Research
Hugging Face paper proposes LoCo, a local preference method to cut reasoning length
LoCo trains models with DPO on state-conditioned preference pairs to decide when to stop reasoning. It reports 41.7% to 45.4% shorter completions on math and 23.4% to 36.3% on code, trained only on math, with code and model promised publicly.
Every outlet that ran it 1 article
Hugging Face PapersResearchLead
Think Less, Locally: Local Preference Learning for Reasoning Compression