Google Research just dropped two new AI agents aimed at the academic research grind: PaperVizAgent (formerly called PaperBanana, which is a much better name) and ScholarPeer. One draws figures for you, the other reviews papers. Both are meant to take some of the drudgery out of the publishing pipeline.
Let’s start with PaperVizAgent, because that’s the one that actually surprised me. Anyone who’s tried to get an LLM to draw a decent figure knows it’s a mess. They can write passable text, but ask for a methodology diagram or a statistical plot and you get something that looks like a child’s crayon drawing. PaperVizAgent tackles this with a five-agent team: a retriever, a planner, a stylist, a visualizer, and a critic. You feed it your method section and a figure caption, and it goes through an iterative loop to produce something publication-ready. The critic agent checks the output against the original text and sends it back for refinement if something’s off.
The examples in the paper show methodology diagrams that look genuinely professional. They beat GPT-Image-1.5 and other baselines by a solid margin. That said, I’m curious how well it handles niche fields with very specific visual conventions. The stylist agent synthesizes aesthetic guidelines from existing literature, so it might struggle if your subfield uses unusual diagram types. But for most standard ML and CS papers, this looks like a huge time saver.
Then there’s ScholarPeer, which is the more controversial of the two. It’s an automated reviewer agent that evaluates papers, including inline diagrams. It generates what Google calls “highly critical, literature-grounded reviews” that outperform existing automated reviewers. That’s a low bar, honestly — most automated review systems are terrible. They miss obvious flaws and produce generic feedback. ScholarPeer seems to do better by grounding its evaluations in cited literature, which is a step in the right direction.
But let’s be real: the peer review system is already strained, and introducing AI reviewers comes with risks. If ScholarPeer becomes widely used, the fear is that journals and conferences might start relying on it too heavily, or that reviewers themselves will use it to generate lazy reviews. Google says it’s designed to assist, not replace, human reviewers. I hope that holds. The paper claims ScholarPeer delivers reviews that are “highly critical,” which is good — you want rigorous evaluation. But I’ve seen enough AI-generated nonsense to be skeptical until I can test it myself.
One thing I appreciate is that both agents are open-source. The code for PaperVizAgent is available, and the papers are on arXiv. That’s the right approach. Let the community poke at them, find the failure modes, and improve them. Closed-source research tools are a dead end.
Overall, these are practical tools that address real pain points. Figure generation is genuinely painful for many researchers, and automated review could help with the submission overload — if used responsibly. I’d like to see more real-world testing before I fully trust ScholarPeer, but PaperVizAgent looks like something I’ll actually use. Just don’t expect either to replace human judgment anytime soon.
Comments (0)
Login Log in to comment.
Be the first to comment!