Co-Design Bench is a reference-free benchmark consisting of 273 tasks derived from 37,463 real user coding sessions. It evaluates agentic models on multimodal editing and repository-level revisions across grounded understanding and functional implementation.
HOW THIS AFFECTS YOU
●
builderUse this benchmark to validate how well your coding agents handle UI-driven or image-based refactoring requests.
●
researcherThis provides a more realistic evaluation framework for agentic coding than standard text-only benchmarks.