RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development
arXiv cs.CL · 20 Oct 2026, 03:55 AEST
RefCraft introduces a benchmark for evaluating visual-reference-conditioned agentic generation in webpage development. The study highlights the challenges faced by current models in effectively utilizing visual references for practical tasks.
Key facts
- Beyond basic tool use and coding skills, effectively incorporating visual references into agentic generation is equally essential for applying AI agents to real-world productivity tasks.
- We introduce RefCraft, a benchmark comprising 300 realistic multimodal webpage-development tasks that combine natural-language instructions with screenshots from one to five reference webpages.
- Agents must selectively interpret and utilize these references while rendering, visually inspecting, and iteratively refining their outputs.
- Across nine frontier (M)LLMs, the strongest system achieves an overall score of only 52.12 out of 100, revealing that visual-reference-conditioned agentic generation remains a substantial challenge for current models in practical webpage-development settings.
Read aloud
RefCraft introduces a benchmark for evaluating visual-reference-conditioned agentic generation in webpage development. The study highlights the challenges faced by current models in effectively utilizing visual references for practical tasks. Beyond basic tool use and coding skills, effectively incorporating visual references into agentic generation is equally essential for applying A I agents to real-world productivity tasks. We introduce RefCraft, a benchmark comprising three hundred realistic multimodal webpage-development tasks that combine natural-language instructions with screenshots from one to five reference webpages. Agents must selectively interpret and utilize these references while rendering, visually inspecting, and iteratively refining their outputs. Across nine frontier M L L M s, the strongest system achieves an overall score of only fifty two point one two out of one hundred, revealing that visual-reference-conditioned agentic generation remains a substantial challenge for current models in practical webpage-development settings.
Sources in this cluster
- RefCraft: Benchmarking Visual-Reference-Conditioned Agentic Generation through Web Development
arXiv cs.CL · 20 Oct 2026, 03:55 AEST