Prompt interference and single-shot accuracy in inference-aware fine-tuning
A post announcing a paper’s acceptance at NeurIPS 2026 points to “prompt interference” as an explanation for why inference-aware fine-tuning often hurts accuracy on a single attempt.
TLDR
The post says pass@k optimization gives harder prompts more weight, potentially leading to gradient conflicts. It presents this “prompt interference” as an explanation for single-shot accuracy losses in inference-aware fine-tuning, announces the paper’s acceptance at NeurIPS 2026 and links to it on arXiv.
