PaperBanana-Interact: Scientific Diagram Refinement with Multi-Turn Human Feedback
The paper targets a specific weak spot in AI-made scientific figures: revision over multiple rounds without degrading the diagram.
The authors introduce MTPaperBananaBench, a 292-image benchmark with 3,518 annotated user requirements for multi-turn diagram refinement. Their user study found that all 14 participants wanted revisions after an initial draft, and 86% preferred the refined diagrams. Baseline systems showed two recurring failures: quality drift over turns and forgetting earlier fixes. PaperBanana-Interact uses a multi-agent critique-and-refine loop and reports higher quality scores and less forgetting than the tested baselines. HF Daily Papers' note
The authors introduce MTPaperBananaBench, a 292-image benchmark with 3,518 annotated user requirements for multi-turn diagram refinement. Their user study found that all 14 participants wanted revisions after an initial draft, and 86% preferred the refined diagrams. Baseline systems showed two recurring failures: quality drift over turns and forgetting earlier fixes. PaperBanana-Interact uses a multi-agent critique-and-refine loop and reports higher quality scores and less forgetting than the tested baselines. HF Daily Papers' note
score 4