Stanford team releases SOTA prompt optimization framework with unsupervised prompt performance measure
BrihiJ · x · 2026-10-29
Stanford researchers (Alexander Spangher, Daniel E. Ho, Diyi Yang, Noah Goodman, Sanmi Koyejo) present a SOTA prompt optimization framework plus a novel unsupervised measure of prompt performance. They argue written rubrics and constitutions are crucial for task improvement, safety following and pluralism — but some desired rubric behaviors are hard to put into words (how do you describe "comic voice"?). Core question: which written constructs can a judge evaluate well, and how good can it get?
More from Research
- Scott Aaronson on AI cracking decades-old math problems — and whether P vs. NP is next — romanyam · 2026-10-29
- Stanford's new CS 312 makes students build DL theory with Claude via experiment prediction — srush_nlp · 2026-10-29
- Andon Labs: AI Now Ties Humans on Physical 3D Understanding Benchmarks — ResultBackground2450 · 2026-10-29
- LMArena launches an arena-style benchmark for agent capabilities — allenainie · 2026-10-29
- H-Net v2: Byte-Level Modeling Gets 3-5x Efficiency Gains, Matches 11T-Token Model with 10x Less Data — SonglinYang4 · 2026-10-29
- EMNLP 2026 talk: explainable and robust automatic fact-checking needs evidence, not just verdicts — IAugenstein · 2026-10-29