Google’s New AI Agents Can Draw Your Figures and Review Your Paper

Google Research just dropped two new AI agents aimed at fixing some of the most tedious parts of academic research: making decent figures and getting through peer review. As someone who’s spent way too many late nights wrestling with matplotlib and waiting on reviewer comments, I’m cautiously optimistic.

The first one is PaperVizAgent (formerly called PaperBanana, which I honestly prefer). It’s a multi-agent system that takes your paper’s method section and a figure caption, then produces a publication-ready diagram. The second is ScholarPeer, an automated reviewer that evaluates papers with a claimed rigor that beats existing baselines.

PaperVizAgent: From text to figures

PaperVizAgent works by orchestrating five specialized AI agents: a retriever, planner, stylist, visualizer, and critic. You feed it your source context (typically the method section) and your communicative intent (the figure caption). The retriever grabs relevant examples from existing literature, the planner organizes the content, the stylist enforces aesthetic guidelines, the visualizer renders the image or generates Python code for statistical plots, and the critic checks for consistency. If something’s off, the critic loops back to the visualizer for iterative refinement.

This is higher than I expected from a system that’s essentially generating figures from scratch. The examples they show look genuinely professional, with clean lines, proper annotations, and consistent color schemes. The iterative refinement loop is key here, because without it, you’d end up with the kind of generic diagrams that plague preprint servers.

But let’s be real: the real test is whether it can handle the messiness of real research. My own papers have figures that went through 20+ revisions because the reviewer wanted a different statistical test or a different layout. PaperVizAgent’s critic agent might catch some of that, but I suspect domain-specific quirks will still require human intervention.

ScholarPeer: The reviewer that never sleeps

ScholarPeer is the more ambitious of the two. It’s designed to automatically evaluate academic papers, including inlined diagrams, with a level of rigor that Google claims beats state-of-the-art automated reviewers. The idea is to help with the peer review bottleneck, where the explosion of submissions has led to reviewer fatigue and inconsistent evaluations.

From the technical description, ScholarPeer uses a multi-agent framework similar to PaperVizAgent, but focused on evaluation rather than generation. It can assess methodology, statistical validity, and even the quality of figures. Google’s evaluations show it outperforms existing baselines on a range of metrics.

This approach has been tried before, and the results have been mixed. Automated review systems tend to be good at catching formatting errors and basic statistical issues, but they struggle with novelty assessment and nuanced methodological critiques. ScholarPeer’s literature-grounded approach might help with the latter, but I’m skeptical about its ability to understand truly novel contributions.

The broader picture

Google is clearly betting on AI agents as active participants in the scientific process, not just as subjects of study. Both PaperVizAgent and ScholarPeer are designed to let researchers focus on innovation rather than administrative overhead. That’s a noble goal.

But there’s a tension here. The same tools that streamline research could also homogenize it. If everyone uses the same AI to generate figures and review papers, we might end up with a monoculture of visual styles and evaluation criteria. The iterative refinement loops in PaperVizAgent might help with diversity, but the underlying models are still trained on the same corpus of existing papers.

I also wonder about the cost. Google hasn’t said much about pricing or availability, but running multi-agent systems with iterative refinement loops isn’t cheap. If these tools end up behind a paywall, they’ll widen the gap between well-funded labs and everyone else.

Bottom line

PaperVizAgent and ScholarPeer are serious attempts to solve real problems in academic research. The multi-agent approach with iterative refinement is a smart design choice, and the early results are impressive. But like any tool, they’ll be judged by how well they work in practice, not in demos.

For now, I’ll keep an eye on the code releases. If PaperVizAgent can save me even a few hours of figure tweaking, it’ll be worth it. As for ScholarPeer, I’ll believe it when I see it handle a paper with a genuinely novel methodology without defaulting to “this doesn’t match the literature.”