The authors introduce projected distillation, which asks a stronger model to minimally edit a student's failed rollout into a correct answer.
Minimal corrections teach hard tasks while preserving broader model skills
Projected distillation repeatedly converts failed student attempts into nearby correct solutions, improving specialization without the same degree of capability loss seen with distant targets.
Big Tech
Shripad V. Deshmukh · Christian Bock · Kangrui Ruan · Abhijeet Awasthi · Vijay Lingam · Rein Houthooft · +1 more
AWS AI Labs · University of Massachusetts
Research Digest··3 min read
Deshmukh et al.
§