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. One draws your figures for you, the other reviews your paper. I’ve been following their work for a while, and this is one of those releases that sounds gimmicky but actually has some substance.

PaperVizAgent: Making your figures less ugly

Let’s be honest — most researchers are terrible at making figures. I’ve seen papers with methodology diagrams that look like they were drawn in MS Paint during a coffee break. PaperVizAgent (they originally called it PaperBanana, which is a much better name) is a multi-agent system that takes your manuscript text and a figure caption, then generates publication-ready illustrations.

The system uses five specialized agents: a retriever that finds relevant examples from existing literature, a planner that organizes the content, a stylist that figures out what looks good, a visualizer that actually renders the image or writes Python code for statistical plots, and a critic that checks if the output matches what you described. If the critic finds problems, it sends feedback back to the visualizer and they iterate until it’s right. That iterative refinement loop is key — it’s what separates this from just slapping a prompt into an image generator.

According to their evaluations, PaperVizAgent consistently beats GPT-Image-1.5, Nano-Banana-Pro (I’m not making that name up), and Paper2Any. The examples they show are genuinely impressive — clean methodology diagrams, properly formatted statistical plots, the kind of stuff that usually requires either hiring a graphic designer or spending hours in Illustrator.

ScholarPeer: The reviewer that actually reads your paper

ScholarPeer is the more interesting of the two, and potentially more controversial. It’s an automated peer reviewer that evaluates papers — including inline diagrams — against the existing literature. The system retrieves relevant papers from the web, cross-references claims, and generates structured reviews with actionable feedback.

Their results show ScholarPeer delivering reviews that are more critical and more grounded in literature than existing automated reviewers. That’s a low bar, honestly — most automated review systems are glorified grammar checkers. But ScholarPeer actually seems to engage with the content. It checks whether your methodology is properly motivated, whether your results are statistically sound, whether you’ve cited relevant prior work.

The obvious concern here is that this could be used to game the system. If everyone starts writing papers optimized for ScholarPeer’s scoring criteria, we end up with a monoculture of AI-friendly research. But Google’s framing is more modest — they position this as a tool for pre-submission quality checks, not as a replacement for human reviewers.

The bigger picture

I’ve been watching the academic publishing space for years, and the volume problem is real. Top conferences like NeurIPS and CVPR are seeing submission numbers double every few years. The reviewer pool can’t keep up. Something has to give.

These agents won’t solve that problem on their own, but they’re a step toward automating parts of the workflow that don’t require human judgment. Figure generation is a perfect candidate for automation — it’s tedious, it’s time-consuming, and it doesn’t require creativity in the way that writing the actual research does. Peer review is trickier, but even partial automation of the grunt work — checking citations, verifying statistical claims, flagging potential issues — could free up reviewers to focus on the substantive evaluation.

The real test will be whether the community actually adopts these tools. Google has released code for PaperVizAgent, which is a good sign. ScholarPeer’s availability is less clear from the announcement. If they keep these locked behind Google Cloud APIs, adoption will be slow. If they open-source them, we might actually see meaningful changes in how researchers produce and evaluate papers.

I’m cautiously optimistic. The agents themselves are technically solid, and the problem they’re addressing is real. But I’ve seen too many AI tools for academia that end up as demos rather than actual workflow changes. Let’s see if this one sticks.