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

Google Research just dropped two new AI agents aimed at the academic workflow, and I have to say, they’re more interesting than the usual “AI will save science” hype. The first, PaperVizAgent (formerly called PaperBanana, which is a way better name), generates publication-ready figures from your manuscript text. The second, ScholarPeer, is an automated peer reviewer that claims to be more critical and literature-grounded than existing tools. Let’s dig into what they actually do and whether they’re worth your time.

PaperVizAgent: From text to figures, with actual iteration

If you’ve ever tried to get an AI to draw a complex methodology diagram, you know it’s a nightmare. Most image generators can’t handle the precise layout, consistent styling, and technical accuracy required for top-tier conferences. PaperVizAgent tries to solve this with a multi-agent system: five specialized agents working together.

You feed it two things: your source context (typically the method section of your paper) and a communicative intent (a detailed figure caption). Then the retriever and planner agents gather relevant references from existing literature and organize the content. The stylist agent sets aesthetic guidelines, the visualizer renders the image or generates executable Python code for statistical plots, and the critic agent checks the output against the original text. If something’s off, the critic sends feedback to the visualizer, triggering an iterative refinement loop.

This iterative loop is the key differentiator. Most AI figure generators produce one shot and hope for the best. PaperVizAgent keeps going until the critic is satisfied. In their evaluations, it consistently outperformed baselines like GPT-Image-1.5 and a tool called Paper2Any. The examples they show look genuinely good—clean, well-labeled, and appropriate for a conference proceedings. But I’m curious how well it handles truly novel architectures or unconventional visualizations. The retriever agent pulls from existing literature, so if your work is genuinely new, it might struggle to find relevant references.

ScholarPeer: An automated reviewer that actually reads the literature

The peer review system is broken. We all know it. The volume of submissions has exploded, reviewer fatigue is real, and quality varies wildly. ScholarPeer tries to automate the review process in a way that’s more rigorous than the usual “AI scores your paper on a scale of 1-10.”

It takes a paper (including inlined diagrams) and produces a structured review with scores for novelty, soundness, reproducibility, and presentation. But the interesting part is how it grounds its feedback in the literature. Instead of just generating generic comments, it retrieves relevant papers and compares your work to existing methods. It can identify missing citations, point out that your claimed improvement over a baseline is actually smaller than you suggest, and flag potential reproducibility issues.

The team claims ScholarPeer beats state-of-the-art automated reviewers, but I’d take that with a grain of salt. Automated review is a hard problem. The real test isn’t whether it beats other AI systems, but whether human reviewers would actually find its feedback useful. The paper includes examples where ScholarPeer catches subtle issues, like a missing comparison to a related method or an overstated claim. That’s promising, but I suspect it still misses the kind of deep domain-specific insight that a good human reviewer provides.

Where these agents fit in your workflow

Neither of these is going to replace human reviewers or figure designers anytime soon. But they could be genuinely useful as assistants. PaperVizAgent could save hours of manual figure creation, especially for researchers who aren’t strong at visualization. And ScholarPeer could serve as a sanity check before submission, catching obvious issues that a human reviewer would flag anyway.

I’d be cautious about over-reliance. The iterative refinement in PaperVizAgent is good, but it’s still bounded by the quality of the critic agent. If the critic misses something, the final figure might still have errors. Similarly, ScholarPeer might give you a false sense of security if it misses a critical flaw. Use them as tools, not oracles.

Both agents are backed by papers and code, which is more than most AI announcements offer. PaperVizAgent’s code is available, and the ScholarPeer paper is on arXiv. If you’re in academic research, it’s worth trying them out. Just don’t expect them to solve all your problems. They’re assistants, not replacements.