Mocha Review
VOL. XII
ALIGNMENT
INTERPRETABILITY
MAY 2026
Research Paper · 2026
Models that
think
before they
speak.
An Essay on Latent Deliberation
延展推理(
extended reasoning
)让语言模型在产出前进行多轮内部审议。本研究展示了
4.7×
的推理深度提升如何转化为显著降低的幻觉率,以及在高风险部署场景下的
可对齐性
增强。
“ The space between question and answer is where alignment lives. ”
By Research Team
Continue Reading
→
M.R.
Editorial · Long Form
§ 04.2