SlideLab
A training-free multi-agent framework that automatically generates research presentation slides directly from academic papers. Specialized collaborative agents handle content planning, visual generation, layout optimization, and grounding verification to complete the slide generation pipeline.
Timeline (last 90 days)
SlideLab is a training-free multi-agent framework capable of automatically generating research presentations directly from research papers.
SlideLab relies on the general capabilities of pre-trained large language models and completes tasks through prompts and a multi-agent collaboration structure, requiring no additional training steps.
SlideLab's workflow involves specialized agents working collaboratively, including a content planning agent, a visual generation agent, a layout optimization agent, and a grounding verification agent.
SlideLab first plans the narrative structure of the entire presentation, ensuring content is organized in logical order rather than simply reproducing the paper's sections.
The grounding verification agent checks whether content genuinely originates from the source paper, comparing each statement against the original paper text to avoid hallucinations.
In a blind human preference study, SlideLab was preferred by human reviewers on 77% of papers, outperforming both open-source and commercial systems.
The number of tokens consumed by SlideLab during inference is approximately one-quarter that of the strongest open-source baseline.
SlideLab is a newly released arXiv paper, and its practical generalizability and adaptability to papers from different disciplines still await broader validation.
All Facts (8)
SlideLab is a training-free multi-agent framework capable of automatically generating research presentations directly from research papers.
50%UnverifiedSlideLab relies on the general capabilities of pre-trained large language models and completes tasks through prompts and a multi-agent collaboration structure, requiring no additional training steps.
50%UnverifiedSlideLab's workflow involves specialized agents working collaboratively, including a content planning agent, a visual generation agent, a layout optimization agent, and a grounding verification agent.
50%UnverifiedSlideLab first plans the narrative structure of the entire presentation, ensuring content is organized in logical order rather than simply reproducing the paper's sections.
50%UnverifiedThe grounding verification agent checks whether content genuinely originates from the source paper, comparing each statement against the original paper text to avoid hallucinations.
50%UnverifiedIn a blind human preference study, SlideLab was preferred by human reviewers on 77% of papers, outperforming both open-source and commercial systems.
50%UnverifiedThe number of tokens consumed by SlideLab during inference is approximately one-quarter that of the strongest open-source baseline.
50%UnverifiedSlideLab is a newly released arXiv paper, and its practical generalizability and adaptability to papers from different disciplines still await broader validation.
50%