The authors built 61 human-reviewed tasks from 34 repositories, spanning five primary programming languages and seven software forms, including command-line tools, services, interactive applications and libraries.
Coding agents struggle to complete repairs as software requirements evolve
FeatureLoop tests 12 models on staged repository tasks, finding that only 15.4% of observed failed stages recover within two corrective revisions.
Chinese Tech
Jia Feng · Alyssa Wu · Cuiyun Gao · Yunlong Wang · Songyuanyi Lu · Zhichao Hu
Hunyuan Team, Tencent · The Chinese University of Hong Kong
Research Digest··2 min read
Feng and colleagues introduce FeatureLoop, a benchmark that tests whether coding agents can extend software, respond to failure reports and preserve earlier functionality.
The Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.