ICML Paper on Tweaking LLM Sampling via Prompts

量子位 · wechat · 2026-07-15

This article discusses an ICML 2026-accepted paper: instead of modifying model architectures or training new models, the authors merely tweaked the prompt to force the model to explicitly output sampling probabilities, thereby mitigating LLM **mode collapse**. Key takeaways include: - The paper proposes **Verbalized Sampling (VS)**, prompting the model to “verbalize” its sampling distribution. - The authors argue that convergent model outputs aren't just an algorithmic issue but are tied to **typicality bias** in preference data: human annotators favor fluent, familiar, and conventional answers. - Experiments show that in creative writing tasks, VS boosts diversity by **1.6 to 2.1 times** compared to standard prompts, without noticeably compromising factual accuracy or safety. - The article also highlights debates within the Reddit community: some feel the work relies too heavily on “prompt engineering” and lacks novelty, while others argue that solid assumptions, reasoning, and experiments make inference-time methods a valid research focus. The piece concludes by drawing parallels to the early days of CoT: what started as just “a single prompt” eventually evolved into a systematic field of study.

Related event: Prompt-only ICML paper ignites debate(2 posts)→

Original post →

More from Research

Research channel →