{"id":383497,"date":"2026-11-08T04:12:00","date_gmt":"2026-11-08T12:12:00","guid":{"rendered":"https:\/\/cms-articles.softonic.io\/en\/?p=383497"},"modified":"2026-11-08T04:12:16","modified_gmt":"2026-11-08T12:12:16","slug":"co-design-bench-is-out-coding-agents-still-need-human-cleanup","status":"publish","type":"post","link":"https:\/\/cms-articles.softonic.io\/en\/co-design-bench-is-out-coding-agents-still-need-human-cleanup\/","title":{"rendered":"Co-Design Bench is out: coding agents still need human cleanup"},"content":{"rendered":"

Co-Design Bench is out. Built by researchers at Alibaba and several Asian universities, it gives coding agents a harder, more realistic test: editing existing web apps<\/strong>, not spinning up polished demos from scratch.<\/p>\n

\r\n
\r\n
\r\n
\r\n \"Visual\r\n <\/div>\r\n
\r\n Visual Studio Code<\/span>\r\n Download<\/a>\r\n <\/div>\r\n
\r\n <\/path><\/path><\/text><\/svg>\r\n <\/div>\r\n <\/div>\r\n
\r\n <\/span>\r\n <\/div>\r\n
\r\n \r\n <\/div>\r\n <\/a>\r\n <\/div>\r\n<\/div>\n\n\n\n

In the paper, the team says it started with 37,463<\/strong> problematic real user requests and narrowed them to 273 tasks. Those tasks include multimodal inputs such as images, receipts, audio, video, tables, and documents, and they span everything from small layout or data edits to full project overhauls.<\/p>\n\n

In Co-Design Bench’s 10-model test, GPT-6 Astra came in first, but even then it fully completed 26.4%<\/strong> of tasks and met 69.9% of the criteria on average, according to the researchers. Claude Opus 5 and Qwen3.8-Max followed behind it. Grok 4.6 landed at the bottom. One wrinkle: Qwen3.8-27B placed second on the larger overhaul tasks, though the authors stop short of saying smaller models are better in general.<\/p>\n\n

If you use coding agents, this is one to watch. The paper says every model<\/strong> struggled when a request mixed layout, design, interaction, and content changes in the same job. Older coding benchmarks are already saturated above 99%. Co-Design Bench still leaves a gap of more than 35 points, using strict browser-based grading and a second judge that checks whether an agent leaned on mock-ups or hardcoded outputs.<\/p>\n\n

The benchmark also tracks with SWE-bench Verified results, the researchers say: GPT-4o at 30.7%<\/strong>, GPT-4.1 at 54.6%, and Qwen3-Coder at 67.4%. That lines up with the broader point here. These tools still need debugging and review from you.<\/p>\n\n

The paper and the benchmark are available online from the researchers.<\/p>","protected":false},"excerpt":{"rendered":"

Co-Design Bench is out. Built by researchers at Alibaba and several Asian universities, it gives coding agents a harder, more realistic test: editing existing web apps, not spinning up polished demos from scratch. In the paper, the team says it started with 37,463 problematic real user requests and narrowed them to 273 tasks. Those tasks … Continue reading “Co-Design Bench is out: coding agents still need human cleanup”<\/span><\/a><\/p>\n","protected":false},"author":9350,"featured_media":383496,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":"","wpcf-pageviews":0},"categories":[1015],"tags":[],"usertag":[],"vertical":[],"content-category":[6771],"class_list":["post-383497","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-news","content-category-ai"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t