Meituan LongCat releases RAP: 25K paper-poster pairs as a visual coding benchmark

meituan · hf · 2026-10-19

Meituan LongCat, with Tsinghua, HKU, and UW, released RAP, a large-scale dataset and benchmark for paper-to-poster generation.

Key points:

• 25K+ verified paper-poster-markup triplets from human-crafted posters at NeurIPS (44.3%), ICLR (33.4%), and ICML (22.3%), spanning 2022-2026
• Frames poster generation as a visual coding task: generate declarative markup with layout, OCR, image assets, and design captions, then render via constrained reconstruction
• Provides both retrieval-augmented generation (template pool + asset replacement + HTML rendering) and an instruction-conditioned VLM approach

Data, models, and demo are publicly released, making RAP the largest open benchmark in this niche.

Original post →

More from Research

• Google closes the loop: agents now run real biology lab robots and found a cancer-binding molecule — aigclink · 2026-10-19
• Study: 60.4% of security agent tokens go to code understanding; AVRI cuts costs up to 23.7% — rohanpaul_ai · 2026-10-19
• NVIDIA paper shows cross-model KV cache transfer skipping prefill, 2.7–25x faster than recomputing — solyarisoftware · 2026-10-19
• HIL-UMI: Robots Learn From Handheld Demos Without Touching Real Hardware — jiqizhixin · 2026-10-19
• Redditor proposes crowdsourced session-data repo to help open LLMs close the frontier gap — kingslayerer · 2026-10-19
• Paper: Bregman divergences as power distances, enabling Euclidean NN queries — FrnkNlsn · 2026-10-19

Research channel →