 | tcballard 0 minutes ago | [–] Interestingly, a tuned length-threshold baseline performs very competitively and even beats Wayfinder on one of the benchmark sets. My intuition is that structural signals become more useful as prompt formats get more heterogeneous (code reviews, logs, tables, long instructions, etc.), but I’d love to see more real-world data. |
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。