Post-training digest: distillation leaks teacher behavior, reward models surpass teachers

A digest of post-training research published October 9–16, 2026. Key findings: distillation transfers a teacher's safety disposition even on unrelated tasks, RM-OPD produces a distilled reward model that beats its specialist teachers, and Apertus 1.5 documents a full four-stage SFT → RLVR → DPO → SDPO stack.

DEV Community

Open source

Post-training digest: distillation leaks teacher behavior, reward models surpass teachers

Imprinting: distillation on unrelated tasks still transfers the teacher's safety disposition
RM-OPD: distilled reward model surpasses the specialist teachers it learned from
Apertus 1.5: SFT on 3.7M samples, RLVR on 350k prompts, DPO and SDPO stages
UltraBERT: cascade distillation helps classification but hurts retrieval at high masking

#HuggingFace#Apertus#UltraBERT

Post-training digest: distillation leaks teacher behavior, reward models surpass teachers

LLM post-training digest: GRPO moves to OCR, DPO and GRPO get debugged

AI/ML Research Digest: self-distillation, MoE pruning and long context

Teachers at Euan Blair's Multiverse report 'horrendous stress' over AI monitoring

Napster Developing AI Teacher Clones to Provide Personalized Homework Support