Building Presentation Intelligence for High-Quality Slide Generation
High-quality presentation generation remains labor-intensive, and existing agentic frameworks rely mostly on foundation-model priors without task-specific training. This paper presents a unified training pipeline integrating cross-validated data synthesis, guided teacher enhancement for supervised fine-tuning, and verifiable aesthetic rewards via GRPO-based reinforcement learning. Experiments on PresentBench and GDPVal show that the pipeline improves both content quality and visual design, with compact models surpassing larger zero-shot baselines.
The paper advances a carefully structured training pipeline for end-to-end slide generation, and the paradigm-selection study (Single-HTML vs. Python vs. multi-page) is empirically solid. However, its contributions are undermined by two issues: the five verifiable aesthetic reward dimensions overlap heavily with the contemporaneous AeSlides work, and the gains on the main benchmark rely on an LLM judge rather than human or rule-based evaluation. While the SFT stage is convincing, the RL stage improves metrics that it directly optimizes, making the reported aesthetic gains partly circular.
The strongest element is the rigorous paradigm selection in Section 3 and Appendix A, which empirically justifies Single-HTML over Python and multi-page alternatives using both automated scoring and qualitative reasoning about math rendering and fault isolation. The SFT ablations are also well-executed: Table 2 demonstrates that mixed-teacher supervision consistently outperforms single-teacher baselines, and the outcome-driven trajectory filtering is lightweight yet generally beneficial. The claim that guidance improves structural regularity is supported by the 70.52 to 71.97 progression on PresentBench.
The paper characterizes AeSlides as limited to "next-slide generation," but the publicly available AeSlides abstract frames it as general single-pass slide generation using the same four core verifiable dimensions (aspect ratio, whitespace, collisions, visual balance), which severely weakens the novelty claim for the RL reward suite. The main benchmark evaluation is performed by Gemini-3-Flash-Preview, introducing opaque judge bias and cost barriers. Additionally, the SFT case study frankly admits that "noticeable aesthetic defects remain after SFT," yet the RL improvements on GDPVal are uneven—e.g., Qwen-3.6-27B total on GDPVal moves only from 3.35 to 3.37 in some RL settings—suggesting the visual refinement problem is not fully solved. Finally, the authors promise to "release the full codebase for the entire pipeline" yet no repository link appears in the paper; only HuggingFace model weights are listed.
Internal evidence supports the claim that teacher mixing and filtering help SFT, but the comparisons to prior work are almost entirely narrative. The authors state that methods such as SlideGen and PPTAgent "rely primarily on the intrinsic capabilities of foundation models without task-specific training," yet no empirical head-to-head comparison is provided under a shared generation framework. The RL evidence is also weak because the five aesthetic metrics are exactly the dimensions the policy is rewarded for, so Table 3 measures training-target alignment rather than independent quality. A fairer evaluation would report benchmark scores — not just aspect scores — after RL, but the paper does not update the overall PresentBench/GDPVal scores post-RL, only the self-referential aesthetic matrix.
The authors provide detailed hyperparameters (Adam, peak LR 5e-6, cosine decay, 64 GPUs, bf16, TP/CP/PP configurations) and release model weights on HuggingFace, which aids partial reproduction. However, the cross-validated synthesis pipeline, teacher prompts (Base/Pro/Max), and trajectory filtering rules are described in prose rather than released as code, despite the explicit promise to "release the full codebase." Reproduction is further gated behind access to proprietary teacher APIs (Qwen-3.8-Max, GLM-5.2) and the Gemini-3-Flash-Preview evaluator. The heavy infrastructure (64 GPUs, Megatron-based Turbo stack, sequence parallelism) also raises the barrier for academic replication, and the core RL training dataset of only 4,802 instances plus filtered subsets is small enough that variance could be high.
PowerPoint presentations (PPTs) are a dominant medium for communication across diverse domains, yet high-quality PPT generation still depends heavily on labor-intensive human effort. Recent advances in large language models (LLMs) for agentic tasks have created both new opportunities and challenges (Chen et al., 2026) for automatic PPT generation. In this work, we present a unified training framework for automatic PPT generation toward presentation intelligence, integrating cross-validated data synthesis, guided teacher enhancement, and verifiable reward optimization. We first construct a cross-validated pipeline for data synthesis, with explicit control over both difficulty and validity. We then develop a teacher-enhancement framework grounded in rules and prior knowledge to substantially strengthen the instructional capacity of teacher models. Building on this, we further improve supervised fine-tuning (SFT) through a mixed-teacher strategy and a lightweight outcome-driven trajectory filtering method. Finally, based on the resulting SFT models, we introduce a suite of verifiable aesthetic metrics together with a GRPO-based reinforcement learning framework to optimize slide-level visual attributes beyond text-only supervision. Experiments show that our enhanced SFT pipeline improves Qwen-3.6-27B and Qwen-3.6-35B-A3B by 7.45 and 9.35 points on PresentBench, respectively, surpassing GLM-5.2 while maintaining strong generalization. Reinforcement learning further improves performance across five aesthetic dimensions, compensating for the limitations of SFT in visual refinement. Overall, the results demonstrate the effectiveness of our framework for high-quality PPT generation.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.