pith. sign in

arxiv: 2607.00407 · v1 · pith:A52ABWLOnew · submitted 2026-07-01 · 💻 cs.AI

Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising

Tianci Liu , Zihan Dong , Linjun Zhang , Haoyu Wang , jing Gao , Emre Kiciman , Ranveer Chandra , Wei-Ting Chen This is my paper

Pith reviewed 2026-07-02 13:08 UTC · model grok-4.3

classification 💻 cs.AI
keywords page-level slide personalizationinverse planningstructural denoisingmulti-agent reinforcement learninglatent design intentsSPIREagentic slide generation
0
0 comments X

The pith

Page-level slide personalization reduces to recovering latent design intents by treating it as inverse planning solved through structural denoising with multi-agent reinforcement learning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper formulates page-level slide personalization as an inverse planning problem to capture latent design intents without any knowledge of the specific tools like PowerPoint or Beamer. It introduces the SPIRE framework, which corrupts clean slides to create a denoising task that two agents solve collaboratively via reinforcement learning. A proof establishes that this structural denoising is a consistent surrogate for the original problem and that the multi-agent setup strictly lowers policy gradient variance. Sympathetic readers would care because existing agent methods rely on templates or verbose instructions and fail at fine-grained, page-level personalization.

Core claim

We formulate PSP as an inverse planning problem and propose SPIRE to solve it approximately by intentionally corrupting visual structures of clean slides, creating a verifiable denoising task solved by two collaborative agents via RL; structural denoising is a consistent surrogate for PSP, and the multi-agent formulation strictly reduces policy gradient variance.

What carries the argument

Structural denoising created by corrupting slide visuals, solved collaboratively by two agents in RL to recover latent design intents.

If this is right

  • Latent design intents for themes and layouts can be learned end-to-end without specifying or controlling the execution tools.
  • The multi-agent RL formulation provides strictly lower policy gradient variance than single-agent alternatives.
  • SPIRE approximates the intractable PSP problem while remaining verifiable through the denoising objective.
  • Experiments confirm superiority over prior template-based or instruction-based agent methods for page-level personalization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same corruption-and-denoise pattern could extend to other agentic creative tasks where execution environments are unknown in advance.
  • Verification through structural consistency might reduce reliance on explicit user feedback loops in personalization systems.
  • If the learned intents prove transferable, the method could support cross-tool slide generation without retraining per software.

Load-bearing premise

Intentionally corrupting the visual structures of clean slides creates a verifiable denoising task whose solution corresponds to recovering latent design intents without knowledge of the executing tools.

What would settle it

A test set of slides with known ground-truth intents where the agents' denoised outputs fail to match the original designs would falsify the consistency of the surrogate.

Figures

Figures reproduced from arXiv: 2607.00407 by Emre Kiciman, Haoyu Wang, Jing Gao, Linjun Zhang, Ranveer Chandra, Tianci Liu, Wei-Ting Chen, Zihan Dong.

Figure 1
Figure 1. Figure 1: Illustration of Spire. Trained with reinforcement learning, Spire learns to infer design intents instead of using prespecified layout templates or lengthy user instructions as existing methods do [10, 53], offering both theoretical and empirical advantages. Recent advances in multi-modal large language models (MLLMs) have en￾abled agentic pipelines that automate slide generation for practical use [9, 10, 2… view at source ↗
Figure 2
Figure 2. Figure 2: The overview of Personalized Slide Personalization (PSP) and the proposed Spire. The Planner and Critic are trained to recover self-supervised Structural Per￾turbation in a decomposed way with reinforcement learning (red). After training, they collaborate with a black-box render executor to conduct PSP. where the integrand consists of two parts. The first term, πθ(z | x, Dref), denotes how the planner πθ p… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of slide generation results. Strong Generalizability. On OOD pages, Spire demonstrates strong gen￾eralizability and achieves the highest judge average, substantially outperforming baselines such as AutoPresent, PSP (o4-mini), PPTAgent, SD 3.5, and PSP (Base). Our method obtains the best scores across all judging dimensions and the second-best visual similarity. These results indicate… view at source ↗
Figure 4
Figure 4. Figure 4: Ablation on Spire revision. Multi-round revision in Spire helps improve both visual similarity and judge scores, indicating better generation qualities. and [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 1
Figure 1. Figure 1: Additional qualitative comparison of slide generation results. Graphic-related categories. For non-image graphic shapes, we define three per￾turbation categories: – Graphic Size (graphic_size): perturbs one size attribute of a graphic element, namely width or height. – Graphic Position (graphic_position): perturbs one spatial coordinate of a graphic element, namely its horizontal coordinate x or vertical c… view at source ↗
Figure 2
Figure 2. Figure 2: Qualitative examples of the proposed perturbation pipeline. Each row shows one original-perturbed slide pair [PITH_FULL_IMAGE:figures/full_fig_p039_2.png] view at source ↗
read the original abstract

Slide design requires personalizing both deck themes and page layouts. Yet, current AI agent-based methods struggle with fine-grained, page-level design. Solely relying on prespecified templates or user verbose instructions, they fail to capture latent design intents, leaving Page-level Slide Personalization (PSP) unresolved. To close this gap, this work formulates PSP as an inverse planning problem. We propose to learn a design intent without assuming any knowledge of the specific executing tools (e.g., PowerPoint, Beamer) being used. However, relinquishing control over these tools makes the problem intractable to optimize end-to-end. To overcome this, we propose SPIRE, a principled framework to solve PSP approximately. By intentionally corrupting the visual structures of clean slides, SPIRE creates a verifiable task to denoise the corruption, whereby two agents learn to collaboratively refine executable designs via reinforcement learning (RL). We present a proof that structural denoising is a consistent surrogate for PSP, and that the multi-agent formulation strictly reduces policy gradient variance in RL. Extensive experiments demonstrate the superiority of SPIRE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper formulates Page-level Slide Personalization (PSP) as an inverse planning problem and introduces the SPIRE framework. SPIRE intentionally corrupts visual structures of clean slides to create a denoising task solved collaboratively by two agents via reinforcement learning, without assuming knowledge of executing tools such as PowerPoint. It asserts a proof that structural denoising is a consistent surrogate for PSP and that the multi-agent RL formulation strictly reduces policy gradient variance, with experiments claimed to demonstrate superiority over prior methods.

Significance. If the consistency proof and variance-reduction result hold with verifiable derivations, and if experiments confirm recovery of latent intents, the work could offer a tool-agnostic approach to agentic design personalization that addresses limitations of template-based or instruction-heavy methods. The multi-agent RL variance claim, if rigorously established, would also be of independent interest in RL.

major comments (2)
  1. [Abstract] Abstract: The central claim rests on the assertion that 'We present a proof that structural denoising is a consistent surrogate for PSP', yet the manuscript supplies no derivation steps, equations, or intermediate results showing that the optimal policy under the denoising objective recovers the original latent design intents (rather than any of the other structures that could satisfy the corruption loss). Without this, the many-to-one mapping concern cannot be ruled out.
  2. [Abstract] Abstract: The claim that 'the multi-agent formulation strictly reduces policy gradient variance in RL' is load-bearing for the proposed solution but is stated without the underlying RL formulation, the single-agent baseline comparison, or the variance-reduction derivation. This prevents assessment of whether the reduction is strict and holds under the inverse-planning objective.
minor comments (1)
  1. [Abstract] The abstract would benefit from a brief statement of the experimental metrics and baselines used to claim superiority, to allow readers to gauge the strength of the empirical support.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful review and for highlighting the need for explicit derivations of our central theoretical claims. We agree that the consistency proof and variance-reduction result require fuller presentation to allow verification. Below we respond to each major comment and commit to revisions that will include the requested steps, equations, and comparisons without altering the core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim rests on the assertion that 'We present a proof that structural denoising is a consistent surrogate for PSP', yet the manuscript supplies no derivation steps, equations, or intermediate results showing that the optimal policy under the denoising objective recovers the original latent design intents (rather than any of the other structures that could satisfy the corruption loss). Without this, the many-to-one mapping concern cannot be ruled out.

    Authors: We accept that the abstract states the existence of the proof without supplying its steps. The full manuscript contains a consistency argument in the theoretical analysis section, but it is presented at a high level. In the revision we will add a dedicated subsection that spells out the full derivation: starting from the inverse-planning objective, showing the equivalence between recovering latent intents and minimizing the expected corruption loss, and proving that the optimal denoising policy is unique (under the structural assumptions of the slide representation) rather than admitting arbitrary many-to-one solutions. This will directly address the many-to-one mapping concern. revision: yes

  2. Referee: [Abstract] Abstract: The claim that 'the multi-agent formulation strictly reduces policy gradient variance in RL' is load-bearing for the proposed solution but is stated without the underlying RL formulation, the single-agent baseline comparison, or the variance-reduction derivation. This prevents assessment of whether the reduction is strict and holds under the inverse-planning objective.

    Authors: We agree that the variance-reduction claim needs the supporting RL formulation and derivation to be verifiable. The manuscript defines the collaborative two-agent policy gradient in Section 4, but the explicit single-agent baseline and the variance expressions are only sketched. In the revision we will expand this section to include: (i) the precise single-agent and multi-agent policy-gradient estimators, (ii) the variance formulas for each, and (iii) the algebraic steps establishing that the multi-agent formulation yields a strictly lower variance under the same inverse-planning objective. A direct comparison table will also be added. revision: yes

Circularity Check

0 steps flagged

No circularity detected from available text

full rationale

The abstract asserts a proof that structural denoising is a consistent surrogate for PSP and that multi-agent RL reduces policy gradient variance, but supplies no equations, derivations, self-citations, or load-bearing steps that can be inspected for reduction to inputs by construction. No self-definitional mappings, fitted inputs renamed as predictions, or uniqueness theorems imported from prior author work are present. The derivation chain cannot be walked beyond the high-level claim, so no circular steps are identifiable; the result is treated as self-contained pending full text.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no information on free parameters, axioms, or invented entities; full text would be required to populate the ledger.

pith-pipeline@v0.9.1-grok · 5746 in / 1025 out tokens · 32668 ms · 2026-07-02T13:08:22.397007+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

94 extracted references · 27 canonical work pages · 10 internal anchors

  1. [1]

    Liu et al

    Anthropic: Claude sonnet 4: Hybrid reasoning model with superior intelligence for high-volume use cases, and 200k context window.https://www.anthropic.com/ claude/sonnet(2025) 16 T. Liu et al

    2025

  2. [2]

    Bai, S., Chen, K., Liu, X., Wang, J., Ge, W., Song, S., Dang, K., Wang, P., Wang, S., Tang, J., Zhong, H., Zhu, Y., Yang, M., Li, Z., Wan, J., Wang, P., Ding, W., Fu, Z., Xu, Y., Ye, J., Zhang, X., Xie, T., Cheng, Z., Zhang, H., Yang, Z., Xu, H., Lin, J.: Qwen2.5-vl technical report (2025)

    2025

  3. [3]

    In: Proceedings of The 5th New Frontiers in Summarization Workshop

    Cao, J., Zhang, X., Li, R., Wei, J., Li, C., Joty, S., Carenini, G.: Multi2: Multi- agent test-time scalable framework for multi-document processing. In: Proceedings of The 5th New Frontiers in Summarization Workshop. pp. 135–156 (2025)

    2025

  4. [4]

    arXiv preprint arXiv:2306.03082 (2023)

    Chen, L., Chen, J., Goldstein, T., Huang, H., Zhou, T.: Instructzero: Efficient instruction optimization for black-box large language models. arXiv preprint arXiv:2306.03082 (2023)

    work page arXiv 2023

  5. [5]

    In: Findings of the Association for Computational Linguistics: ACL 2023

    Chen, Z., Liu, G., Zhang, B.W., Yang, Q., Wu, L.: Altclip: Altering the language encoder in clip for extended language capabilities. In: Findings of the Association for Computational Linguistics: ACL 2023. pp. 8666–8682 (2023)

    2023

  6. [6]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Comanici, G., Bieber, E., Schaekermann, M., Pasupat, I., Sachdeva, N., Dhillon, I., Blistein, M., Ram, O., Zhang, D., Rosen, E., et al.: Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities. arXiv preprint arXiv:2507.06261 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  7. [7]

    A Survey on Code Generation with LLM-based Agents

    Dong, Y., Jiang, X., Qian, J., Wang, T., Zhang, K., Jin, Z., Li, G.: A survey on code generation with llm-based agents. arXiv preprint arXiv:2508.00083 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  8. [8]

    In: Forty-first international conference on machine learning (2024)

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: Forty-first international conference on machine learning (2024)

    2024

  9. [9]

    In: Proceedings of the AAAI Confer- ence on Artificial Intelligence

    Fu, T.J., Wang, W.Y., McDuff, D., Song, Y.: Doc2ppt: Automatic presentation slides generation from scientific documents. In: Proceedings of the AAAI Confer- ence on Artificial Intelligence. vol. 36, pp. 634–642 (2022)

    2022

  10. [10]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Ge, J., Wang, Z.Z., Zhou, X., Peng, Y.H., Subramanian, S., Tan, Q., Sap, M., Suhr, A., Fried, D., Neubig, G., et al.: Autopresent: Designing structured visuals from scratch. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 2902–2911 (2025)

    2025

  11. [11]

    arXiv preprint arXiv:2412.06771 (2024)

    Hahn, M., Zeng, W., Kannen, N., Galt, R., Badola, K., Kim, B., Wang, Z.: Proactive agents for multi-turn text-to-image generation under uncertainty. arXiv preprint arXiv:2412.06771 (2024)

    work page arXiv 2024

  12. [12]

    Advances in Neural Information Processing Systems37, 86309–86345 (2024)

    Hu, W., Shu, Y., Yu, Z., Wu, Z., Lin, X., Dai, Z., Ng, S.K., Low, B.K.H.: Local- ized zeroth-order prompt optimization. Advances in Neural Information Processing Systems37, 86309–86345 (2024)

    2024

  13. [13]

    In: IJCAI

    Hu, Y., Wan, X.: Ppsgen: Learning to generate presentation slides for academic papers. In: IJCAI. pp. 2099–2105 (2013)

    2099

  14. [14]

    arXiv preprint arXiv:2601.15286 (2026)

    Jaiswal, S., Prabhudesai, M., Bhardwaj, N., Qin, Z., Zadeh, A., Li, C., Fragki- adaki, K., Pathak, D.: Iterative refinement improves compositional image genera- tion. arXiv preprint arXiv:2601.15286 (2026)

    work page arXiv 2026

  15. [15]

    DECKBench: Benchmarking Multi-Agent Frameworks for Academic Slide Generation and Editing

    Jang, D., Heisler, M.L., Xing, L., Li, Y., Wang, E., Xiong, Y., Zhang, Y., Fan, Z.: Deckbench: Benchmarking multi-agent frameworks for academic slide generation and editing. arXiv preprint arXiv:2602.13318 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  16. [16]

    Talk to Your Slides: High-Efficiency Slide Editing via Language-Driven Structured Data Manipulation

    Jung, K., Cho, H., Yun, J., Yang, S., Jang, J., Choo, J.: Talk to your slides: Language-driven agents for efficient slide editing. arXiv preprint arXiv:2505.11604 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Lan, G.: First-order and stochastic optimization methods for machine learning, vol. 1. Springer (2020) 17

    2020

  18. [18]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Li, C., Xue, M., Zhang, Z., Yang, J., Zhang, B., Yu, B., Hui, B., Lin, J., Wang, X., Liu, D.: Start: Self-taught reasoner with tools. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. pp. 13523– 13564 (2025)

    2025

  19. [19]

    arXiv preprint arXiv:2502.11560 (2025)

    Li, W., Wang, X., Li, W., Jin, B.: A survey of automatic prompt engineering: An optimization perspective. arXiv preprint arXiv:2502.11560 (2025)

    work page arXiv 2025

  20. [20]

    arXiv preprint arXiv:2512.04529 (2025)

    Liang, X., Zhang, X., Xu, Y., Sun, S., You, C.: Slidegen: Collaborative multimodal agents for scientific slide generation. arXiv preprint arXiv:2512.04529 (2025)

    work page arXiv 2025

  21. [21]

    arXiv preprint arXiv:2510.05571 (2025)

    Liu, C., Yang, Y., Zhou, K., Zhang, Z., Fan, Y., Xie, Y., Qi, P., Wang, X.E.: Presenting a paper is an art: Self-improvement aesthetic agents for academic pre- sentations. arXiv preprint arXiv:2510.05571 (2025)

    work page arXiv 2025

  22. [22]

    In: ACM SIGGRAPH 2024 Conference Papers

    Ma, J., Liang, J., Chen, C., Lu, H.: Subject-diffusion: Open domain personalized text-to-image generation without test-time fine-tuning. In: ACM SIGGRAPH 2024 Conference Papers. pp. 1–12 (2024)

    2024

  23. [23]

    arXiv preprint arXiv:2508.06916 (2025)

    Ma, S., Guo, Y., Su, J., Huang, Q., Zhou, Z., Wang, Y.: Talk2image: A multi-agent system for multi-turn image generation and editing. arXiv preprint arXiv:2508.06916 (2025)

    work page arXiv 2025

  24. [24]

    arXiv preprint arXiv:2405.13095 (2024)

    Maheshwari, H., Bandyopadhyay, S., Garimella, A., Natarajan, A.: Presentations are not always linear! gnn meets llm for document-to-presentation transformation with attribution. arXiv preprint arXiv:2405.13095 (2024)

    work page arXiv 2024

  25. [25]

    Advances in Neural Information Processing Systems36, 53038–53075 (2023)

    Malladi, S., Gao, T., Nichani, E., Damian, A., Lee, J.D., Chen, D., Arora, S.: Fine- tuning language models with just forward passes. Advances in Neural Information Processing Systems36, 53038–53075 (2023)

    2023

  26. [26]

    In: Proceedings of the 18th Con- ference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers)

    Mondal, I., Shwetha, S., Natarajan, A., Garimella, A., Bandyopadhyay, S., Boyd- Graber, J.: Presentations by the humans and for the humans: Harnessing llms for generating persona-aware slides from documents. In: Proceedings of the 18th Con- ference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 2664–2...

    2024

  27. [27]

    OpenAI: Introducing chatgpt (2022),https://openai.com/blog/chatgpt

    2022

  28. [28]

    OpenAI: Gpt-image-1.https://platform.openai.com/docs/models/gpt-image- 1(2025)

    2025

  29. [29]

    OpenAI: Introducing gpt-5.https://openai.com/index/introducing- gpt- 5/ (2025)

    2025

  30. [30]

    arXiv preprint arXiv:2505.21497 (2025)

    Pang, W., Lin, K.Q., Jian, X., He, X., Torr, P.: Paper2poster: Towards multimodal poster automation from scientific papers. arXiv preprint arXiv:2505.21497 (2025)

    work page arXiv 2025

  31. [31]

    arXiv preprint arXiv:2504.06838 (2025)

    Park, S., Jeong, J., Kim, Y., Lee, J., Lee, N.: Zip: An efficient zeroth-order prompt tuning for black-box vision-language models. arXiv preprint arXiv:2504.06838 (2025)

    work page arXiv 2025

  32. [32]

    In: Findings of the Association for Computational Linguistics: EMNLP 2025

    Qi, Y., Tian, J., Liu, T., Li, R., Wei, T., Liu, H., Tang, X., Cheng, M., He, J.: Learning to instruct: Fine-tuning a task-aware instruction optimizer for black-box llms. In: Findings of the Association for Computational Linguistics: EMNLP 2025. pp. 7707–7733 (2025)

    2025

  33. [33]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Qu, Y., Fang, S., Wang, Y., Wang, X., Chen, Z., Xie, H., Zhang, Y.: Igd: In- structional graphic design with multimodal layer generation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 18218–18228 (2025)

    2025

  34. [34]

    Interna- tional journal of surgical pathology14(4), 285–305 (2006)

    Rojo, M.G., García, G.B., Mateos, C.P., García, J.G., Vicente, M.C.: Critical com- parison of 31 commercially available digital slide systems in pathology. Interna- tional journal of surgical pathology14(4), 285–305 (2006)

    2006

  35. [35]

    An overview of gradient descent optimization algorithms

    Ruder, S.: An overview of gradient descent optimization algorithms. arXiv preprint arXiv:1609.04747 (2016) 18 T. Liu et al

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  36. [36]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Shao, Z., Wang, P., Zhu, Q., Xu, R., Song, J., Bi, X., Zhang, H., Zhang, M., Li, Y., Wu, Y., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  37. [37]

    HybridFlow: A Flexible and Efficient RLHF Framework

    Sheng, G., Zhang, C., Ye, Z., Wu, X., Zhang, W., Zhang, R., Peng, Y., Lin, H., Wu, C.: Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv: 2409.19256 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  38. [38]

    In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies

    Sun, E., Hou, Y., Wang, D., Zhang, Y., Wang, N.X.: D2s: Document-to-slide gener- ation via query-based text summarization. In: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. pp. 1405–1418 (2021)

    2021

  39. [39]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Tang, W., Xiao, J., Jiang, W., Xiao, X., Wang, Y., Tang, X., Li, Q., Ma, Y., Liu, J., Tang,S.,etal.:Slidecoder:Layout-awarerag-enhancedhierarchicalslidegeneration from design. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. pp. 9026–9050 (2025)

    2025

  40. [40]

    Anatomical sciences ed- ucation9(6), 583–602 (2016)

    Trelease, R.B.: From chalkboard, slides, and to e-learning: How computing tech- nologies have transformed anatomical sciences education. Anatomical sciences ed- ucation9(6), 583–602 (2016)

    2016

  41. [41]

    arXiv preprint arXiv:2504.05306 (2025)

    Venkatesh, K., Dunlop, C., Yanardag, P.: Crea: A collaborative multi-agent frame- work for creative image editing and generation. arXiv preprint arXiv:2504.05306 (2025)

    work page arXiv 2025

  42. [42]

    Xie, E., Waterfield, D., Kennedy, M., Zhang, A.: Slidebot: A multi-agent frame- workforgeneratinginformative,reliable,multi-modalpresentations.arXivpreprint arXiv:2511.09804 (2025)

    work page arXiv 2025

  43. [43]

    arXiv preprint arXiv:2602.01511 , year=

    Xu, R., Liu, T., Dong, Z., You, T., Hong, I., Yang, C., Zhang, L., Zhao, T., Wang, H.: Alternating reinforcement learning for rubric-based reward modeling in non- verifiable llm post-training. arXiv preprint arXiv:2602.01511 (2026)

    work page arXiv 2026

  44. [44]

    arXiv preprint arXiv:2505.21660 (2025)

    Xu, X., Xu, X., Chen, S., Chen, H., Zhang, F., Chen, Y.C.: Pregenie: An agen- tic framework for high-quality visual presentation generation. arXiv preprint arXiv:2505.21660 (2025)

    work page arXiv 2025

  45. [45]

    arXiv preprint arXiv:2502.15412 (2025)

    Xu, Y., Ma, X., Qiu, J., Zhao, H.: Textual-to-visual iterative self-verification for slide generation. arXiv preprint arXiv:2502.15412 (2025)

    work page arXiv 2025

  46. [46]

    IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models

    Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: Ip-adapter: Text compati- ble image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  47. [47]

    DAPO: An Open-Source LLM Reinforcement Learning System at Scale

    Yu, Q., Zhang, Z., Zhu, R., Yuan, Y., Zuo, X., Yue, Y., Dai, W., Fan, T., Liu, G., Liu, L., et al.: Dapo: An open-source llm reinforcement learning system at scale. arXiv preprint arXiv:2503.14476 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  48. [48]

    Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking

    Zelikman, E., Harik, G., Shao, Y., Jayasiri, V., Haber, N., Goodman, N.D.: Quiet- star: Language models can teach themselves to think before speaking. arXiv preprint arXiv:2403.09629 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [49]

    Advances in Neural Information Processing Systems35, 15476–15488 (2022)

    Zelikman, E., Wu, Y., Mu, J., Goodman, N.: Star: Bootstrapping reasoning with reasoning. Advances in Neural Information Processing Systems35, 15476–15488 (2022)

    2022

  50. [50]

    arXiv preprint arXiv:2512.20292 (2025)

    Zeng, W., Ouyang, M., Cui, L., Ng, H.T.: Slidetailor: Personalized presentation slide generation for scientific papers. arXiv preprint arXiv:2512.20292 (2025)

    work page arXiv 2025

  51. [51]

    In: Findings of the Association for Computational Linguistics: EMNLP 2024

    Zhan, H., Chen, C., Ding, T., Li, Z., Sun, R.: Unlocking black-box prompt tun- ing efficiency via zeroth-order optimization. In: Findings of the Association for Computational Linguistics: EMNLP 2024. pp. 14825–14838 (2024)

    2024

  52. [52]

    In: 19 Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Zhang, K., Li, J., Li, G., Shi, X., Jin, Z.: Codeagent: Enhancing code generation with tool-integrated agent systems for real-world repo-level coding challenges. In: 19 Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 13643–13658 (2024)

    2024

  53. [53]

    In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing

    Zheng, H., Guan, X., Kong, H., Zhang, W., Zheng, J., Zhou, W., Lin, H., Lu, Y., Han, X., Sun, L.: Pptagent: Generating and evaluating presentations beyond text- to-slides. In: Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. pp. 14413–14429 (2025)

    2025

  54. [54]

    With more and more customers opting out of cookies, the amount of data for wisdom of crowd declines

    Zhu, D., Meng, R., Song, Y., Wei, X., Li, S., Pfister, T., Yoon, J.: Paperbanana: Automating academic illustration for ai scientists. arXiv preprint arXiv:2601.23265 (2026) Supplementary Material of Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising Tianci Liu1,∗, Zihan Dong2, Linjun Z...

    work page arXiv 2026

  55. [55]

    Title Suppressed Due to Excessive Length 33

    GRAPHIC_COLOR: palette + color roles. Title Suppressed Due to Excessive Length 33

  56. [56]

    GRAPHIC_POSITION: alignment + margins + spacing rhythm

  57. [57]

    GRAPHIC_SIZE: relative visual weight + hierarchy

  58. [58]

    IMAGE_POSITION: anchoring + alignment + proximity to related text

  59. [59]

    IMAGE_SIZE: prominence + proportions + image-to-text balance

  60. [60]

    TEXT_COLOR: contrast + emphasis conventions

  61. [61]

    TEXT_POSITION: grid/columns + padding + wrap region

  62. [62]

    </think> ## STEP 3: FEEDBACK CHECKLIST Output a SINGLE <comment> block containing the feedback for ALL 8 aspects

    TEXT_SIZE: typographic hierarchy + consistency. </think> ## STEP 3: FEEDBACK CHECKLIST Output a SINGLE <comment> block containing the feedback for ALL 8 aspects. Ensure every aspect from Step 2 is included in the list inside the block: <comment> <aspect>GRAPHIC_COLOR</aspect>: <current>...</current><target>...</target> ... <aspect>TEXT_SIZE</aspect>: <cur...

  63. [63]

    For position/size revisions, give concrete ratios/ percentages and specify the frame (of panel/of slide)

    Use proportions only (no px/pt). For position/size revisions, give concrete ratios/ percentages and specify the frame (of panel/of slide). Prefer bbox x:.. y:.. w:.. h:.. when precision is needed; otherwise provide at least one clear numeric target (e.g., target margin/height/width) plus an alignment rule

  64. [64]

    Always output ALL 8 aspects, in the exact order above

  65. [65]

    Otherwise provide a specific revision to improve consistency

    If acceptable, use <target>GOOD</target>. Otherwise provide a specific revision to improve consistency

  66. [66]

    Be concise: <current> and <target> should each be 1 short sentence

  67. [67]

    move a bit

    Avoid vague revisions (e.g., "move a bit", "make bigger"). Always include at least one numeric target. User Prompt for Critique Generation (Spire) Evaluate the **Generated Slide** using the **Slide Request** below. Note: The request may include: (1) media image(s) to place in the slide, and (2) style reference slide(s) used as a STYLE GUIDE. Critique visu...

  68. [68]

    GRAPHIC_COLOR: [palette/contrast]

  69. [69]

    GRAPHIC_POSITION: [proportions with frames]

  70. [70]

    GRAPHIC_SIZE: [ratios]

  71. [71]

    IMAGE_POSITION: [zones]

  72. [72]

    IMAGE_SIZE: [ratios]

  73. [73]

    TEXT_COLOR: [contrast/emphasis]

  74. [74]

    TEXT_POSITION: [divisions]

  75. [75]

    Use simple language that a 10-year-old child can understand

    TEXT_SIZE: [hierarchy ratios; **legibility floor: all font sizes >= 10% of container height**] </think> ## STEP 3: COMPLETE DESIGN PLAN <plan> Write one paragraph in this order to state the complete design plan: background -> containers (e.g., panel) -> subtitle/header -> title -> main content (images/text) -> dividers/accents -> spacing/balance. Use simp...

  76. [79]

    Title Suppressed Due to Excessive Length 35

    Omit minor, complex visual effects (like subtle textures or complex shadows). Title Suppressed Due to Excessive Length 35

  77. [80]

    Do NOT use abstract theme keys (like ’accent_1’) as this causes ambiguity

    Be specific about colors: Use common color names (e.g., ’white’, ’black’, ’warm gold’) or specific RGB values. Do NOT use abstract theme keys (like ’accent_1’) as this causes ambiguity

  78. [81]

    User Prompt for Plan Generation (Spire) Generate a complete 3-step design plan (Analysis, Strategy, Plan) based on the user request below

    Be specific about containters: When introducing a container (like ’panel’ or ’ title_band’), name it using its ID (e.g., \...a container named **panel**\). User Prompt for Plan Generation (Spire) Generate a complete 3-step design plan (Analysis, Strategy, Plan) based on the user request below. Adhere strictly to the 3-step format defined in your system pr...

  79. [82]

    **Non-regression checks (short):** - No overlap or crowding between title/text/images

    TEXT_SIZE: ... **Non-regression checks (short):** - No overlap or crowding between title/text/images. - Text contrast is high against background. - Clear hierarchy (title > body). - Legibility floor satisfied. </think> ## STEP 3: COMPLETE DESIGN PLAN <plan> Write one paragraph in this order to state the complete design plan: background -> containers (e.g....

  80. [83]

    For any element placement, specify a bbox as x:

    Use proportions ONLY (no px/pt). For any element placement, specify a bbox as x:.. y:.. w:.. h:.. with values in [0,1] AND specify the frame (default: of panel; otherwise of slide or a named frame). x/y refer to the TOP-LEFT corner. Do not use center-point notation

Showing first 80 references.