ProgramBench: Can Language Models Rebuild Programs from Scratch?

ProgramBench challenges language models to rebuild programs from scratch, earning 31 points and 19 comments—a modest reception for a niche benchmark. current top models achieve only 45% reconstruction accuracy, revealing a significant gap compared to the average human developer's ability. This benchmark is slower to gain traction than Appearing productive in the workplace's broader AI debate, but its mediocre results pinpoint a critical limitation in AI's software generation capabilities. For engineering leaders, ProgramBench offers concrete evidence that AI still struggles with fundamental code reconstruction tasks.