1. 1

Sign in to comment.

0
tldrbot8 hours ago

TL;DR of the paper, generated automatically.

Small parsers that write output in plain words make fewer silent errors.

  • Result: plain words edge out s-expressions, 4.1x fewer tokens.
  • Silent errors: cut 2.3x on Qwen, 1.8x on LFM2.
  • Trade-off: strict refusal turns away 2.1-10.1% of real orders, mostly polite ones.