Search papers, labs, and topics across Lattice.
This study addresses the challenges of multi-turn refinement in scientific diagram generation, revealing that initial drafts often fail to meet user expectations, prompting requests for further revisions. By introducing MTPaperBananaBench, a benchmark for evaluating multi-turn diagram generation, the authors highlight two critical failure modes: quality drift and forgetting of previously implemented features. The proposed PaperBanana-Interact system effectively mitigates these issues, showing significant improvements in diagram quality and reduced forgetting across multiple iterations.
Multi-agent refinement can enhance diagram quality across multiple iterations, countering common pitfalls like quality drift and forgetting.
Recent efforts have aimed to automate scientific diagram generation from paper content (Lin et al., 2026; Zhu et al., 2026a). However, fully satisfying an author's visual and communicative preferences in a single turn is challenging: in our formative user study (N = 14), all participants requested further revisions after viewing an initial draft, and 86% of them rated the refined diagrams as more satisfactory. Despite the clear demand, the multi-turn workflow remains largely underexplored. To bridge this gap, we present MTPaperBananaBench, a benchmark for multi-turn diagram generation containing 292 images annotated with 3,518 user requirements. To reduce expensive human studies and enable scalable benchmarking, we construct a user simulator that, at each turn, identifies unsatisfied requirements and converts k of them into natural language feedback. Evaluating both requirement satisfaction and overall diagram quality reveals two key failure modes shared across baseline multiturn systems: (1) quality drift, where diagram quality progressively declines over turns, and (2) forgetting, where previously implemented features are lost in subsequent turns. To address these issues, we introduce PaperBanana-Interact, a multi-agent system that refines diagrams via an internal critique-and-refine loop. PaperBanana-Interact consistently improves rather than degrades diagram quality across turns, outperforming baselines by 11.9-18.6 points in quality score and reducing forgetting by 3.7-6.2 points.
