Why Pass@k Optimization Can Degrade Pass@1: Prompt Interference in LLM Post-training

Researchers identify a critical trade-off in AI model training where optimizing for Pass@k metrics (multiple attempts) degrades Pass@1 performance (single attempt). The study reveals this occurs due to gradient conflicts when the training process reweights toward low-success prompts, creating interference that hurts single-shot performance.

• →Pass@k optimization methods improve multi-sample performance but consistently degrade single-attempt (Pass@1) performance in large language models.
• →The degradation occurs due to gradient conflicts caused by prompt interference during training.
• →Pass@k optimization implicitly reweights training toward low-success prompts, which can negatively interfere with Pass@1 gradients.
• →This trade-off has significant practical implications since Pass@1 remains operationally important due to cost, latency, and reliability constraints.
• →The research provides theoretical characterization of when this degradation occurs in verifiable AI tasks like math reasoning and code generation.