推文

@fireandstart · 2026-10-14 06:07

Useful result, and the 14-point gap on the sparsest prompt is especially striking. I’d also compare calibration and abstention: can the decision endpoint flag uncertain cases for review, and does that change the end-to-end error rate or cost? The output shape only earns its place if it helps the whole workflow on representative data.

曝光 17 · 评论 1 · 点赞 0 · 书签 0 · 曝光/时 4.283587399153833