Prompt-only ICML paper ignites debate

An ICML 2026 paper has sparked debate because it improves LLM diversity without changing model architecture or training a new model. Instead, it uses prompt modifications to make the model verbalize sampling probabilities, aiming to mitigate mode collapse and raising questions about whether such prompt-engineering work belongs at a top ML conference.

2026-07-13 ~ 2026-07-15 · 2 related posts

• ICML Paper Sparks Debate: Is Prompt Engineering Top-Tier ML Material? — Mean_Revolution1490 · 2026-07-13
• ICML Paper on Tweaking LLM Sampling via Prompts — 量子位 · 2026-07-15