EMNLP 2026 Findings

ConvDeck: Conversational Paper-to-Slide Generation via Stage-Specific User Feedback

Tarik Can Ozden* Sachidanand VS* Furkan Horoz* Ozgur Kara Dilek Hakkani-Tür† Junho Kim† James M. Rehg†
University of Illinois Urbana-Champaign
* Equal contribution  ·  † Corresponding author
arXiv GitHub Explore

ConvDeck turns a research paper into a presentation you can shape through conversation: refining the narrative outline, then polishing the rendered deck. Feedback is routed to the stage where each decision is actually made.

Overview of the ConvDeck pipeline: Preprocessing, Outline Generation, Conversational Outline Refinement, Slide Generation, and Conversational Slide Refinement.
The ConvDeck pipeline. Given a paper PDF and optional inputs (target audience, presentation duration), ConvDeck produces a final deck through five stages, with two stage-specific conversational refinement loops: one over the outline before rendering, and one over the rendered slides.
See it in action

ConvDeck, presented by ConvDeck

This presentation was generated by ConvDeck from its own paper, then refined through the same conversational stages it describes.

A 16-slide deck generated by ConvDeck from its own paper. Use the arrows (or ← / → keys) to browse, and Expand for fullscreen.

The idea

Rethinking where the conversation happens

Prior systems either generate a deck in a single pass, or bolt a single conversational loop onto the finished slides, forcing every request through a post-hoc editor. ConvDeck instead distributes interaction across the pipeline, aligning each refinement step with the component best suited to act on it.

Comparison to prior paper-to-slide paradigms: (a) Non-conversational, (b) Post-editing, (c) ConvDeck.
Comparison to prior paradigms. (a) Non-conversational methods produce a deck in a single pass, with no interaction. (b) Post-editing methods add one refinement loop after the deck is generated, so all feedback is absorbed by a single post-hoc editing stage. (c) ConvDeck distributes interaction through stage-specific loops.
(a) Non-conversational

One pass, no dialogue

The system reads the paper and emits the full deck in a single generation. If the output misses your intent, there is no way to steer it.

(b) Post-editing

Feedback comes too late

A single conversation runs on the finished slides. Narrative order, content coverage must be patched through surface-level edits, which is inefficient and unreliable.

(c) ConvDeck

Feedback at the right stage

Two loops: revise the narrative and structure on the outline before rendering, then refine figures, layout, and wording on the rendered deck.

How it works

Five stages, two conversations

ConvDeck's outline and slide generation follow a narrative-driven, multi-agent design (building on ArcDeck). What's new is that the two conversational stages — Stage 3 and Stage 5 — put the user in the loop exactly where each kind of decision is made.

1

Preprocessing

Docling parses the paper into a markdown representation and an asset dictionary of figures and tables with their captions, while a citation-key dictionary preserves in-text citations for later stages.

2

Outline Generation

A narrative-driven planner builds a draft outline: an RST discourse tree captures rhetorical relations, a Commitment Builder turns the target audience and duration into a global intent, and a Slide Planner ⇄ Narrative Critic cycle drafts and revises the outline.

3

Conversational Outline Refinement Conversational

Before any slide is rendered, you revise narrative flow, section emphasis, slide ordering, and content coverage. An Outline Refiner adds, edits, splits, merges, removes, and reorders slides — so structural changes propagate cheaply across the whole deck.

4

Slide Generation

The refined outline becomes a draft deck: a Slide Deck Constructor selects figures and layouts and writes concise bullets, and an Aesthetic Refiner adds emphasis and fills sparse slides. Slides are compiled to editable PPTX via JavaScript and PPTXGenJS.

5

Conversational Slide Refinement Conversational

Now looking at the rendered slides, a Slide Refiner corrects what only the visual output reveals — text overflow, undersized figures, crowded layouts — through localized edits: insert, delete, modify, split, merge, reorder, reposition, resize, and typography.

The mechanism

Think, Speak, Act

Inspired by ReSpAct, ConvDeck's refiners don't blindly edit. They reason about your feedback and then either speak — asking a clarifying question or explaining a revision when a request is ambiguous — or act, applying localized edits through dedicated tools. Every edit is a local update, so the deck is never regenerated from scratch.

Conversational refinement: the refiner reasons about user feedback, then speaks with the user or acts through tools to return an updated outline or deck.
Conversational refinement overview. Given a draft outline or deck, the user provides feedback. The refiner reasons about the input, either speaks with the user or acts through tools (editing operations, web and arXiv search), and returns an updated outline or deck for the next round.
Think

Reason first

The agent reasons about the request and what information is missing before touching the outline or deck.

Speak

Clarify in dialogue

When feedback is underspecified or the goal is ambiguous, it asks a clarifying question and discusses revision options rather than guessing.

Act

Edit & retrieve

It applies edits through dedicated editing functions, and can search the web or arXiv to bring in related work or newer results on request.

Evaluation

Does routing feedback by stage actually help?

On the 100-paper ArcBench benchmark, across GPT-5, Gemini 3 Pro, and Qwen3-VL backbones and two independent VLM judges, stage-specific conversational feedback raises user-goal satisfaction while preserving narrative coherence, content quality, and visual presentation.

79%
Initially-unmet goals satisfied overall
Highest of any conversational system
+54
Points of outline-goal satisfaction from Stage 3
37.7% → 92.0% before vs. after refinement
+18
Points of slide-goal satisfaction from Stage 5
68.4% → 86.4% before vs. after refinement
96%
Goals met in natural human-interaction studies
93% on participant-defined, paper-specific goals
Conversational stageBeforeAfterΔ
Outline refinement (Stage 3)37.7%92.0%+54.3
Slide refinement (Stage 5)68.4%86.4%+18.0

Per-stage goal improvement. Each stage improves mainly the categories it is designed to address — outline-level changes are cheap and global before rendering, while figure, layout, and wording fixes depend on the rendered output only Stage 5 can observe.

Preferred over every baseline by both VLM judges on the stronger GPT-5 and Gemini 3 Pro backbones, and the automated method that comes closest to author-prepared slides.
An ablation confirms the two loops are complementary: removing both leaves a one-shot deck satisfying just 18% of goals — a 64-point drop.
Affordable interactivity: a 98% render success rate, ~$0.82 per paper, and only ~10% compute overhead over its non-conversational backbone, since each round applies localized edits.
Cite

Citation

ConvDeck is accepted to the Findings of the Association for Computational Linguistics: EMNLP 2026.

@article{ozden2026convdeck,
  title     = {ConvDeck: Conversational Paper-to-Slide Generation
               via Stage-Specific User Feedback},
  author    = {Ozden, Tarik Can and VS, Sachidanand and Horoz, Furkan and
               Kara, Ozgur and Hakkani-T{\"u}r, Dilek and Kim, Junho and
               Rehg, James M.},
  journal   = {arXiv preprint arXiv:2609.00226},
  year      = {2026}
}