SlopCodeBench: Top Coding Agents Struggle as Requirements Evolve

UW–Madison's SlopCodeBench tests coding agents on evolving requirements, finding none of 15 models solved all 36 tasks, the best agent passed only 14.8% of checkpoints, and AI code was 2.3x more verbose than human code.

2026-10-19 ~ 2026-10-19 · 2 related posts

• SlopCodeBench: no model solves all 36 problems, AI code 2.3x more verbose than human repos — aj_marshall · 2026-10-19

1 near-duplicate retellings: aj_marshall