Original title: ExBind: A Controlled Diagnostic Benchmark for Visual-to-Executable Correspondence
Ziqian Wang, Yuxiao Cheng, Tingxiong Xiao, Jinli Suo · September 2026
When AI systems edit websites, diagrams, or tables, they must connect what they see on screen to the exact piece of underlying code. If an AI points to the wrong element, its edit fails, but regular tests cannot pinpoint why. To fix this, researchers built ExBind, a diagnostic test suite that isolates this visual-to-code connection. ExBind generates test cases across formats like SVG, DOM, canvas, trees, graphs, and tables. An AI is shown an element and must output only the strict code target. The benchmark then checks if the choice satisfies structural rules. The researchers found that Qwen2.5-VL-3B picked real targets 98.4% of the time but only achieved 76.4% exact accuracy. Meanwhile, Qwen3-VL-4B reached 100.0% validity and 98.8% exact accuracy. Shuffling target orders changed individual outcomes without altering error patterns.
Imagine you're guiding a friend to pick a specific spice from a cluttered shelf. They might grab a valid spice bottle but not the one you intended. In this paper, ExBind tests if AI can identify the exact spice by giving it visual clues and checking the reference it outputs.
When AI systems help edit code by pointing to visual elements, they can select a valid but incorrect part, causing silent errors. This benchmark helps developers identify and fix such pointing inaccuracies, making coding assistants more reliable.
ExBind is designed for controlled diagnosis, not for population-scale ranking or end-to-end editing evaluation, so it does not assess overall system performance or general application scenarios.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
ExBind: A Controlled Diagnostic Benchmark for Visual-to-Executable Correspondence — Ziqian Wang, Yuxiao Cheng, Tingxiong Xiao et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.