OSPD sharpens persona consistency through self-distillation

OSPD is a clever post-training recipe that turns privileged context into supervision, making persona fidelity more about internalized behavior than ever-longer prompts.

• –Role-aware KL switching focuses imitation on character-critical tokens while preserving diversity in generic dialogue.
• –Progressive trait masking pushes identity, personality, knowledge, style, and history into model parameters.
• –OSPD improves CharacterBench from 3.00 with SFT to 3.51 and CharacterEval persona consistency from 2.90 to 3.49.
• –On-policy rollouts appear more valuable than teacher scale: OSPD beats a 72B off-policy distillation baseline despite using a much smaller teacher.
• –The tradeoff is compute: training takes about 18.6 hours versus 4.2 for SFT, and results may depend strongly on model scale and LLM-judge evaluation.