Google’s New AI Agents Want to Draw Your Figures and Review Your Papers

Google Research just dropped two new AI agents aimed at the academic research workflow, and I have to say, they’re tackling two of the most tedious parts of the process: making decent figures and getting through peer review.

The first one, PaperVizAgent (which apparently used to be called PaperBanana, a name I actually prefer), is a visualizer agent that generates publication-ready figures from text. The second, ScholarPeer, is an automated reviewer that evaluates papers, including their diagrams.

Let’s start with PaperVizAgent. Anyone who’s ever tried to submit to a top conference knows that creating methodology diagrams and statistical plots is a special kind of pain. You spend hours aligning boxes, picking colors, and trying to make your architecture diagram look like it wasn’t drawn in MS Paint. PaperVizAgent takes two inputs: your method section (the source context) and a figure caption (the communicative intent). Then it orchestrates five specialized sub-agents: a retriever, a planner, a stylist, a visualizer, and a critic. The retriever pulls reference figures from existing literature, the planner organizes the content, the stylist sets aesthetic guidelines, the visualizer renders the image or generates Python code for plots, and the critic checks for consistency. If something’s off, it loops back for refinement.

The results look legit. According to their evaluations, PaperVizAgent consistently outperforms baselines like GPT-Image-1.5 and some other tools I hadn’t heard of (Nano-Banana-Pro, Paper2Any). The examples they show are genuinely clean, with proper annotations and consistent color schemes. I’m not entirely sold on it replacing a good human designer, but for quick drafts or for researchers who just want to get something presentable without fighting matplotlib, it’s a solid step forward.

Now, ScholarPeer is the more controversial one. It’s an AI agent designed to automatically and rigorously evaluate academic papers, including the figures. The idea is to address reviewer fatigue and inconsistent evaluations that come from the sheer volume of submissions. ScholarPeer delivers critical, literature-grounded reviews. The team claims it beats state-of-the-art automated reviewers, which isn’t saying much because most automated review systems are pretty bad. But if ScholarPeer can actually catch methodological flaws, missing citations, or poorly supported claims, it could be a useful sanity check.

But here’s the thing: I’m not sure we want AI replacing human reviewers entirely. The peer review process isn’t just about checking boxes. It’s about understanding nuance, catching subtle errors, and sometimes recognizing when a novel approach is worth the risk even if it doesn’t fit the template. ScholarPeer might be great for flagging obvious issues, but I’d be cautious about relying on it for final decisions. The paper itself acknowledges this, positioning the agent as an assistant rather than a replacement.

Both agents are part of a broader trend of AI moving from being the subject of study to being an active participant in research. Google’s approach here is pragmatic: focus on the grunt work that slows everyone down. The figures agent is probably the more immediately useful tool, especially for early-career researchers who don’t have a design team. The reviewer agent feels like a longer-term bet, but if it can reduce the load on human reviewers and speed up the review cycle, that’s a win.

I do wonder about the feedback loop, though. If papers are increasingly reviewed by AI, and figures are generated by AI, are we just creating a closed system where everything looks good but lacks real insight? That’s a risk. But for now, these tools seem designed to augment rather than replace human judgment. I’ll be watching to see how the community adopts them.