WebRISE Catches What Visual AI Benchmarks Miss in Web Generation

Ask any multimodal LLM to build you a checkout page, and it will probably produce something that looks great. Buttons are styled, layouts are clean, the visual score is high. But click "Add to Cart" with nothing selected, and the checkout button stays enabled when it shouldn't. The total price never updates. The page looks right but behaves wrong. This is exactly the gap that a new benchmark called WebRISE was built to measure.

The evaluation gap nobody was talking about

The field of MLLM-generated web artifacts has exploded. Models are now routinely asked to generate full interactive web pages from text prompts, wireframe sketches, screenshots, or even screen-recording videos. Benchmarks have kept pace on the visual side, scoring pixel fidelity and layout quality. But established benchmarks focus on algorithmic correctness and are blind to the visual fidelity and interactive integrity that define modern user experiences.

The deeper problem is that existing evaluation approaches check local evidence: did this button render? Does this element exist in the DOM? They never ask whether the sequence of interactions a user would actually perform leads to the right page states. A form that renders perfectly but never disables its submit button on empty input is broken -- and prior benchmarks would not catch it.

Interaction Contract Graphs: the core idea

WebRISE introduces a new abstraction called an Interaction Contract Graph (ICG). Think of it as a formal specification of how a web page is supposed to behave, expressed as a graph of states and transitions. Each node is an observable page state (e.g., "cart is empty", "item selected", "checkout enabled"). Each edge is a user-intent transition (e.g., "user clicks Add to Cart"). Each transition carries DOM assertions (does the right element exist and have the right attribute?) and visual assertions (does the screenshot match the expected appearance?).