four orthogonal modules:ContentQuiz\(Fid\., four QA types: contribution / method / experiment / limitation\),SpatialCheck\(SCR, VLM\-judged clean\-rate over four hard\-violation families\),Aesthetics\(Aes\.,00–55ordinal over five design axes\), andDeckDesign\(Des\.,00–33Likert over five info\-architecture principles in deck and slide layers; App\.[L](https://arxiv.org/html/2609.00194#A12)\); the four modules ask, respectively,*what*facts a deck conveys,*whether*they render cleanly,*how each slide looks*, and*how the deck organises information*\. The in\-domain split contains 100 papers stratified by discipline \(CS/ML 25, Bio\-Med 15, Physical\-Sci 25, Econ/Finance 25, Soc\-Sci/HCI 10\) and complexity \(30 short / 40 medium / 30 long\); construction details, source freeze, and per\-task call budget are in App\.[K](https://arxiv.org/html/2609.00194#A11)\.PresentBench\[[3](https://arxiv.org/html/2609.00194#bib.bib1)\]is reserved for the cross\-domain stress test \(Sec\.[4\.5](https://arxiv.org/html/2609.00194#S4.SS5), App\.[F](https://arxiv.org/html/2609.00194#A6)\); excluded from the main table due to known judge cross\-model drift\.

Models, comparisons, and protocol\.We hold the agent prompt and tool budget fixed across GPT\-5\.4, Gemini\-3\.1, and Claude\-4\.6, and compareReDeckagainst three published deck\-level slide agents rerun end\-to\-end under matched per\-task LLM\-call\-count cap:SlideGen\[[28](https://arxiv.org/html/2609.00194#bib.bib14)\],SlideTailor\[[59](https://arxiv.org/html/2609.00194#bib.bib16)\]\(code fromnusnlp/SlideTailor\), andDeepPresenter\[[62](https://arxiv.org/html/2609.00194#bib.bib13)\]\(code fromicip\-cas/PPTAgent; the most recent system, descended fromPPTAgent\[[61](https://arxiv.org/html/2609.00194#bib.bib4)\]\)\. In\-house refinement strategies \(Single\-Shot,Self\-Refine\[[33](https://arxiv.org/html/2609.00194#bib.bib6)\],Reflexion\[[40](https://arxiv.org/html/2609.00194#bib.bib7)\], per\-turnScreenshot\) appear in Sec\.[4\.3](https://arxiv.org/html/2609.00194#S4.SS3)as feedback\-paradigm ablations sharing our parser, tool set, and prompt scaffolding so that the contrast isolates the feedback channel\. Each \(model, system, task\) cell uses three seeds; main\-table entries are means, with seed variability and task\-paired inference reported in App\.[N](https://arxiv.org/html/2609.00194#A14)\. All systems run under a matched150150LLM\-call\-count cap \(App\.[M](https://arxiv.org/html/2609.00194#A13)\); token volume, cost, and latency are not matched and are reported separately in App\.[O](https://arxiv.org/html/2609.00194#A15)\. The fourDeckQuizVLM judges run on a single fixed GPT\-5\.4 backbone for all evaluated systems, ensuring consistent scoring across rows\[[63](https://arxiv.org/html/2609.00194#bib.bib44)\], with deterministic decoding \(App\.[I](https://arxiv.org/html/2609.00194#A9)\)\. Component, feedback\-frequency, scheduling, transfer, and human\-validation analyses are reported in Sec\.[4\.3](https://arxiv.org/html/2609.00194#S4.SS3)–[4\.5](https://arxiv.org/html/2609.00194#S4.SS5)\.

### Main results

Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)reports the fourDeckQuizmodules on the in\-domain100100tasks×\\times33seeds, comparingReDeckagainst three published slide agents under the matched per\-task LLM\-call\-count cap\. The ablations in Sec\.[4\.3](https://arxiv.org/html/2609.00194#S4.SS3)attribute the gap to specific components of the loop\.

Table 1:Performance comparison with published slide agents onDeckQuiz\(3 models×\\times100 tasks×\\times3 seeds\) under matched per\-task LLM\-call\-count caps\. Best results per column are inbold\.
Figure 3:Refinement convergence dynamics across scheduling and observation\-frequency variants\.\(a\)Open\-issue count per turn:ReDeck’s adaptive scheduler drops sharply between turns00–11then decays to a low plateau; all\-probes\-every\-turn drifts*upward*;N∈\{4,8\}N\{\\in\}\\\{4,8\\\}decay but plateau higher\.\(b\)Per\-turn resolution rate:ReDeckclimbs steepest, plateauing∼77%\{\\sim\}77\\%; per\-step variantsN=4∼74%N\{=\}4\{\\sim\}74\\%,N=8∼60%N\{=\}8\{\\sim\}60\\%; all\-probes flattens at∼20%\{\\sim\}20\\%\.\(c\)New issues raised this turn: near\-zero forReDeck, plateaus high for all\-probes, and decays slowly for low\-frequency variants\.n=300n\{=\}300\(100×3100\\times 3seeds\); final quality in Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)\.Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)shows thatReDeckachieves the highest score on all fourDeckQuizmodules for each agent backbone\. GPT\-5\.4 task\-paired inference confirms all 12 comparisons after Holm correction, while the other two backbones serve as descriptive replications \(App\.[N](https://arxiv.org/html/2609.00194#A14)\)\. BecauseReDeck’s initial draft is already strong onFid\.andDes\., we also evaluate both refinement procedures from the ReDeck and DeepPresenter initial decks\. The refinement advantage persists across both starts \(App\.[N](https://arxiv.org/html/2609.00194#A14), Tab\.[11](https://arxiv.org/html/2609.00194#A14.T11)\)\. Component attribution is in Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2), and a real repair trajectory is presented in App\.[B](https://arxiv.org/html/2609.00194#A2)\.

### Ablations

All ablations use GPT\-5\.4 on the same 100 tasks and three seeds as the corresponding main result\. Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)compares turn\-level natural\-language and screenshot critique, isolates the two feedback components, varies the observation interval overN∈\{1,4,8\}N\\\!\\in\\\!\\\{1,4,8\\\}, and contrasts adaptive scheduling with evaluation of all 237 probes at every turn\.

Table 2:Ablation on different feedback mechanisms\. Each block isolates a single design choice; the shaded row is theReDeckconfiguration of Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)\. Per\-turn convergence dynamics for the last two blocks are in Fig\.[3](https://arxiv.org/html/2609.00194#S4.F3)\.The feedback\-paradigm results show that both turn\-level alternatives underperformReDeck, with the largest differences appearing in SCR andDes\.The component ablations further show that the two feedback channels serve different roles: action\-level render feedback substantially improves SCR but reducesDes\.relative toT0T\_\{0\}, whereas the turn\-level critic provides a small design improvement without improving spatial correctness\. Combining the two channels yields the best result on all four metrics, consistent with their complementary local and global functions\.

Performance decreases consistently as the observation interval increases fromN=1N\{=\}1toN=4N\{=\}4andN=8N\{=\}8, indicating that feedback frequency affects both spatial correction and overall deck quality\. Likewise, evaluating all probes at every turn performs close toT0T\_\{0\}and substantially below adaptive scheduling\. Fig\.[3](https://arxiv.org/html/2609.00194#S4.F3)and App\.[Q](https://arxiv.org/html/2609.00194#A17)show that this setting raises additional visual findings whose subsequent repairs introduce geometric regressions\. Together, these results indicate that frequent render observation and selective probe scheduling jointly improve refinement stability\. Seed variability is summarized in App\.[N](https://arxiv.org/html/2609.00194#A14)\.

Fig\.[4](https://arxiv.org/html/2609.00194#S4.F4)further decomposes open issues by family, showing that growth inBvisualB\_\{\\text\{visual\}\}under exhaustive scheduling is accompanied by increasingBgeomB\_\{\\text\{geom\}\}\.

Figure 4:Per\-family open\-issue composition across the refinement trajectory\.Each coloured layer is the count of open issues from one probe family at the end of turntt\.Left: underReDeck, every family decays; step\-level render feedback closesBgeomB\_\{\\text\{geom\}\}fastest, then the persistent list steadily reducesAA,CC,DD, andEE\.Right: under all\-probes\-every\-turn,BvisualB\_\{\\text\{visual\}\}grows monotonically as the critic raises new style findings every turn; the edits triggered by these findings in turn introduce layout regressions, causingBgeomB\_\{\\text\{geom\}\}to rise in parallel\.
### Cost and latency

The hierarchical loop incurs substantially greater inference cost than single\-shot generation\. On the same100100\-task GPT\-5\.4 split as Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1),ReDeck’s end\-to\-end wall\-clock is35863586s \(8\.5×8\.5\{\\times\}DeepPresenter\) and per\-task USD cost is $1\.741\.74\(14\.5×14\.5\{\\times\}\)\. To control for inference cost, we increase DeepPresenter’s post\-T0T\_\{0\}call cap from150150to21752175, approximately matchingReDeck’s mean post\-T0T\_\{0\}cost of $1\.741\.74rather than each task’s exact expenditure\. At this budget, DeepPresenter reaches77\.0/78\.0/3\.20/63\.077\.0/78\.0/3\.20/63\.0onFid\./SCR/Aes\./Des\., compared withReDeck’s88\.6/91\.5/3\.64/71\.288\.6/91\.5/3\.64/71\.2\. A routedReDeck\-Lite variant retains high\-level planning and final submission on GPT\-5\.4 while sending recurrent execution and verification to a nano backbone; it costs $0\.280\.28per task and retains89\.889\.8SCR\. App\.[O](https://arxiv.org/html/2609.00194#A15)reports the cost, latency, and quality trade\-offs\.

### Robustness and validity

Cross\-domain transfer onPresentBench\.We rerunSlideGen,SlideTailor,DeepPresenter, andReDeckon the publicPresentBench\[[3](https://arxiv.org/html/2609.00194#bib.bib1)\]subset across all five domains \(*academia*,*advertising*,*economics*,*education*,*talk*; 3 models×\\times2 seeds\)\. Fig\.[5](https://arxiv.org/html/2609.00194#S4.F5)shows the resulting per\-domain win\-rate matrix together with the aggregate cross\-domain rank correlation; full per\-domain numbers are in App\.[F](https://arxiv.org/html/2609.00194#A6)\. We quantify ranking stability across the five domains using average pairwise Kendall’sτ\\tau\[[22](https://arxiv.org/html/2609.00194#bib.bib64)\]over the\(52\)=10\\binom\{5\}\{2\}\{=\}10domain pairs, withτ≥0\.85\\tau\\geq 0\.85specified as the threshold for strong cross\-domain consistency\.ReDeckis the top\-ranked system in all five domains \(Fig\.[5](https://arxiv.org/html/2609.00194#S4.F5), App\.[F](https://arxiv.org/html/2609.00194#A6)\) with per\-domain win rate0\.720\.72on*academia*and0\.660\.66–0\.680\.68on the most distant domains; the only deviation from the pooled order is a 2nd\-vs\-3rd swap on*talk*, whereSlideTailor’s template\-driven design priors marginally outscoreDeepPresenter\(0\.510\.51vs\.0\.500\.50\), and the resulting aggregateτ=0\.87\\tau=0\.87remains above the0\.850\.85threshold\.

![Refer to caption](https://arxiv.org/html/2609.00194v1/fig_cross_domain.png)Figure 5:Per\-domain win\-rate matrix onPresentBench\.Each cell: average pairwise win rate of the row system against the other three within that domain \(33models×\\times22seeds\)\. Right margin: row meanw¯\\overline\{w\}; subtitle: average pairwise Kendall’sτ\\tauacross the\(52\)=10\\binom\{5\}\{2\}\{=\}10domain pairs\. The only rank\-order change is a 2nd\-vs\-3rd swap on*talk*\(SlideTailornarrowly aboveDeepPresenter\), yieldingτ=0\.87\\tau=0\.87\.External evaluation\.To test whether the ranking depends onDeckQuiz, we evaluate the same systems withDECKBenchandSlidesGen\-Bench, which change the data and evaluator\. Against DeepPresenter,ReDeckimprovesDECKBenchLayoutQ, DeckFaith, and DeckFid by0\.0320\.032,0\.0230\.023, and0\.0220\.022, respectively, and improvesSlidesGen\-BenchQuizBank accuracy by2\.62\.6percentage points and computational aesthetics by3\.23\.2points\. The protocols, common\-set sizes, and system\-level results are in App\.[G](https://arxiv.org/html/2609.00194#A7); these metrics remain related to slide quality, but are not derived from theReDeckprobe taxonomy\.

Transfer to PPTX\.The refinement loop requires editable geometry and repeatable rendering rather than HTML APIs specifically\. Replacing DOM queries withpython\-pptxshape geometry and browser rendering with LibreOffice improves a PPTXT0T\_\{0\}from78\.0/58\.0/2\.80/60\.078\.0/58\.0/2\.80/60\.0to83\.0/82\.0/3\.30/66\.083\.0/82\.0/3\.30/66\.0on the fourDeckQuizmodules \(100 tasks×\\times3 seeds; App\.[H](https://arxiv.org/html/2609.00194#A8)\)\.

Human pairwise study\.Because the probe library𝒫\\mathcal\{P\}andDeckQuizwere built by the same authors, three independent annotators blindly compareReDeckwith the strongest baseline, DeepPresenter, on 45 held\-out GPT\-5\.4 tasks\. Majority vote favorsReDeckon overall presentation quality in33/4533/45cases \(0\.730\.73, Wilson 95% CI\[0\.59,0\.84\]\[0\.59,0\.84\]\) and on source faithfulness in31/4531/45cases \(0\.690\.69,\[0\.54,0\.80\]\[0\.54,0\.80\]\); the protocol and full counts are in App\.[D](https://arxiv.org/html/2609.00194#A4)\.

Internal\-validity audits\.Appendix audits guard against pipeline artefacts: probe\-taxonomy coverage \(App\.[J](https://arxiv.org/html/2609.00194#A10)\), VLM\-SCR calibration against non\-author labels and a deterministic DOM cross\-check \(App\.[E](https://arxiv.org/html/2609.00194#A5)\), and per\-principle\-layer regression onDes\.\(App\.[P](https://arxiv.org/html/2609.00194#A16)\)\. Title/DOI matching finds no paper overlap betweenDeckQuizandPresentBench; none of these checks reverses the headline ranking\.

## Conclusion

We have presentedReDeck, a step\-level render\-grounded refinement framework for document\-to\-slide generation\. By decomposing slide revision into atomic edit actions and returning renderer\-derived observations after each step,ReDeckcollapses the perception–action gap from an entire turn to a single edit, allowing the agent to detect and repair local layout failures at the moment they arise\. The turn\-level adaptive deck critic complements this with global design direction, and the submission\-level validation gate prevents regressions from persisting across turns\.

Experiments across three model backbones show that this dual\-scale feedback consistently outperforms existing slide agents on content fidelity, spatial correctness, aesthetics, and information architecture\. Ablations confirm that step\-level render feedback and the turn\-level critic are individually insufficient but jointly super\-additive, while shared\-start controls isolate the refinement loop from initial\-deck quality\.

More broadly,ReDeckillustrates a general principle for iterative generation of persistent, inspectable artifacts: feedback should be delivered at the granularity where the corresponding problem can be most reliably observed and fixed\. We believe this principle extends beyond slides to other domains where symbolic edits have nonlinear rendered consequences, including web pages, documents, diagrams, and interactive interfaces\.

## Limitations

ReDeckrequires editable geometry and repeatable rendering\. We validate HTML/CSS and PPTX, but not PDF/OCR\-only editing, Keynote, animations, masking, or font\-fallback behavior\. Native end\-to\-end comparisons also combine each system’s initial generator with its refinement loop; the shared\-start experiment isolates loop quality across two initial\-deck sources but does not make the native generators identical\.

The full system costs14\.5×14\.5\\timesmore and runs8\.5×8\.5\\timesslower than default DeepPresenter\. A mean\-cost\-matched control andReDeck\-Lite clarify the quality–cost trade\-off, but routing mitigates rather than removes it\. Finally,DeckQuizand the adaptive critic share related quality constructs\. External benchmarks, non\-author SCR labels, a deterministic DOM cross\-check, and blinded human preference reduce this coupling without eliminating evaluator dependence\. The 45\-task human study remains moderate in size, and subjective visual harmony is still difficult to automate\.

## References

- \[1\]\(2024\)Enhancing presentation slide generation by llms with a multi\-staged end\-to\-end approach\.arXiv preprint arXiv:2406\.06556\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1)\.
- \[2\]R\. A\. Bradley and M\. E\. Terry\(1952\)Rank analysis of incomplete block designs: i\. the method of paired comparisons\.Biometrika39\(3/4\),pp\. 324–345\.Cited by:[Appendix D](https://arxiv.org/html/2609.00194#A4.p2.1)\.
- \[3\]X\. Chen, J\. Zhu, P\. Li, H\. Wang, S\. Yang, and M\. Guo\(2026\)PresentBench: a fine\-grained rubric\-based benchmark for slide generation\.arXiv preprint arXiv:2603\.07244\.Cited by:[Table 10](https://arxiv.org/html/2609.00194#A11.T10.28.1.2.1),[Appendix K](https://arxiv.org/html/2609.00194#A11.p4.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p1.1),[§4\.5](https://arxiv.org/html/2609.00194#S4.SS5.p1.1)\.
- \[4\]X\. Chen, M\. Lin, N\. Schärli, and D\. Zhou\(2024\)Teaching large language models to self\-debug\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Link](https://arxiv.org/abs/2304.05128)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[5\]W\. Chiang, L\. Zheng, Y\. Sheng, A\. N\. Angelopoulos, T\. Li, D\. Li, H\. Zhang, B\. Zhu, M\. Jordan, J\. E\. Gonzalez, and I\. Stoica\(2024\)Chatbot arena: an open platform for evaluating LLMs by human preference\.InInternational Conference on Machine Learning \(ICML\),External Links:[Link](https://arxiv.org/abs/2403.04132)Cited by:[Appendix D](https://arxiv.org/html/2609.00194#A4.p2.1),[Appendix I](https://arxiv.org/html/2609.00194#A9.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[6\]W\. S\. Cleveland and R\. McGill\(1984\)Graphical perception: theory, experimentation, and application to the development of graphical methods\.Journal of the American Statistical Association79\(387\),pp\. 531–554\.Cited by:[Appendix L](https://arxiv.org/html/2609.00194#A12.p1.1)\.
- \[7\]J\. Cohen\(1960\)A coefficient of agreement for nominal scales\.Educational and Psychological Measurement20\(1\),pp\. 37–46\.Cited by:[Appendix J](https://arxiv.org/html/2609.00194#A10.p5.1)\.
- \[8\]W\. Feng, W\. Zhu, T\. Fu, V\. Jampani, A\. Akula, X\. He, S\. Basu, X\. E\. Wang, and W\. Y\. Wang\(2023\)LayoutGPT: compositional visual planning and generation with large language models\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:[Link](https://arxiv.org/abs/2305.15393)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1),[§3\.1](https://arxiv.org/html/2609.00194#S3.SS1.p3.1)\.
- \[9\]T\. Fu, W\. Y\. Wang, D\. McDuff, and Y\. Song\(2022\)DOC2PPT: automatic presentation slides generation from scientific documents\.InAAAI Conference on Artificial Intelligence,External Links:[Link](https://arxiv.org/abs/2101.11796)Cited by:[Appendix K](https://arxiv.org/html/2609.00194#A11.p1.1),[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p1.1)\.
- \[10\]J\. Ge, Z\. Z\. Wang, X\. Zhou, Y\. Peng, S\. Subramanian, Q\. Tan, M\. Sap,et al\.\(2025\)AutoPresent: designing structured visuals from scratch\.InCVPR,Cited by:[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[11\]Z\. Gou, Z\. Shao, Y\. Gong, Y\. Shen, Y\. Yang, N\. Duan, and W\. Chen\(2024\)CRITIC: large language models can self\-correct with tool\-interactive critiquing\.ICLR\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[12\]Y\. Guo, Z\. Zhang, Y\. Liang, D\. Zhao, and N\. Duan\(2023\)PPTC benchmark: evaluating large language models for powerpoint task completion\.InNeurIPS Datasets and Benchmarks Track,External Links:[Link](https://arxiv.org/abs/2311.01767)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[13\]Z\. Guo, W\. Liu, M\. Xie, J\. Xu, Z\. Huang, M\. Tian, J\. Xu, Y\. Shen, Q\. Qian, M\. Wu, X\. Wang, C\. Lv, H\. Wang, H\. Yao, X\. Zheng, and X\. Huang\(2025\)RECAST: expanding the boundaries of llms’ complex instruction following with multi\-constraint data\.arXiv preprint arXiv:2505\.19030\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[14\]K\. Gupta, J\. Lazarow, A\. Achille, L\. S\. Davis, V\. Mahadevan, and A\. Shrivastava\(2021\)LayoutTransformer: layout generation and completion with self\-attention\.InIEEE/CVF International Conference on Computer Vision \(ICCV\),External Links:[Link](https://arxiv.org/abs/2006.14615)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1)\.
- \[15\]S\. Hong, M\. Zhuge, J\. Chen, X\. Zheng, Y\. Cheng, C\. Zhang, J\. Wang, Z\. Wang, S\. K\. S\. Yau, Z\. Lin, L\. Zhou, C\. Ran, L\. Xiao, C\. Wu, and J\. Schmidhuber\(2024\)MetaGPT: meta programming for a multi\-agent collaborative framework\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Link](https://arxiv.org/abs/2308.00352)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[16\]W\. Hong, W\. Wang, Q\. Lv, J\. Xu, W\. Yu, J\. Ji, Y\. Wang, Z\. Wang, Y\. Dong, M\. Ding, and J\. Tang\(2024\)CogAgent: a visual language model for GUI agents\.InIEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),External Links:[Link](https://arxiv.org/abs/2312.08914)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[17\]D\. Huang, J\. M\. Zhang, M\. Luck, Q\. Bu, Y\. Qing, and H\. Cui\(2024\)AgentCoder: multi\-agent\-based code generation with iterative testing and optimisation\.arXiv preprint arXiv:2312\.13010\.External Links:[Link](https://arxiv.org/abs/2312.13010)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[18\]J\. Huang, X\. Chen, S\. Mishra, H\. S\. Zheng, A\. W\. Yu, X\. Song, and D\. Zhou\(2024\)Large language models cannot self\-correct reasoning yet\.ICLR\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[19\]N\. Inoue, K\. Kikuchi, E\. Simo\-Serra, M\. Otani, and K\. Yamaguchi\(2023\)LayoutDM: discrete diffusion model for controllable layout generation\.InIEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),External Links:[Link](https://arxiv.org/abs/2303.08137)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1)\.
- \[20\]D\. Jang, M\. L\. Heisler, L\. Xing, Y\. Li, E\. Wang, Y\. Xiong, Y\. Zhang, and Z\. Fan\(2026\)DECKBench: benchmarking multi\-agent frameworks for academic slide generation and editing\.arXiv preprint arXiv:2602\.13318\.Cited by:[Table 10](https://arxiv.org/html/2609.00194#A11.T10.28.1.5.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[21\]C\. E\. Jimenez, J\. Yang, A\. Wettig, S\. Yao, K\. Pei, O\. Press, and K\. Narasimhan\(2024\)SWE\-bench: can language models resolve real\-world GitHub issues?\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Link](https://arxiv.org/abs/2310.06770)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[22\]M\. G\. Kendall\(1938\)A new measure of rank correlation\.Biometrika30\(1/2\),pp\. 81–93\.Cited by:[§4\.5](https://arxiv.org/html/2609.00194#S4.SS5.p1.1)\.
- \[23\]S\. Kim, J\. Suk, S\. Longpre, B\. Y\. Lin, J\. Shin, S\. Welleck, G\. Neubig, M\. Lee, K\. Lee, and M\. Seo\(2024\)Prometheus 2: an open source language model specialized in evaluating other language models\.InConference on Empirical Methods in Natural Language Processing \(EMNLP\),External Links:[Link](https://arxiv.org/abs/2405.01535)Cited by:[Appendix I](https://arxiv.org/html/2609.00194#A9.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[24\]S\. M\. Kosslyn\(2007\)Clear and to the point: 8 psychological principles for compelling PowerPoint presentations\.Oxford University Press\.Cited by:[Appendix L](https://arxiv.org/html/2609.00194#A12.p1.1)\.
- \[25\]K\. Krippendorff\(2004\)Content analysis: an introduction to its methodology\.2 edition,Sage Publications\.Cited by:[Appendix C](https://arxiv.org/html/2609.00194#A3.p4.1)\.
- \[26\]H\. Laurençon, L\. Tronchon, and V\. Sanh\(2024\)Unlocking the conversion of web screenshots into HTML code with the WebSight dataset\.arXiv preprint arXiv:2403\.09029\.External Links:[Link](https://arxiv.org/abs/2403.09029)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1)\.
- \[27\]Y\. Li, Z\. Zeng, Z\. Zhou, X\. Gao, M\. Tian, Y\. Yang, M\. Cheng, Q\. Dai, Y\. Yang, L\. Qiu, Z\. Wang, Z\. Yang, L\. Wang, J\. Li, and C\. Luo\(2026\)BizGenEval: a systematic benchmark for commercial visual content generation\.arXiv preprint arXiv:2603\.25732\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[28\]X\. Liang, X\. Zhang, Y\. Xu, S\. Sun, and C\. You\(2025\)SlideGen: collaborative multimodal agents for scientific slide generation\.arXiv preprint arXiv:2512\.04529\.Cited by:[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p1.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p2.1)\.
- \[29\]W\. Lidwell, K\. Holden, and J\. Butler\(2010\)Universal principles of design, revised and updated\.Rockport Publishers\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[30\]H\. Lightman, V\. Kosaraju, Y\. Burda, H\. Edwards, B\. Baker, T\. Lee, J\. Leike, J\. Schulman, I\. Sutskever, and K\. Cobbe\(2024\)Let’s verify step by step\.InICLR,Cited by:[§3\.2](https://arxiv.org/html/2609.00194#S3.SS2.p1.1)\.
- \[31\]X\. Liu, H\. Yu, H\. Zhang, Y\. Xu, X\. Lei, H\. Lai, Y\. Gu, H\. Ding, K\. Men, K\. Yang,et al\.\(2024\)AgentBench: evaluating llms as agents\.ICLR\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[32\]C\. Lv, J\. Zhou, W\. Zhao, J\. Xu, Z\. Huang, M\. Tian, S\. Dou, T\. Gui, L\. Tian, X\. Zhou, X\. Zheng, X\. Huang, and J\. Zhou\(2026\)Learning query\-specific rubrics from human preferences for deepresearch report generation\.arXiv preprint arXiv:2602\.03619\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[33\]A\. Madaan, N\. Tandon, P\. Gupta, S\. Hallinan, L\. Gao, S\. Wiegreffe, U\. Alon, N\. Dziri, S\. Prabhumoye, Y\. Yang,et al\.\(2023\)Self\-refine: iterative refinement with self\-feedback\.InNeurIPS,Cited by:[Appendix R](https://arxiv.org/html/2609.00194#A18.p2.1),[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p2.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p2.1)\.
- \[34\]R\. E\. Mayer\(2009\)Multimedia learning\.2 edition,Cambridge University Press\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[35\]R\. E\. Mayer \(Ed\.\)\(2014\)The cambridge handbook of multimedia learning\.2 edition,Cambridge University Press\.Cited by:[Appendix L](https://arxiv.org/html/2609.00194#A12.p1.1)\.
- \[36\]T\. C\. Ozden, S\. VS, F\. Horoz, O\. Kara, J\. Kim, and J\. M\. Rehg\(2026\)Narrative\-driven paper\-to\-slide generation via arcdeck\.arXiv preprint arXiv:2604\.11969\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[37\]Y\. Pan, C\. Hu, X\. Huang, C\. Huang, M\. Zhao, Y\. Bi, X\. Zhang, A\. Zeng, and L\. Hu\(2026\)AeSlides: incentivizing aesthetic layout in llm\-based slide generation via verifiable rewards\.arXiv preprint arXiv:2604\.22840\.Cited by:[Table 10](https://arxiv.org/html/2609.00194#A11.T10.28.1.4.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[38\]Q\. Qian, C\. Huang, J\. Xu, C\. Lv, M\. Wu, W\. Liu, X\. Wang, Z\. Wang, Z\. Huang, M\. Tian, J\. Xu, K\. Hu, H\. Wang, Y\. Hu, X\. Huang, and X\. Zheng\(2026\)Benchmark2\{\}^\{2\}: systematic evaluation of llm benchmarks\.arXiv preprint arXiv:2601\.03986\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[39\]G\. Reynolds\(2008\)Presentation Zen: simple ideas on presentation design and delivery\.New Riders\.Cited by:[Appendix L](https://arxiv.org/html/2609.00194#A12.p1.1)\.
- \[40\]N\. Shinn, F\. Cassano, A\. Gopinath, K\. Narasimhan, and S\. Yao\(2023\)Reflexion: language agents with verbal reinforcement learning\.InNeurIPS,Cited by:[Appendix R](https://arxiv.org/html/2609.00194#A18.p2.1),[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p2.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p2.1)\.
- \[41\]C\. Si, Y\. Zhang, R\. Li, Z\. Yang, R\. Liu, and D\. Yang\(2025\)Design2Code: benchmarking multimodal code generation for automated front\-end engineering\.InConference of the North American Chapter of the Association for Computational Linguistics \(NAACL\),External Links:[Link](https://arxiv.org/abs/2403.03163)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1),[§3\.1](https://arxiv.org/html/2609.00194#S3.SS1.p3.1)\.
- \[42\]E\. Sun, Y\. Hou, D\. Wang, Y\. Zhang, and N\. X\. R\. Wang\(2021\)D2S: document\-to\-slide generation via query\-based text summarization\.InNAACL,Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1)\.
- \[43\]R\. Tanaka, K\. Nishida, K\. Nishida, T\. Hasegawa, I\. Saito, and K\. Saito\(2023\)SlideVQA: a dataset for document visual question answering on multiple images\.InAAAI Conference on Artificial Intelligence,External Links:[Link](https://arxiv.org/abs/2301.04883)Cited by:[Appendix K](https://arxiv.org/html/2609.00194#A11.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[44\]E\. R\. Tufte\(1983\)The visual display of quantitative information\.Graphics Press\.Cited by:[Appendix L](https://arxiv.org/html/2609.00194#A12.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[45\]G\. Wang, Y\. Xie, Y\. Jiang, A\. Mandlekar, C\. Xiao, Y\. Zhu, L\. Fan, and A\. Anandkumar\(2024\)Voyager: an open\-ended embodied agent with large language models\.Transactions on Machine Learning Research \(TMLR\)\.External Links:[Link](https://arxiv.org/abs/2305.16291)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[46\]P\. Wang, L\. Li, Z\. Shao, R\. X\. Xu, D\. Dai, Y\. Li, D\. Chen, Y\. Wu, and Z\. Sui\(2024\)Math\-Shepherd: verify and reinforce LLMs step\-by\-step without human annotations\.InAnnual Meeting of the Association for Computational Linguistics \(ACL\),External Links:[Link](https://arxiv.org/abs/2312.08935)Cited by:[§3\.2](https://arxiv.org/html/2609.00194#S3.SS2.p1.1)\.
- \[47\]X\. Wang, M\. Tian, Y\. Zeng, Z\. Huang, J\. Yuan, B\. Chen, J\. Xu, M\. Zhou, W\. Liu, M\. Wu,et al\.\(2026\)Reward hacking in the era of large models: mechanisms, emergent misalignment, challenges\.arXiv preprint arXiv:2604\.13602\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[48\]X\. Wang, B\. Li, Y\. Song, F\. F\. Xu, X\. Tang, M\. Zhuge, J\. Pan, Y\. Song, B\. Li, J\. Singh,et al\.\(2025\)OpenHands: an open platform for AI software developers as generalist agents\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Link](https://arxiv.org/abs/2407.16741)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[49\]J\. Wei, X\. Wang, D\. Schuurmans, M\. Bosma, B\. Ichter, F\. Xia, E\. Chi, Q\. V\. Le, and D\. Zhou\(2022\)Chain\-of\-thought prompting elicits reasoning in large language models\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:[Link](https://arxiv.org/abs/2201.11903)Cited by:[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[50\]M\. Wertheimer\(1938\)Laws of organization in perceptual forms\.InA Source Book of Gestalt Psychology,W\. D\. Ellis \(Ed\.\),pp\. 71–88\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[51\]E\. B\. Wilson\(1927\)Probable inference, the law of succession, and statistical inference\.Journal of the American Statistical Association22\(158\),pp\. 209–212\.Cited by:[Appendix D](https://arxiv.org/html/2609.00194#A4.p5.1)\.
- \[52\]Q\. Wu, G\. Bansal, J\. Zhang, Y\. Wu, B\. Li, E\. Zhu, L\. Jiang, X\. Zhang, S\. Zhang, J\. Liu, A\. H\. Awadallah, R\. W\. White, D\. Burger, and C\. Wang\(2024\)AutoGen: enabling next\-gen LLM applications via multi\-agent conversation\.InConference on Language Modeling \(COLM\),External Links:[Link](https://arxiv.org/abs/2308.08155)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[53\]T\. Xiong, X\. Wang, D\. Guo, Q\. Ye, H\. Fan, Q\. Gu, H\. Huang, and C\. Li\(2025\)LLaVA\-Critic: learning to evaluate multimodal models\.InIEEE/CVF Conference on Computer Vision and Pattern Recognition \(CVPR\),External Links:[Link](https://arxiv.org/abs/2410.02712)Cited by:[Appendix E](https://arxiv.org/html/2609.00194#A5.p2.1),[Appendix I](https://arxiv.org/html/2609.00194#A9.p1.1)\.
- \[54\]J\. Yang, C\. E\. Jimenez, A\. Wettig, K\. Lieret, S\. Yao, K\. Narasimhan, and O\. Press\(2024\)SWE\-agent: agent\-computer interfaces enable automated software engineering\.NeurIPS\.Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[55\]Y\. Yang, W\. Li, H\. Ren, Z\. Lu, K\. Wang, Z\. Huang,et al\.\(2026\)SlidesGen\-bench: evaluating slides generation via computational and quantitative metrics\.arXiv preprint arXiv:2601\.09487\.Cited by:[Table 10](https://arxiv.org/html/2609.00194#A11.T10.28.1.3.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1)\.
- \[56\]Y\. Yang, W\. Jiang, Y\. Wang, Y\. Song, Y\. Wang, and C\. Zhang\(2025\)Auto\-slides: an interactive multi\-agent system for creating and customizing research presentations\.External Links:2509\.11062,[Link](https://arxiv.org/abs/2509.11062)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1)\.
- \[57\]S\. Yao, D\. Yu, J\. Zhao, I\. Shafran, T\. L\. Griffiths, Y\. Cao, and K\. Narasimhan\(2023\)Tree of thoughts: deliberate problem solving with large language models\.InAdvances in Neural Information Processing Systems \(NeurIPS\),External Links:[Link](https://arxiv.org/abs/2305.10601)Cited by:[§3\.1](https://arxiv.org/html/2609.00194#S3.SS1.p3.1)\.
- \[58\]S\. Yao, J\. Zhao, D\. Yu, N\. Du, I\. Shafran, K\. Narasimhan, and Y\. Cao\(2023\)ReAct: synergizing reasoning and acting in language models\.InInternational Conference on Learning Representations \(ICLR\),External Links:[Link](https://arxiv.org/abs/2210.03629)Cited by:[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p2.1),[§3\.4](https://arxiv.org/html/2609.00194#S3.SS4.p1.1)\.
- \[59\]W\. Zeng, M\. Ouyang, L\. Cui, and H\. T\. Ng\(2025\)SlideTailor: personalized presentation slide generation for scientific papers\.arXiv preprint arXiv:2512\.20292\.Note:AAAI 2026Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p1.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p2.1)\.
- \[60\]B\. Zheng, B\. Gou, J\. Kil, H\. Sun, and Y\. Su\(2024\)GPT\-4V\(ision\) is a generalist web agent, if grounded\.InInternational Conference on Machine Learning \(ICML\),External Links:[Link](https://arxiv.org/abs/2401.01614)Cited by:[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.
- \[61\]H\. Zheng, X\. Guan, H\. Kong, J\. Zheng, W\. Zhou, H\. Lin, Y\. Lu, B\. He, X\. Han, and L\. Sun\(2025\)PPTAgent: generating and evaluating presentations beyond text\-to\-slides\.InEMNLP,External Links:[Link](https://arxiv.org/abs/2501.03936)Cited by:[Table 10](https://arxiv.org/html/2609.00194#A11.T10.28.1.6.1),[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p2.1)\.
- \[62\]H\. Zheng, G\. Mo, X\. Yan, Q\. Yuan, W\. Zhang, X\. Chen, Y\. Lu, H\. Lin, X\. Han, and L\. Sun\(2026\)DeepPresenter: environment\-grounded reflection for agentic presentation generation\.arXiv preprint arXiv:2602\.22839\.Note:ACL 2026Cited by:[Appendix K](https://arxiv.org/html/2609.00194#A11.p4.1),[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p2.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p1.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p2.1)\.
- \[63\]L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. P\. Xing, H\. Zhang, J\. E\. Gonzalez, and I\. Stoica\(2023\)Judging LLM\-as\-a\-judge with MT\-Bench and chatbot arena\.InAdvances in Neural Information Processing Systems \(NeurIPS\) Datasets and Benchmarks Track,External Links:[Link](https://arxiv.org/abs/2306.05685)Cited by:[Appendix I](https://arxiv.org/html/2609.00194#A9.p1.1),[§2](https://arxiv.org/html/2609.00194#S2.p3.1),[§4\.1](https://arxiv.org/html/2609.00194#S4.SS1.p2.1)\.
- \[64\]S\. Zhou, F\. F\. Xu, H\. Zhu, X\. Zhou, R\. Lo, A\. Sridhar, X\. Cheng, Y\. Bisk, D\. Fried, U\. Alon,et al\.\(2024\)WebArena: a realistic web environment for building autonomous agents\.ICLR\.Cited by:[§1](https://arxiv.org/html/2609.00194#S1.p2.1),[§2](https://arxiv.org/html/2609.00194#S2.p2.1)\.

## Appendix AInitial\-Draft Pipeline and Rendering Environment

This appendix documents two implementation layers that sit*outside*theReDeckrefinement loop: \(i\) the bootstrap pipeline that produces its session\-start deckd0d\_\{0\}, and \(ii\) the rendering environment that turns editable deck state into rendered geometryrtr\_\{t\}and diagnosticsCtC\_\{t\}\. The refinement loop itself — the compositional action space, Step\-Level Render Feedback, the Adaptive Deck Critic, and the submission gateρ\\rho— is specified in §[3](https://arxiv.org/html/2609.00194#S3)\. EveryReDeckablation in Sec\.[4\.3](https://arxiv.org/html/2609.00194#S4.SS3)starts from the samed0d\_\{0\}and uses the same renderer and extractor\. Published baselines use their native initial\-generation pipelines in Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1); App\.[N](https://arxiv.org/html/2609.00194#A14)reports the corresponding initial\-deck attribution analysis\.

Bootstrap: producing the session\-start deckd0d\_\{0\}\.Given a source paper, the bootstrap runs four ordered steps and emits a complete HTML/CSS deck that becomesd0d\_\{0\}in the notation of §[3\.1](https://arxiv.org/html/2609.00194#S3.SS1)\.*\(i\) Source parsing*converts the paper \(PDF or LaTeX source\) into structured sections via layout analysis combined with an LLM section classifier; tables, figures, and equations are extracted separately and stored as referenceable entities with their captions and source locations\.*\(ii\) Blueprint planning*issues a single LLM call that reads the parsed structure and emits a JSON*deck blueprint*that assigns each slide a role \(title, outline, content, figure, comparison, conclusion\) and selects the source spans that populate it\.*\(iii\) Layout selection*maps each blueprint entry to one template from a library of 15 base layouts \(single\-column, two\-column, figure\-dominant, table\-dominant, etc\.\), conditioned on the slide role, content density, and figure count from the blueprint\.*\(iv\) Code generation*issues one code\-generation LLM call per slide that consumes the blueprint entry, layout template, and source spans; per\-slide independence makes this step embarrassingly parallel across the deck\. The result is the deck sourced0d\_\{0\}that theReDeckfamily of variants in Sec\.[4\.3](https://arxiv.org/html/2609.00194#S4.SS3)shares as its starting state\. Anything labelled “no refinement” or “T0T\_\{0\}” in the experiments refers to thisd0d\_\{0\}rendered without any loop applied on top\. Published baselines \(SlideGen,SlideTailor,DeepPresenter\) are run on their released code under the matched per\-task LLM\-call\-count cap and use their own native initial\-generation pipeline\.

Rendering environment: producingrtr\_\{t\}andCtC\_\{t\}\.Generated HTML slides are rendered to PNG via a headless browser \(Playwright/Chromium at1280×7201280\\times 720, the same canvas as the case study in App\.[B](https://arxiv.org/html/2609.00194#A2)\)\. A DOM\-based spatial extractor traverses the rendered DOM tree and computes bounding boxes, an overlap matrix, overflow measurements, and off\-canvas indicators; together with the compiler diagnosticsCtC\_\{t\}collected during rendering, these are the rendered\-side components of the hidden statest=\(dt,rt,Ct\)s\_\{t\}=\(d\_\{t\},r\_\{t\},C\_\{t\}\)defined in §[3\.1](https://arxiv.org/html/2609.00194#S3.SS1)\. Step\-Level Render Feedback \(§[3\.2](https://arxiv.org/html/2609.00194#S3.SS2)\) reads from this extractor to construct the observationot=\(Δviol​\(dt,d0\),At\)o\_\{t\}=\(\\Delta\_\{\\text\{viol\}\}\(d\_\{t\},d\_\{0\}\),\\,A\_\{t\}\)delivered after every atomic action, and the Adaptive Deck Critic \(§[3\.3](https://arxiv.org/html/2609.00194#S3.SS3)\) reads from the same extractor \(alongside the rendered PNGs\) when its deterministic\-geometry probes updateLt\+1L\_\{t\+1\}\. For finalSCRreporting, all systems are rendered to PNG and evaluated by the same SpatialCheck VLM judge \(App\.[E](https://arxiv.org/html/2609.00194#A5)\); the DOM extractor is used forReDeck’s internal step\-level feedback rather than as the cross\-systemSCRscorer\.

## Appendix BStep\-Level Feedback Case Study

Fig\.[6](https://arxiv.org/html/2609.00194#A2.F6)records two logged repair trajectories from the same starting slide \(a scaling\-laws content slide rendered in the Slate Modern theme\), differing only in whether step\-level feedback is enabled\. The initial states0s\_\{0\}triggers three independent hard layout issues inBgeomB\_\{\\text\{geom\}\}at once:i1i\_\{1\}a low\-contrast title \(foreground / background lift below readability threshold\),i2i\_\{2\}a bullet column whose right edge overlaps the chart container, andi3i\_\{3\}a takeaway callout positioned with𝚕𝚎𝚏𝚝=940\\mathtt\{left\}=940,𝚠𝚒𝚍𝚝𝚑=400\\mathtt\{width\}=400on a12801280\-wide canvas, so its right edge lies6060px off\-canvas\. The left column shows the trajectory without step\-level feedback: the agent receives rendered critique only at turn boundaries, so multiple edits can accumulate before any visual feedback is observed\. The right column showsReDeckwith step\-level render feedback on the same state, where each atomic edit is rendered and checked before the trajectory advances\.

Without step\-level render feedback \(left\)\.The baseline receives a turn\-boundary verdict \(bullet/chart overlap; callout looks lost at the bottom\) and applies both suggestions in a single batch ase1′e\_\{1\}^\{\\prime\}: narrows the bullet column from800800to580580px \(well intentioned, addressesi2i\_\{2\}\) and simultaneously raises the callout fromy=540y=540toy=300y=300to “increase visibility\.” Without per\-step rendering it does not see that the new callout position now sits on top of the chart atx=−1240x=940\\\!\-\\\!1240,y=−480y=300\\\!\-\\\!480\. The next critic verdict \(callout overlaps chart\) drivese2′e\_\{2\}^\{\\prime\}: enlarge the chart to720720px wide so the callout is no longer in front of it; the new chart right edge is atx=1340x=1340, off\-canvas, but the agent has no way to see this\. A third verdict \(lower\-left feels empty, add metrics\) drivese3′e\_\{3\}^\{\\prime\}: a three\-card metric strip is inserted at𝚝𝚘𝚙=470\\mathtt\{top\}=470, colliding with the bullet text that already extends toy≈480y\\\!\\approx\\\!480\. After three well\-intentioned edits the deck has*five*hard issues rather than three, because no observation interrupts the regression chain inside the turn\. In the ablation logs, the same pattern appears in Case0202: the initial deck has2424issues, and after one NL\-critique repair turn without rendered feedback the count grows to3131as text\-motivated edits trigger cascading overflow and density regressions before the next turn boundary can report them\.

With step\-level render feedback \(right\)\.Within turn00ReDeckissues four atomic edits\.e1e\_\{1\}narrows the bullet column to580580px alone \(no batched callout move\); render feedback compares the post\-edit render againsts0s\_\{0\}, observes no new hard issue, and accepts \(s0→s1s\_\{0\}\\\!\\to\\\!s\_\{1\}\)\.e2e\_\{2\}then attempts to widen the chart to720720px to use the freed horizontal room; the new chart right edge lands atx=1340x=1340, breaching the canvas, and step\-level render feedback reports the regression immediately; the agent choosesrollback, returning the deck tos1s\_\{1\}before the trajectory continues \(the inset on the right column shows the rolled\-back attempt struck through with a red cross\)\.e3e\_\{3\}takes a different shape: the takeaway becomes a full\-width bottom band \(𝚕𝚎𝚏𝚝=40\\mathtt\{left\}=40,𝚠𝚒𝚍𝚝𝚑=1200\\mathtt\{width\}=1200\), resolvingi3i\_\{3\}\.e4e\_\{4\}darkens the title text from𝚛𝚐𝚋⁡\(180,188,195\)\\mathtt\{rgb\}\(180,188,195\)to the Slate Modern primary𝚛𝚐𝚋⁡\(38,50,56\)\\mathtt\{rgb\}\(38,50,56\), resolvingi1i\_\{1\}\. At the turn\-00boundary the deck critic confirmsBgeomB\_\{\\text\{geom\}\}as clean and raises a newAnarrativeA\_\{\\text\{narrative\}\}finding: the chart alone is hard to parse in three seconds\. Turn 1 contains a single edit:e5e\_\{5\}inserts a three\-card metric strip below the chart that surfaces the key numbers\. The final rendered state retains zero hard geometric issues after the narrative repair\.

What the comparison shows\.The same draft and the same kind of well\-intentioned edits produce opposite outcomes once step\-level render feedback is removed\. Three behaviours are visible in the contrast: \(i\) step\-level render feedback catches the off\-canvas chart \(e2e\_\{2\}\) one edit after it appears, so detection costs one rollback rather than a multi\-edit pile\-up; \(ii\) rollback returns the trajectory to the most recent rendered\-clean state, so the agent can try a different edit shape \(e3e\_\{3\}\) instead of retrying the same parameter; \(iii\) the deck critic and step\-level render feedback observe disjoint signals — the critic raises a turn\-boundary narrative finding the geometry channel cannot see \(e5e\_\{5\}\), and the next step\-level check confirms that the narrative repair did not silently regress geometry\.

Figure 6:A real repair trajectory with and without step\-level render feedback\.Both columns start from the same draft with three hard spatial issues\.Left: without step\-level feedback, well\-intentioned edits silently introduce new issues \(count grows→53\\\!\\to\\\!5\)\.Right:ReDeck’s per\-edit observations catch the off\-canvas regression immediately; the agent rolls back and retries a different edit shape, converging to zero hard issues\.
## Appendix CDeckQuizHuman Alignment Study

This appendix calibrates the fourDeckQuizjudges against human annotation at the*item*level \(per quiz question, per slide\), distinct from the*deck\-level*forced\-choice study in App\.[D](https://arxiv.org/html/2609.00194#A4)\. Its purpose is to localise any deck\-level ranking disagreement to the responsible module rather than to the suite as a whole\.

Items\.We sample 10 evaluation items per module \(40 items in total\): 10 ContentQuiz questions stratified across the four QA types \(3 contribution, 3 method, 2 experiment, 2 limitation\), 10 SpatialCheck slides covering the four hard\-violation families, 10 Aesthetics slides spanning the five design axes, and 10 DeckDesign slides spanning the five principles\. Items are drawn from a held\-out subset of the GPT\-5\.4 main split that the judge prompts have never seen during development\.

Annotators and labelling\.Three annotators with prior slide\-design experience \(none involved in probe or rubric design\) label each item independently using the same rubric the judge consumes\. ContentQuiz items receive a 0/1 correctness label; SpatialCheck items receive a per\-family violation count; Aesthetics items receive a00–55ordinal score on each design axis; DeckDesign items receive a 0–3 Likert score\. Total labelling load is40×3=12040\\times 3=120judgments\.

Statistics\.For each module we report \(i\) Krippendorff’sα\\alpha\[[25](https://arxiv.org/html/2609.00194#bib.bib60)\]across the three annotators \(inter\-rater agreement\) and \(ii\) Krippendorff’sα\\alphabetween the judge and the annotator majority \(judge–human alignment\)\. We useαjudge–human≥0\.5\\alpha\_\{\\text\{judge\-\-human\}\}\\geq 0\.5as the minimum reliability threshold; modules below threshold are demoted from Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)to a diagnostic role\.

Table 3:DeckQuizper\-module judge–human alignment on the 10\-item\-per\-module calibration subset \(40 items total, 3 annotators, 120 judgments\)\. Modules below the specifiedαjudge–human≥0\.5\\alpha\_\{\\text\{judge\-\-human\}\}\\geq 0\.5threshold would be demoted to diagnostic use\.Fig\.[7](https://arxiv.org/html/2609.00194#A3.F7)pairs the inter\-raterαannotator\\alpha\_\{\\text\{annotator\}\}with the judge–humanαjudge–human\\alpha\_\{\\text\{judge\-\-human\}\}per module so that both reliability of the human reference and alignment of the automated judge against it can be read at once; the specified threshold runs through the chart as a horizontal reference, and any module whose judge–human bar drops below it is annotated as a demotion candidate\.

Figure 7:DeckQuizjudge–human alignment per module\.Light grey bar: inter\-rater Krippendorff’sα\\alphaacross the three annotators on a given module’s 10\-item subset\. Navy bar:α\\alphabetween the automated judge and the annotator majority on the same items\. The dashed reference atα=0\.50\\alpha=0\.50is the specified demotion threshold; on this calibration subset all four modules clear the threshold \(the lowest,Aestheticsatαjudge–human=0\.57\\alpha\_\{\\text\{judge\-\-human\}\}=0\.57, sits closest to it and is the first candidate for demotion if a future re\-calibration drops it below0\.500\.50\)\.
## Appendix DDeck\-Level Human Pairwise Study

This appendix details the forced\-choice deck\-level study referenced in Sec\.[4\.5](https://arxiv.org/html/2609.00194#S4.SS5)\. Its purpose is external validation against the strongest baseline, distinct from the per\-item judge calibration in App\.[C](https://arxiv.org/html/2609.00194#A3)\.

Items and pairing\.A 45\-task subset is drawn from the 100\-task GPT\-5\.4 split\. For each task we form one forced\-choice pair\[[2](https://arxiv.org/html/2609.00194#bib.bib63),[5](https://arxiv.org/html/2609.00194#bib.bib45)\],ReDeckversus DeepPresenter, yielding270270individual judgments across two dimensions and three annotators\.

Annotators and blinding\.Three annotators with prior slide\-design or HCI experience, none of whom contributed to the probe library, critic prompt, orDeckQuizrubric design\. Decks are rendered to PNG at the same1280×7201280\\times 720canvas used by the loop \(App\.[A](https://arxiv.org/html/2609.00194#A1)\) and then up\-sampled to a fixed\-DPI PDF for printing; file names are stripped of system identity, left/right order is randomised independently per pair, and the same pair is never shown twice in a row to the same annotator\.

Task and dimensions\.Each annotator answers two forced\-choice questions per pair: \(D1\)*which deck better presents the source paper overall*\(presentation quality\), and \(D2\)*which deck is more faithful to the source paper*\(faithfulness\)\. A third "no preference" option is allowed but counted as0\.50\.5for each side; cases with≥2/3\\geq 2/3"no preference" are reported separately and excluded from the ranking\.

Statistics\.We aggregate the three ratings by task\-level majority vote and reportReDeck’s preference rate with Wilson 95% CIs\[[51](https://arxiv.org/html/2609.00194#bib.bib62)\]\. No task has a majority “no preference” outcome\.

Table 4:Blinded human preference forReDeckover DeepPresenter on 45 tasks \(3 annotators per task, majority vote\)\.
## Appendix ESpatialCheck VLM\-Judge Validation

This appendix calibrates the vision\-language judge that produces the SpatialCheck*clean\-rate*\(SCR\) reported in Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1), so that SCR differences between systems can be attributed to deck quality rather than judge noise\. Unlike the DOM\-based extractor used insideReDeck’s step\-level render feedback \(App\.[A](https://arxiv.org/html/2609.00194#A1)\), which operates only on HTML/CSS source, the SCR judge consumes the rendered PNG of each slide and therefore applies uniformly to PPTX, PDF, and HTML decks across all comparison systems\.

Judge\.A GPT\-5\.4 vision model receives one rendered slide at a time together with a fixed rubric enumerating four hard\-violation families and returns a strict\-JSON list of per\-family violations; this follows the rendered\-screenshot vision\-judge paradigm developed for multimodal evaluation\[[53](https://arxiv.org/html/2609.00194#bib.bib47)\]\. The four families areS1 text\_clipping\(visible text cut off by a container or canvas edge\),S2 element\_overlap\(two or more readable elements occlude one another\),S3 out\_of\_bounds\(an element lies partially or wholly outside the slide canvas\), andS4 image\_clipping\(an image is cropped by its bounding container so that semantic content is lost\)\. A slide counts as*clean*iff its violation list is empty across all four families; SCR is the fraction of clean slides over the deck\.

Sample\.We sample 150 rendered slides from the shared 100\-paper task split, stratified across systems, agent backbones, and the four hard\-violation families\.

Reference labels\.Two non\-author raters independently label each slide and adjudicate disagreements before seeing the VLM predictions\.

Metrics\.We report inter\-rater Cohen’sκ\\kappaand VLM sensitivity, specificity, andF1F\_\{1\}against the adjudicated reference\. Dirty/violation is the positive class\.

Table 5:SpatialCheck validation against an adjudicated non\-author reference on 150 slides\.The primary clean/dirty balanced accuracy is0\.890\.89\. As a deterministic consistency check, we also score all six HTML variants with the DOM extractor used by the step\-level channel\. Across 100 tasks×\\times3 seeds×\\times6 variants, DOM\-SCR preserves the full VLM\-SCR ordering, with 92% slide\-level agreement \(κ=0\.83\\kappa=0\.83\[0\.79, 0\.86\], MCC=0\.84=0\.84\[0\.80, 0\.87\]\), a deck\-level bias of−0\.8\-0\.8points \[−1\.3,−0\.3\-1\.3,\-0\.3\], and MAE1\.41\.4points \[1\.0, 1\.8\]\. This is a second implementation of the same violation taxonomy, not an independent task definition\.

## Appendix FPresentBench Domain\-Wise Results

Full PresentBench results broken down by domain are reported in Fig\.[5](https://arxiv.org/html/2609.00194#S4.F5)\. On the scientific*academia*domain \(which overlaps with our in\-domain training distribution\),ReDeckachieves the highest win rate \(0\.72\); on the most distant domains \(*talk*,*advertising*\), the win rate is lower \(0\.66–0\.68\) but still above chance, indicating partial transfer of the feedback loop’s benefits to non\-scientific content\. Among the three baselines, the pooled order isDeepPresenter\>\>SlideTailor\>\>SlideGen; the only per\-domain deviation is on*talk*, whereSlideTailornarrowly outscoresDeepPresenter\(0\.510\.51vs\.0\.500\.50\), consistent withSlideTailor’s template\-driven design priors transferring better to less content\-dense talk slides thanDeepPresenter’s content\-completeness optimisation\. This single 2nd\-vs\-3rd swap is the sole contributor to the aggregate Kendall’sτ=0\.87\\tau=0\.87reported in Fig\.[5](https://arxiv.org/html/2609.00194#S4.F5)\.

## Appendix GExternal Evaluation

We evaluate the same five systems with two published protocols whose data and scoring are not derived fromDeckQuiz\.DECKBenchuses the common 281 of 294 completed paper–slide pairs;SlidesGen\-Benchuses the common 181 of 187 completed topics\. Three seeds are averaged within each native unit before aggregation\.

Table 6:External evaluation on the common completed sets\. LQ, DFaith, and DFid denoteDECKBenchLayoutQ, DeckFaith, and DeckFid; Quiz and Aes\. areSlidesGen\-Benchmetrics\.Relative to DeepPresenter,ReDeckgains0\.0320\.032\[0\.018, 0\.046\] on LQ,0\.0230\.023\[0\.010, 0\.036\] on DFaith,0\.0220\.022\[0\.008, 0\.036\] on DFid,2\.62\.6percentage points \[0\.8, 4\.4\] on Quiz, and3\.23\.2points \[1\.5, 4\.9\] on Aes\. Quiz measures answerability from extracted slide content rather than complete source fidelity; the results therefore triangulate, rather than duplicate, the fourDeckQuizmodules\.

## Appendix HTransfer to PPTX

For PPTX,ReDeckreplaces browser DOM queries withpython\-pptxshape geometry and renders each state through LibreOffice\. The action loop, critic, rollback, and submission gate are unchanged\.

Table 7:PPTX refinement on GPT\-5\.4 \(100 tasks×\\times3 seeds\)\.The gains are\+5\.0/\+24\.0/\+0\.50/\+6\.0\+5\.0/\+24\.0/\+0\.50/\+6\.0onFid\./SCR/Aes\./Des\.; the SCR gain has a 95% CI of \[19\.1, 28\.9\]\. This supports transfer to a second editable geometry/rendering stack, but does not establish support for PDF/OCR\-only editing, Keynote, animations, masking, or font\-fallback behavior\.

## Appendix IJudge Configuration

All fourDeckQuizVLM judges are run on a single fixed backbone \(GPT\-5\.4, vision\-enabled variant for the image\-reading judgesSCR,Aes\., andDes\.\) regardless of which agent backbone \(GPT\-5\.4, Gemini\-3\.1, Claude\-4\.6\) produced the deck\. This ensures that all systems are scored by the same judge under identical conditions, controlling for judge variation across rows of Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)\. When the agent backbone is also GPT\-5\.4, the judge and the generator share a model family; we treat potential same\-family self\-preference as a residual limitation and validate against human annotations in App\.[C](https://arxiv.org/html/2609.00194#A3)\. This design follows the methodology ofLLM\-as\-a\-judgestudies that document position bias, verbosity bias, and self\-enhancement bias\[[63](https://arxiv.org/html/2609.00194#bib.bib44),[5](https://arxiv.org/html/2609.00194#bib.bib45)\], and aligns with rubric\-based open evaluator practice\[[23](https://arxiv.org/html/2609.00194#bib.bib46),[53](https://arxiv.org/html/2609.00194#bib.bib47)\]\. The ContentQuiz QA grader forFid\.uses the same backbone in text\-only mode against the held\-out QA bank in App\.[K](https://arxiv.org/html/2609.00194#A11)\. Evaluation calls use deterministic decoding \(temperature00, top\-pp11, fixed seed where the API exposes one\) and a strict structured\-JSON output schema; parse failures trigger a single retry, after which the call is recorded as missing and excluded from the corresponding cell mean \(less than0\.3%0\.3\\%of calls in the main run\)\. Agent calls, in contrast, keep each system’s released defaults \(sampling temperature, top\-pp, and tool\-use settings as shipped\) so that we benchmark each baseline at its publicly recommended operating point rather than a re\-tuned variant\. The judge prompt templates are released with the codebase atapp/prompts/\.

## Appendix JPilot Audit, Taxonomy Mapping, and Probe Coverage

This appendix establishes that the six probe families in §[3\.3](https://arxiv.org/html/2609.00194#S3.SS3)\(AAnarrative,BgeomB\_\{\\text\{geom\}\}layout,BvisualB\_\{\\text\{visual\}\}visual style,CCcompleteness,DDcorrectness,EEsource fidelity\) are derived from rather than imposed on the failure distribution, and documents their provenance from external benchmarks\.

External provenance\.Tab\.[8](https://arxiv.org/html/2609.00194#A10.T8)maps eachReDeckprobe family to the external failure dimensions from which it is derived\.

Table 8:External failure taxonomies projected onto theReDeckprobe families\.Sample\.We hand\-label 100 refinement turns drawn at random from preliminaryReDeckruns collected before the probe library used in the main evaluation was finalised\. Each turn contributes one to several discrete failures; the unit of analysis is the*failure*, of which the 100 turns yield 264 instances\.

Schema and inter\-rater agreement\.Two authors independently assign each failure to one of the six families or to an*other*bucket using a written codebook released with the supplementary code\. Disagreements are resolved by adjudication after both passes are complete to avoid leakage between annotators\. We report Cohen’sκ\\kappaover the six\-way label set as inter\-rater agreement\.

Coverage criterion\.The taxonomy is accepted if \(i\) the residual*other*rate is at most5%5\\%of failures and \(ii\) Cohen’sκ≥0\.6\\kappa\\geq 0\.6\[[7](https://arxiv.org/html/2609.00194#bib.bib61)\]\. Failure to meet either criterion forces taxonomy revision before the probe library is committed\.

Table 9:Pilot audit: distribution of 264 hand\-labelled failures over the six probe families and the residual*other*bucket, with two\-author Cohen’sκ\\kappa\.FamilyCountShareAAnarrative3814\.4%BgeomB\_\{\\text\{geom\}\}layout7227\.3%BvisualB\_\{\\text\{visual\}\}visual style4115\.5%CCcompleteness5219\.7%DDcorrectness3111\.7%EEsource fidelity176\.4%*Other*\(residual\)134\.9%Two\-author Cohen’sκ\\kappa0\.78Fig\.[8](https://arxiv.org/html/2609.00194#A10.F8)renders the same distribution as a donut chart with the residual*other*slice exploded for emphasis; the centre of the donut carries the two\-authorκ\\kappatogether with the total failure count, and the acceptance criterion is annotated above the chart so that audit success or failure can be read off without referring back to the table\.

Figure 8:Pilot\-audit failure distribution over the six probe families\.Each ring slice is one probe family, sized by its share of the 264 hand\-labelled failures from the 100\-turn pilot; the small grey wedge labelled*Other*is the residual unclassified bucket, exploded to make its size visible against the≤5%\\leq 5\\%ceiling\. The donut centre reports the two\-author Cohen’sκ\\kappaused as the inter\-rater check; the green badge above the chart confirms that both acceptance criteria \(other≤5%\\textit\{other\}\\leq 5\\%andκ≥0\.60\\kappa\\geq 0\.60\) are satisfied\.
## Appendix KDeckQuizConstruction and Protocol

Source pool\.The 100 evaluation papers are drawn from five disciplines \(CS/ML 25, Bio\-Med 15, Physical\-Sci 25, Econ/Finance 25, Soc\-Sci/HCI 10\)\. Across the full evaluation split, 30 papers are short \(≤\\leq8 pages\), 40 are medium \(9–15 pages\), and 30 are long \(\>\>15 pages\)\. All papers are from arXiv, posted between 2024\-01\-01 and 2026\-03\-31, and are selected to have no pre\-existing official presentation slides so that no system in Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)could have memorised the target deck during pre\-training\. The paper\-grounded QA construction follows the document\-VQA tradition\[[43](https://arxiv.org/html/2609.00194#bib.bib59)\]and the slide\-from\-paper benchmarking lineage\[[9](https://arxiv.org/html/2609.00194#bib.bib58)\], but localises questions to per\-paper extracted facts so that the deck rather than a topic\-level summary is being evaluated\.

Question generation: what should a reader leave the deck knowing?ContentQuiz operationalises a reader\-centric question:*after watching the deck once, can the audience answer the questions a reviewer would ask about this paper?*For each paper we instantiate this with a four\-type taxonomy that mirrors the canonical reviewer rubric:*Contribution*questions probe the paper’s main claims and novelty;*Method*questions probe the proposed approach;*Experiment*questions probe the key results and the conditions under which they hold;*Limitation*questions probe the caveats and scope restrictions\. For each type the generator is conditioned on the paper alone \(not the deck\) and asked to write a question whose answer must be present in a faithful deck; this decouples question difficulty from the system being evaluated and preventsFid\.from rewarding decks that happen to share a stylistic prior with the judge\.

Span\-extraction templates\.Each question is paired with the source span that contains the answer\. Spans are extracted by a template\-guided LLM call that localises the answer to a contiguous passage of≤\\leq200 tokens in the source paper, supporting item\-level grounding and audit of the generated question and gold answer\.

Why quiz, not checklist\.PresentBench\[[3](https://arxiv.org/html/2609.00194#bib.bib1)\]and theDeepPresentereval suite\[[62](https://arxiv.org/html/2609.00194#bib.bib13)\]score content fidelity by a per\-deck*checklist*of binary “is propertyXXpresent?” rubric items\. A checklist conflates two failure modes that an iterative refinement loop must distinguish: \(a\) the deck omits a substantive fact the audience needs, and \(b\) the deck phrases a fact in a form the rubric did not anticipate\. A reader\-centric quiz collapses \(b\) because the question is the canonical phrasing and the four answer choices fix the scoring surface, while \(a\) remains diagnosable via per\-type accuracy\. MC distractors are drawn from related papers in the same discipline \(not generic foils\) so that a deck which talks about the right*topic*but omits the right*paper\-specific*content cannot pass by topic match alone — a failure mode that affects every refinement system that paraphrases freely\.

Protocol\.For each paper,55questions per type are generated \(2020total\), each with44answer choices \(one correct, three same\-discipline distractors\)\. Every question must be answerable from a well\-made deck alone, without consulting the full paper\. The judge LLM \(App\.[I](https://arxiv.org/html/2609.00194#A9)\) reads the rendered deck PNGs and selects an answer for each question;Fid\.is the fraction of correct answers, averaged across all four types\.

Answer uniqueness and No\-Context screening\.A calibration pass removes questions whose correct answer can be guessed from world knowledge alone\. Under the No\-Context condition, the judge receives only the question and answer choices; questions answered correctly in more than half of three trials are regenerated\.

Statistics\.The final bank contains 2,000 questions across 100 papers \(20 per paper\); per\-type distribution: 500 Contribution / 500 Method / 500 Experiment / 500 Limitation; mean question length 18\.3 tokens, mean answer\-choice length 8\.7 tokens\.

License\.Evaluation code is released under the MIT license\. Dataset artifacts are distributed according to per\-item source licenses documented in a license manifest accompanying the release; for sources that do not permit redistribution, we release identifiers and reconstruction scripts only\.

Comparison against existing slide benchmarks\.Tab\.[10](https://arxiv.org/html/2609.00194#A11.T10)positionsDeckQuizagainst existing deck\-level benchmarks along the axes that matter for refinement\-loop attribution: how spatial health is measured, how content fidelity is scored, how design quality is anchored, and whether the suite resolves per\-loop\-component \(step\-level vs\. turn\-level\) signal beyond a single end\-to\-end deck score\. The differentiating column is*per\-component attribution*: most existing benchmarks score the final artifact as a whole, making it difficult to isolate the contribution of individual loop components\.DeckQuizis designed to pair a deterministic spatial task definition for hard layout failures with a calibrated rendered\-PNG VLM judge, paper\-level QA fidelity, and a per\-layer principle decomposition, enabling per\-loop\-component ablation analysis\.

Table 10:Comparison ofDeckQuizagainst existing deck\-level evaluation suites\.*Spatial*: rendered\-PNG VLM judge \(LLM\) vs\. deterministic DOM/render\-based geometric check \(DOM\) vs\. absent \(no\)\.*Content*: per\-paper QA against extracted facts \(QA\) vs\. rubric / checklist LLM scoring \(rubric\) vs\. embedding similarity to a reference deck \(sim\) vs\. absent \(no\); “\+” means both modes are supported\.*Aesthetic*: VLM rubric over multiple design axes \(LLM\) vs\. low\-level visual statistics or verifiable layout reward \(signal\) vs\. absent \(no\)\.*Human anchor*: forced\-choice human study calibrating the judge\.*Per\-component attr\.*: whether the suite resolves per\-loop\-component signal beyond end\-to\-end ranking\.*Trajectory*: whether the suite reports per\-turn or per\-iteration metrics\. Numbers from each benchmark’s published paper / repository; AeSlides counts individual page\-level samples \(†\\dagger\), not decks; SlidesGen\-Bench counts generated outputs \(‡\\ddagger\); “—” marks dimensions not specifically curated\.
## Appendix LDeckDesignPrinciples and Scoring

This appendix specifies the five principles whose aggregate is theDes\.column of Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)and Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)\. The principles are organised into two layers that capture the*information architecture*of a slide deck—an orthogonal dimension to geometry \(SpatialCheck\), aesthetics \(Aesthetics\), and content fidelity \(ContentQuiz\)\. The theoretical grounding draws on Mayer’s multimedia principles\[[35](https://arxiv.org/html/2609.00194#bib.bib30)\], Kosslyn’s cognitive\-load guidelines\[[24](https://arxiv.org/html/2609.00194#bib.bib31)\], Cleveland & McGill’s graphical perception hierarchy\[[6](https://arxiv.org/html/2609.00194#bib.bib32)\], Tufte’s data–ink ratio\[[44](https://arxiv.org/html/2609.00194#bib.bib28)\], and Reynolds’ presentation\-Zen heuristics\[[39](https://arxiv.org/html/2609.00194#bib.bib33)\]\. Layer11\(deck\-level\) requires seeing the full slide sequence and is judged once per deck; Layer22\(slide\-level\) is judged per slide and averaged\.

Layer 1: Deck\-level \(P1, P5\)\.*P1 Narrative Progression\.*The deck builds a coherent argumentative arc \(context→\\togap→\\toapproach→\\toevidence→\\toimplication\), and each slide advances the narrative by introducing substantive information the audience needs at that point; content\-free slides \(section dividers, outlines, thank\-you placeholders\) break progression and are penalised\.*P5 Cross\-Slide Consistency\.*Recurring concepts \(variable names, model names, colour encodings, terminology\) are referred to with consistent symbols and visual encodings throughout the deck so it reads as one integrated document rather than a patchwork of independently authored slides\.

Layer 2: Slide\-level \(P2–P4\)\.*P2 Slide Segmentation*\(Mayer’s segmenting principle\): each slide conveys exactly one coherent message unit, neither overloaded \(multiple topics on one slide\) nor underloaded \(title\-only divider\)\.*P3 Signal & Emphasis*\(Mayer’s signaling principle\): the slide uses visual hierarchy \(size, weight, colour, position\) to direct attention to the key takeaway, evaluated along three aspects—\(a\) action\-oriented assertion titles vs\. generic topic labels, \(b\) differential body hierarchy vs\. uniform template styling, and \(c\) explicit takeaway elements \(summary bars, callout boxes\)\.*P4 Structural Correspondence*\(Cleveland & McGill, Tufte\): the visual form matches the content’s logical structure—tables for comparisons, charts for quantities, timelines for processes, prose for arguments\.

Scoring\.Each principlePiP\_\{i\}\(i=1,…,5i\\\!=\\\!1\{,\}\\dots\{,\}5\) is scored on a00–33Likert scale with explicit anchors \(00absent /11partial /22adequate /33exemplary; the per\-anchor descriptors for each principle are released with the supplementary code\)\. TheDes\.score in Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)is the principle\-count\-weighted layer mean rescaled to\[0,100\]\[0,100\]:

Des\.=\(15​∑i=15Pi\)/3×100=25​P¯deck\+35​P¯slide,\\textsc\{Des\.\}\\;=\\;\\Bigl\(\\tfrac\{1\}\{5\}\\textstyle\\sum\_\{i=1\}^\{5\}P\_\{i\}\\Bigr\)\\,/\\,3\\,\\times\\,100\\;=\\;\\tfrac\{2\}\{5\}\\,\\bar\{P\}\_\{\\text\{deck\}\}\\;\+\\;\\tfrac\{3\}\{5\}\\,\\bar\{P\}\_\{\\text\{slide\}\},whereP¯deck=P1\+P52⋅1003\\bar\{P\}\_\{\\text\{deck\}\}=\\frac\{P\_\{1\}\+P\_\{5\}\}\{2\}\\\!\\cdot\\\!\\frac\{100\}\{3\}aggregates the two deck\-level principles andP¯slide=P2\+P3\+P43⋅1003\\bar\{P\}\_\{\\text\{slide\}\}=\\frac\{P\_\{2\}\+P\_\{3\}\+P\_\{4\}\}\{3\}\\\!\\cdot\\\!\\frac\{100\}\{3\}aggregates the three slide\-level principles\. Layer weights2/52/5and3/53/5therefore reflect principle counts, not editorial preference, and the same weighting is used in the per\-layer regression audit \(App\.[P](https://arxiv.org/html/2609.00194#A16)\)\.

## Appendix MProbe Library Composition and Per\-Task Call Budget

The probe library consists of 40 probe group files containing 237 atomic checks, partitioned into six families:AAnarrative \(44 checks\),BgeomB\_\{\\text\{geom\}\}layout \(37 checks\)222The B family is stored as a single registry entry of 122 atomic checks; we partition it at analysis time intoBgeomB\_\{\\text\{geom\}\}\(deterministic geometric checks:overlap,text\_overflow,pw\_oob; 37 checks\) andBvisualB\_\{\\text\{visual\}\}\(the remaining 85 rubric\-judged style checks\), reflecting their distinct gating policy \(BgeomB\_\{\\text\{geom\}\}hard,BvisualB\_\{\\text\{visual\}\}soft\)\.,BvisualB\_\{\\text\{visual\}\}visual style \(85 checks\),CCcompleteness \(22 checks\),DDcorrectness \(33 checks\), andEEsource fidelity \(16 checks\)\. TheDeckQuizquestion bank contains100100papers×\\times44QA types×\\times55questions=2000=2000items\. All systems run under a matched per\-task LLM\-call\-count cap of150150calls \(agent \+ turn\-level \+ step\-level combined\); systems that exhaust the cap early submit their current state\. The prompt templates, probe definitions, thresholds, and question bank used in the held\-out evaluation are released as part of the codebase so that the runs in Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)and Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)can be re\-executed against the same artifacts\.

Adaptive schedulerπ𝒫\\pi\_\{\\mathcal\{P\}\}\.The 237 atomic checks are organised into 40 probe groups across the six families\. Each probe group has a parent ID \(e\.g\.,B03\) and contains 3–8 atomic checks with fine\-grained IDs \(e\.g\.,B03\.1“content elements overlap making text unreadable”\)\. At each turntt, the scheduler receives rendered PNGs, the current open\-issue listLtL\_\{t\}, the set of modified slides, and the full catalog, and selects33–55atomic check IDs based on visual inspection and issue history\. For every open issue inLtL\_\{t\}on a modified slide, the scheduler must select at least one check from the corresponding probe group, ensuring re\-verification\. Selected check IDs sharing a parent are executed as a single LLM call\. Any open issue whose probe group was not selected is automatically carried forward aspersistedinLt\+1L\_\{t\+1\}\.

## Appendix NInitial\-Deck Attribution and Task\-Paired Inference

Native comparisons combine each system’s initial generator with its refinement procedure\. To isolate loop quality, we cross the ReDeck and DeepPresenter refinement procedures with both initial\-deck sources\. For formal GPT\-5\.4 comparisons, we first average the three seeds within each task, then compute task\-bootstrap 95% CIs and paired permutation tests over the 100 tasks\. Holm correction covers the 12ReDeck\-versus\-baseline comparisons; Gemini\-3\.1 and Claude\-4\.6 are descriptive replications\.

Table 11:Initial\-deck attribution and task\-paired inference on GPT\-5\.4\. Panel \(a\) applies both refinement procedures to both initial\-deck sources \(100 tasks×\\times3 seeds\)\. Panel \(b\) reportsReDeckminus baseline with task\-bootstrap 95% CIs after averaging seeds within task\.\(a\) Shared\-start refinement comparison

\(b\) Task\-paired main\-table differences

In Panel \(a\),ReDeckimproves SCR over the DeepPresenter loop by19\.519\.5points fromReDeckT0T\_\{0\}and20\.220\.2points from DeepPresenterT0T\_\{0\}, with no clear draft\-by\-loop interaction \(Δ​Δ=−0\.7\\Delta\\Delta=\-0\.7points\)\. Seven of the eight metric contrasts have 95% CIs excluding zero; only the\+0\.03\+0\.03Aes\.contrast from DeepPresenterT0T\_\{0\}is inconclusive\. The loop advantage therefore persists across both initial\-deck sources\.

In Panel \(b\), all 12 CIs exclude zero and all paired permutation tests remain significant after Holm correction\. The smallest contrast isAes\.versus DeepPresenter \(\+0\.17\+0\.17\[0\.02, 0\.32\], raw and Holm\-adjustedp=0\.024p=0\.024\)\. Across main\-table cells, seed\-level SEs range from0\.50\.5to1\.21\.2on the 0–100 metrics and0\.030\.03to0\.070\.07onAes\.; ablation\-cell SEs range from0\.80\.8to1\.71\.7and0\.050\.05to0\.080\.08, respectively\. Human\-study intervals are computed over task\-level majority votes rather than seeds\.

## Appendix OAbsolute Inference Cost

Tab\.[12](https://arxiv.org/html/2609.00194#A15.T12)reports native per\-task cost and latency on the GPT\-5\.4 100\-task split\. Tab\.[13](https://arxiv.org/html/2609.00194#A15.T13)then compares DeepPresenter andReDeckat approximately equal mean post\-T0T\_\{0\}refinement cost\.

Table 12:Cost and latency per generated deck on the GPT\-5\.4 100\-task in\-domain split \(mean over 3 seeds\)\. Latency is end\-to\-end wall\-clock under each system’s native per\-slide parallelism; token counts sum agent, turn\-level, and step\-level calls\.Table 13:Refinement\-cost control on the sharedReDeckT0T\_\{0\}\(100 tasks×\\times3 seeds\)\. The cost\-scaled DeepPresenter cap matches mean post\-T0T\_\{0\}cost, not exact per\-task spend\.At the21752175\-call endpoint,ReDeckexceeds the cost\-scaled DeepPresenter loop by\+11\.6/\+13\.5/\+0\.44/\+8\.2\+11\.6/\+13\.5/\+0\.44/\+8\.2onFid\./SCR/Aes\./Des\.; all four task\-bootstrap CIs exclude zero\. The tested scaling policy improves DeepPresenter SCR from72\.072\.0to78\.078\.0but does not close the gap\.

Routed execution\.ReDeck\-Lite retains high\-level planning and final submission on GPT\-5\.4 while routing recurrent edit execution and both verification channels to GPT\-5\.4\-nano\. It reduces mean cost from $1\.74 to $0\.28 and latency from35863586s to14501450s, while changingFid\./SCR/Aes\./Des\.by−0\.8/−1\.7/−0\.03/−0\.8\-0\.8/\-1\.7/\-0\.03/\-0\.8\. The full system therefore remains substantially more expensive than DeepPresenter, and routing mitigates rather than removes this trade\-off\.

## Appendix PPer\-Layer Regression Honesty

DeckDesign aggregates five principles into two layers \(deck\-level: P1*narrative progression*, P5*cross\-slide consistency*; slide\-level: P2*slide segmentation*, P3*signal & emphasis*, P4*structural correspondence*; see App\.[L](https://arxiv.org/html/2609.00194#A12)\)\. Reporting only the aggregate*Des\.*score in Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)can mask cases whereReDecklifts one layer while regressing on another\. This appendix therefore reportsReDeck’s per\-layer changeΔL=ScoreLTfinal−ScoreLT0\\Delta\_\{L\}=\\mathrm\{Score\}\_\{L\}^\{T\_\{\\text\{final\}\}\}\-\\mathrm\{Score\}\_\{L\}^\{T\_\{0\}\}on the GPT\-5\.4 main split\. We do not assign artificial zero\-delta rows to published baselines because they do not shareReDeck’s explicitT0→TfinalT\_\{0\}\\\!\\to\\\!T\_\{\\text\{final\}\}repair trajectory\.

Falsifying outcome\.IfΔL<0\\Delta\_\{L\}<0forReDeckon any layer, the aggregate gain in Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)is contaminated by a regression that the main number conceals; we still report the main number but flag the affected principle in the discussion\. A negativeΔ\\Deltaconfined to a single principle within a positive layer is reported as a sub\-row\.

Table 14:ReDeckper\-principle and per\-layer change in DeckDesign score from the initial deck \(T0T\_\{0\}\) to the submitted deck \(TfinalT\_\{\\text\{final\}\}\) on the GPT\-5\.4 main split\. Per\-principle and per\-layer entries use the same00–100100*Des\.*scale as Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1); layer values are within\-layer principle means and the aggregate\+8\.4\+8\.4is the principle\-count weighted mean of the two layers \(25×6\.7\+35×9\.5=8\.38\\tfrac\{2\}\{5\}\{\\times\}6\.7\+\\tfrac\{3\}\{5\}\{\\times\}9\.5=8\.38, with weights2/52/5and3/53/5from App\.[L](https://arxiv.org/html/2609.00194#A12)\)\. A negativeΔ\\Deltaon any principle within a positive layer is reported as a sub\-row\.All five per\-principle deltas are strictly positive \(\+5\.8≤ΔPi≤\+11\.2\+5\.8\\leq\\Delta\_\{P\_\{i\}\}\\leq\+11\.2\), so the aggregate\+8\.4\+8\.4in Tab\.[1](https://arxiv.org/html/2609.00194#S4.T1)is not concealing a regression on any single principle \(let alone an entire layer\); the slide\-level layer gains slightly more than the deck\-level layer \(\+9\.5\+9\.5vs\.\+6\.7\+6\.7\) because step\-level render feedback dissolves slide\-local layout violations that disproportionately depress P2 \(segmentation\) and P3 \(signal\), while P4 \(structural correspondence\) and the deck\-level pair remain bottlenecked by content choices the loop does not aggressively rewrite\.

## Appendix QTrajectory Event\-Flow

The issue lifecycle across the refinement trajectory is characterised by four status transitions per issue:open→\\toresolved\(issue addressed by the agent’s edits\),resolved→\\toregressed\(a previously resolved issue reappears due to a later edit\),open→\\topersisted\(issue carried forward because the scheduler did not probe its family this turn\), andopen→\\todropped\(issue no longer detected, not explicitly addressed\)\.

UnderReDeck, the dominant transition isopen→\\toresolved\(68% of issue\-turn pairs\), followed byopen→\\topersisted\(22%, indicating the adaptive scheduler correctly deprioritised stable issues\), withresolved→\\toregressedoccurring in only 8\.2% of cases\. The rollback mechanism accounts for the low regression rate: 73% of potential regressions are caught by per\-step observation and rolled back before they enter the issue list\.

Under the all\-probes\-every\-turn variant, theresolved→\\toregressedrate rises to 23\.5% because probing every family every turn introduces new soft findings whose repair edits break previously clean layout, as described in the scheduling block of Sec\.[4\.3](https://arxiv.org/html/2609.00194#S4.SS3)\.

Per\-family trajectory diagnostics\.The aggregate curves of Fig\.[3](https://arxiv.org/html/2609.00194#S4.F3)do not show*which*family drives the drift; the same runs disaggregated by source family in Fig\.[4](https://arxiv.org/html/2609.00194#S4.F4)\(no new compute\) make this visible\. The six families match the implementation:AAnarrative,BgeomB\_\{\\text\{geom\}\}layout \(pw\_oob/overlap/overflow\),BvisualB\_\{\\text\{visual\}\}visual style,CCcompleteness,DDcorrectness, andEEsource fidelity; hard families \(BgeomB\_\{\\text\{geom\}\},CC,DD\) fail the submission gate, the others do not\. Three observations follow from the stacked composition: \(i\) underReDeck,BgeomB\_\{\\text\{geom\}\}collapses fastest because step\-level render feedback catches it before the turn boundary, and the soft families \(AA,CC,EE\) decay steadily once issues enter the persistent list and structural choices stabilise; \(ii\) underAll\-probes\-every\-turn,BvisualB\_\{\\text\{visual\}\}*grows*monotonically — with no scheduling, the per\-turn critic raises new style findings on every slide every turn and never declares one done; \(iii\) the rewrite edits those findings trigger then break layout, liftingBgeomB\_\{\\text\{geom\}\}in step withBvisualB\_\{\\text\{visual\}\}\(annotated arrow in Fig\.[4](https://arxiv.org/html/2609.00194#S4.F4), right\); this propagation of style\-motivated edits into layout regressions, not any single hard family failing, drives the open\-issue drift in Fig\.[3](https://arxiv.org/html/2609.00194#S4.F3)\(a\)\. The factual familyDDstays small in absolute count but has the highest*reopen*rate per issue underAll\-probes\-every\-turn, because edits to neighbouring text break previously verified claims; this is where step\-level render feedback matters most beyond layout\. Per\-family event\-flow plots, reopen matrices, and the same decomposition forN∈\{4,8\}N\{\\in\}\\\{4,8\\\}are available in the released codebase\.

## Appendix RDiagnostic Mini\-Ablations

Tab\.[15](https://arxiv.org/html/2609.00194#A18.T15)reports two single\-row diagnostics that the four\-block ablation in Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)folds together: \(i\) splitting the per\-turn NL\-critique row by judge source \(Self\-Refine vs\. Reflexion\) and \(ii\) replacing the rendered\-image input to the turn\-level critic with the deck source code only\. GPT\-5\.4×\\times100 tasks×\\times3 seeds, on the same scaffolding as Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)\.

Table 15:Single\-row diagnostics for the per\-turn NL\-critique row of Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)\(rows 1–2\) and the turn\-level critic input modality \(row 3\)\. Numbers are means over three seeds on the same100100\-task split as Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)\.Rows 1–2 disaggregate the per\-turn NL\-critique row of Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)\(76\.8/68\.2/3\.18/60\.576\.8/68\.2/3\.18/60\.5\) into its two implementations: Self\-Refine\[[33](https://arxiv.org/html/2609.00194#bib.bib6)\]uses the agent as its own critic, Reflexion\[[40](https://arxiv.org/html/2609.00194#bib.bib7)\]uses an external judge\. The gap between them is below11point on every module and within the observed seed\-level variability, so the NL\-critique modality, not its source, is the binding constraint\. Row 3 keeps the fullReDeckloop but feeds the turn\-level critic the deck source code only \(no rendered image\); relative to the defaultReDeckconfiguration in Tab\.[2](https://arxiv.org/html/2609.00194#S4.T2)\(88\.6/91\.5/3\.64/71\.288\.6/91\.5/3\.64/71\.2\) this drops SCR by17\.317\.3and Des\. by6\.16\.1while leaving Fid\. approximately flat, validating the dual\-space hypothesis \(§[3\.3](https://arxiv.org/html/2609.00194#S3.SS3)\) that text\-only critique cannot localise spatial violations even when the action channel keeps step\-level rendering intact\.