A system described in a paper submitted to arXiv on July 1, 2026 takes a scientific paper and returns a coherent multi-slide deck, a poster, and a video. OmniPresent surpasses strong baselines on both accuracy and visual appeal, according to its authors. It also removes a piece of friction that academic culture has quietly been reading as information.
What does OmniPresent actually build from a paper?
OmniPresent ingests a scientific paper and emits a suite of presentation artifacts: slides, posters, and a video, all kept semantically consistent with one another rather than allowed to drift.
The system’s core design choice is what the paper calls a renderable HTML representation for centralized content planning. The source paper is parsed into a structured, HTML-based representation, and every output format renders from that same canonical source rather than being generated as three independent prompts. That distinction is the real contribution. The existing business slide-generator category, tools like Beautiful.ai and Tome aimed at sales decks and pitch materials, takes a user prompt and produces a single set of slides. OmniPresent takes a full scientific paper and produces multiple, mutually consistent formats from it. The paper frames this as the first system built specifically for scientific papers with semantic consistency across modalities.
Read against the rest of an academic workflow, this targets a specific stretch. A finished manuscript already travels through editing, peer review, and a preprint server before it reaches an audience. The talk, the poster, and the conference video were the part of that journey still done largely by hand, the final step between a complete paper and a presented one. OmniPresent is built for exactly that step.
How does it keep the slides, poster, and video from contradicting each other?
OmniPresent runs a self-correcting verify-and-repair loop that detects and resolves conflicts across the generated slides, poster, and video before the suite is finalized, rather than generating each artifact and hoping they agree.
Generate a poster and a deck independently and they will disagree: a different headline number, a different framing of the contribution, a figure captioned one way on the slide and another way on the poster. The verify-and-repair loop is the engineering answer, checking for cross-modal inconsistencies and revising them. The paper describes the loop as actively resolving conflicts, which is a telling framing. It implies that without intervention, independent generation produces incoherent suites. Coherence across formats is the hard problem; any single format is not.
Can we even measure how good a generated presentation is?
To evaluate itself, the OmniPresent team released OmniPreBench, a dataset of over a thousand papers each paired with human-authored presentation artifacts, and reported that the system surpasses strong baselines on both accuracy and visual appeal.
The word accuracy does quiet work in that sentence. Most likely it means faithfulness to the source: no fabricated results, no misattributed claims, numbers that match the manuscript. Visual appeal is the softer axis and the harder one to defend. A summary is right or wrong about its source; a presentation makes rhetorical choices about what to foreground, and those choices have no single correct answer. Two competent decks from the same paper can emphasize different things and both be defensible. That makes a benchmark score on visual appeal a weaker claim than the same score on accuracy, and worth reading accordingly.
What was the conference talk actually signaling?
Here the argument leaves what the paper claims and becomes a reading of what the talk has been doing. The honest claim is that a polished seminar talk has functioned, in part, as an effort signal: visible proof that a scholar spent the hours distilling their work into something a room could follow.
Making a good talk is work. Compressing a paper into twenty minutes of clear slides forces decisions about what matters, and the time spent is partly why a crisp talk reads as evidence of command. Doing that compression well is a craft, and the craft exposes understanding. A scholar who has internalized the work chooses what to cut and what to emphasize, builds an arc a room can follow, and knows which figure earns its slide and which one is decoration. Those choices are legible. A practiced audience reads them as command, often without naming what they are reading. The slides themselves were never the whole signal. The choices embedded in them were.
The venue carries weight too. Academic conferences have grown, since the 1960s, into a sector on the order of a hundred billion pounds a year globally, according to Wikipedia’s overview of the conference industry. The scale matters because an enormous share of scholarly evaluation, from hiring talks to invited lectures to conference presentations, is mediated by this one artifact. When the artifact becomes cheap to produce, every one of those settings inherits the change.
What happens when the deck stops being a differentiator?
When every researcher can produce a visually competent deck, visual competence stops discriminating between scholars who understand their work and those who do not.
The floor rises to competent, and once it does, the worst deck in the seminar stops telling you anything. There was a time when a genuinely confusing slide deck was a weak but real signal about the speaker’s grasp of their own work. That information is about to become free to fabricate. The risk is not that bad talks become good. It is that mediocre work can be dressed in the visual vocabulary of rigorous work, and the dressing now costs nothing. A field whose evaluation partly relied on spotting the unprepared speaker by their slides loses that diagnostic and gains nothing equivalent in return.
Where does the pressure go once slides are free?
With the slides costing almost nothing to produce, the live question-and-answer period after a talk becomes one of the last moments where a scholar has to reason about their work without a model in the loop.
Evaluation pressure migrates to Q&A, because that is where depth has always shown and where the deck can no longer carry the weight. But the migration is not neutral. Q&A rewards quickness under pressure, and quickness under pressure is its own filter, one that favors fluency in a particular language, comfort with adversarial questioning, and a presentation style that not every competent researcher shares. The talk-as-effort-signal rewarded sustained preparation. Replacing it with a signal that rewards extemporaneous performance changes who gets read as authoritative, and not always toward the better scholar. One imperfect proxy is being traded for another, with the trade going unnamed.
What do tenure panels and hiring committees actually lose?
Committees that have treated a strong invited talk as evidence of a scholar’s command of their field are losing a proxy, and the honest position is that no replacement has been named.
As of the OmniPresent paper, there is no evidence that hiring committees, tenure panels, or grant reviewers have formally reckoned with AI-generated decks in their criteria. The proxy is being eroded before anyone has documented its loss. The replacement, if there is one, is more expensive than what it replaces: reading the work closely, probing depth in person, watching how a scholar handles a question they did not prepare for. All of that is slower and harder to do at volume than sitting through a polished talk, and the convenience of the talk-as-signal was always that it was cheap to consume. That cheapness depended on the talk being expensive to produce. Remove the expense and you remove the reason the signal was reliable.
The talk has always done two jobs at once: communicating findings, and proving the author did the work worth communicating. OmniPresent handles the first job competently. It does nothing for the second. What changes is that it removes the reason the second job ever got done for free, which was that producing the first job well took real effort. Once the effort becomes optional, the proof disappears with it, and the people who relied on that proof have not yet noticed it is gone.
Frequently Asked Questions
Could OmniPresent be adapted for legal briefs, policy reports, or journalism?
The current system is trained and benchmarked on scientific papers, so applying it to another domain would require a new parser for that source format and a paired dataset comparable to OmniPreBench. The renderable HTML representation is not generic; it encodes the structure of research manuscripts, and the verify-and-repair loop is tuned for conflicts among slides, posters, and videos of that kind.
What kind of error can the verify-and-repair loop miss?
It can miss mistakes that are consistent across all three artifacts because they originate in the shared HTML representation, such as a misread figure or a misattributed claim that gets rendered the same way on the slide, poster, and video. The loop resolves cross-modal conflicts, not errors that only appear when a human compares the output to the original paper.
What should a lab change in its workflow once decks are generated automatically?
Move review effort from slide polishing to source checking and Q&A rehearsal. A high accuracy score on OmniPreBench means the deck is faithful to the paper, not that the speaker understands it, so the final quality gate should be a human read of both the paper and the generated artifacts, plus preparation for live questions.
How does OmniPreBench differ from a standard text-summarization benchmark?
It judges a generated suite against human-authored presentation artifacts from over a thousand papers, so it tests resemblance to scholarly presentation conventions rather than factual correctness alone. Unlike a summarization benchmark that checks against one reference, it must also measure consistency across slides, poster, and video, which makes alignment harder and disagreement about visual choices harder to score.
What development would make the remaining Q&A signal unreliable?
If a model could generate plausible live answers to audience questions in the speaker’s voice, the Q&A period would stop being an unaided test of reasoning. OmniPresent only produces static artifacts, so the live exchange is still human; once that boundary falls, almost no part of the talk would reliably reflect the speaker’s own preparation.
