Improving the academic workflow: Introducing two AI agents for better figures and peer review

Google Research just dropped two research prototypes aimed squarely at the parts of academia nobody enjoys: making publication-quality figures and slogging through peer review. The first, PaperVizAgent (previously called PaperBanana), turns a manuscript's methods section and a figure caption into a polished diagram or statistical plot. The second, ScholarPeer, reads a paper and writes a review the way a skeptical senior researcher would — checking claims, hunting for missing baselines, and grounding everything in live literature search.

PaperVizAgent isn't a single model prompted once. It's five agents working in sequence: a retriever that pulls relevant reference figures from existing literature, a planner that organizes the visual content, a stylist that applies academic formatting conventions, a visualizer that actually renders the image or writes the plotting code, and a critic that checks the result against the source text and kicks it back for revision if something's off. That loop matters. On a 0-100 faithfulness-conciseness-readability-aesthetics scale calibrated against real human-made figures (with 50.0 as the human baseline), PaperVizAgent scored 60.2 overall — beating GPT-Image-1.5, Nano-Banana-Pro, and Paper2Any, and the only system tested that cleared the human baseline at all. It was especially strong on conciseness and aesthetics, and held its own on statistical plots too.

ScholarPeer takes a different approach to the same problem: reviewer burnout. Instead of generating review text directly, like most automated reviewers do, it splits the job into context-gathering and active verification. A