Savar Bhasin

Code LLMs can recognize incorrect instructions but still follow them, leading to irrecoverable semantic errors that defy traditional evaluation metrics.