Can shorter answers be smarter? Researchers say yes.

New method DLER pairs simple truncation with better RL tricks to cut output length by >70% while improving accuracy over prior baselines, such as batch reward normalization, higher clipping, dynamic sampling.

It also scales… pic.twitter.com/75UXfQ0a6K

— 机器之心 JIQIZHIXIN (@jiqizhixin) 9 novembre 2025

Can shorter answers be smarter? Researchers say yes. New method DLER pairs simple truncation with better RL tricks to cut output length by >70% while improving accuracy over prior baselines, such as batch reward normalization, higher clipping, dynamic sampling. It also scales

→ View original post on X — @jiqizhixin