UCLA x Google unveil PaperBanana-Interact: multi-turn chat refinement for scientific diagrams
kaiwei_chang · x · 2026-09-03
UCLA and Google Cloud AI researchers released PaperBanana-Interact, a system that makes AI-generated scientific diagrams conversational: generate a draft, then refine structure, components, and arrows via multi-turn chat.
Key points from the paper:
- A formative user study (N=14) found every participant requested revisions after the first draft, and 86% rated refined diagrams more satisfactory
- They release MTPaperBananaBench, a multi-turn diagram benchmark with 292 images and 3,518 annotated user requirements, benchmarked via a user simulator to cut human-study costs
- Evaluation exposes two failure modes in multi-turn systems: quality drift (progressive decline over turns) and forgetting (previously implemented features lost)
- Their fix is a multi-agent internal critique-and-refine loop
Code, data, and demo are coming soon; the paper's own figure was generated by the system.
Related event: UCLA and Google Release PaperBanana-Interact for Scientific Charts(2 posts)→
More from Research
- ACS Research Launches Free Platform Giving Researchers Access to Llama 3.1 Base Models — jankulveit · 2026-10-21
- LessWrong Essay Introduces LM Psychology: Just Talking to an AI Still Beats Fancy Evals — jankulveit · 2026-10-21
- Stanford researcher clashes over genomics foundation model's cherry-picked variant prediction benchmarks — anshulkundaje · 2026-10-21
- Gemini Agents Crack 59-Year-Old Fusion Math Conjecture in Hours — DynamicWebPaige · 2026-10-21
- Bombadil: a foundation model that decodes plants' bio-electrical signals — tdhopper · 2026-10-21
- TrackEverything: CMU and Meta open-source a 3D tracker for all points in 1000+ frame videos — AdamWHarley · 2026-10-21