The AI Was Right. The Answer Was Still Wrong.

Developer Akanksha Sharma explores the persistent issue of AI models failing to follow specific constraints during coding tasks. While modern LLMs are increasingly capable of solving complex programming problems, they often struggle with negative constraints or specific formatting requirements—such as avoiding certain methods or returning only raw code. To investigate this, Sharma developed a benchmark that evaluates models based on two distinct metrics: task correctness and instruction compliance. By testing multiple models with identical prompts, the study aims to identify patterns in how AI handles constraints and whether technical accuracy is compromised by a failure to follow instructions. The author emphasizes that for AI to be truly useful in professional development environments, it must adhere to specific project constraints rather than just providing a functional solution. The article invites developers to share their experiences with AI instruction failures to help refine future benchmarking efforts.