Talk to Your Slides: Language-Driven Agents for Efficient Slide Editing

TL;DR

Talk-to-Your-Slides presents a language-driven agent for efficient slide editing by directly manipulating slide objects via a high-level/low-level architecture. The system combines instruction understanding, document understanding, document editing, and code generation to outperform GUI-based baselines in speed, fidelity, and cost on TSBench, a 379-instruction slide-editing benchmark. Experiments show the approach achieves higher execution success, instruction fidelity, and lower cost than GUI-based baselines, while highlighting that GUI-based methods still outperform in certain layout tasks due to visual context and OCR limitations. The work argues for future hybrid systems that balance precise, structured editing with visual awareness to maximize robustness and practicality for automated presentation editing.

Abstract

Editing presentation slides remains one of the most common and time-consuming tasks faced by millions of users daily, despite significant advances in automated slide generation. Existing approaches have successfully demonstrated slide editing via graphic user interface (GUI)-based agents, offering intuitive visual control. However, such methods often suffer from high computational cost and latency. In this paper, we propose Talk-to-Your-Slides, an LLM-powered agent designed to edit slides %in active PowerPoint sessions by leveraging structured information about slide objects rather than relying on image modality. The key insight of our work is designing the editing process with distinct high-level and low-level layers to facilitate interaction between user commands and slide objects. By providing direct access to application objects rather than screen pixels, our system enables 34.02% faster processing, 34.76% better instruction fidelity, and 87.42% cheaper operation than baselines. To evaluate slide editing capabilities, we introduce TSBench, a human-annotated dataset comprising 379 diverse editing instructions paired with corresponding slide variations in four categories. Our code, benchmark and demos are available at https://anonymous.4open.science/r/Talk-to-Your-Slides-0F4C.

Figures (14)

• Figure 1: Comparison of slide editing methods on translating 50-page lecture slides from Korean to English. (a) Manual translation requires day(s) and consumes graduate-student labor. (b) A GUI-based agent reduces human effort but incurs high cost and occupies the host machine during execution. However, (c) our approach runs in the background at a low cost and in a relatively short time.
• Figure 2: Overview of the Talk-to-Your-Slides framework. The system consists of four modules: instruction understanding, document understanding, document editing, and code generator.
• Figure 3: Example output generated by the instruction understanding module.
• Figure 4: Example output of document understanding. The yellow sections contain information about the parsed object's name, type, location, size, and other details. The runs highlighted in green demonstrate that different text formatting styles can exist within a single text box.
• Figure 5: Examples of instructions across four categories.