跪拜 Guibai
← All articles
Artificial Intelligence · LLM · Deep Learning

OPD: How Large Models Now Distill Judgment, Not Just Answers

By 吴佳浩Alben ·
Read original on juejin.cn ↗ Google Translate ↗ Alt translation

RLHF pipelines are expensive and brittle, and high-quality human preference data is drying up. OPD offers a cheaper, scalable path to alignment by turning an existing strong model into a reusable judge, making it practical to align a whole fleet of smaller, deployable models.

Summary

Traditional knowledge distillation copies a Teacher model's output probabilities, transferring facts but not taste. OPD changes the signal: the Teacher ranks multiple candidate answers from the Student, and the Student learns to maximize the probability of that ranking. This shifts the training objective from fitting a static token distribution to learning a relative preference relationship.

The process is online because the Student generates new candidates after each parameter update, and the Teacher re-scores them in a co-evolving loop. A full PyTorch implementation shows a simulated Student's Teacher score rising from near zero to 0.88 over 150 rounds, without ever seeing a correct answer. A production-ready example uses Qwen as the Student and GPT-5 as the ranking Teacher, with TRL's DPOTrainer handling the Bradley-Terry loss.

Industry adoption is driven by cost: running full RLHF for every Student size is impractical. OPD reuses a single strong Teacher's judgment to train an entire family of smaller models, cutting the per-model alignment cost. The technique also enables preference transfer chains, where a 14B model trained via OPD can itself become the Teacher for a 3B model.

Takeaways
OPD replaces the Teacher's logit output with a ranking signal over multiple Student-generated candidates.
The Student learns a relative preference (B > A > C) rather than an absolute token distribution.
Training is online: the Student generates new candidates each round, and the Teacher re-scores them dynamically.
A minimal PyTorch implementation shows a simulated Student's alignment score rising from ~0 to 0.88 in 150 rounds without any correct-answer labels.
A production pipeline can pair a small open model like Qwen2.5-7B with GPT-5 as the ranking Teacher, using TRL's DPOTrainer.
OPD enables preference transfer chains: a 14B model trained this way can then teach a 3B model.
Without a variance constraint (like a KL penalty), the Student will cheat by inflating its exploration variance, collapsing generation quality.
Conclusions

OPD reframes alignment as a ranking problem, which is a better fit for subjective quality than the point-estimate approach of traditional distillation.

The online loop creates a co-evolution dynamic that can amplify small initial preference signals into strong behavioral shifts.

Preference transfer chains suggest a future where alignment is a cascading, automated process rather than a one-off human annotation effort per model size.

The Bradley-Terry loss is the common mathematical thread linking RLHF, DPO, and OPD; the innovation is in who provides the ranking and when.

Concepts & terms
Online Preference Distillation (OPD)
A training method where a Teacher model ranks multiple candidate answers generated by a Student, and the Student updates its parameters to make higher-ranked answers more probable. The process is 'online' because the Student generates fresh candidates after each update, and the Teacher re-evaluates them.
Bradley-Terry Model
A statistical model for pairwise comparisons. It estimates the probability that one item is preferred over another as a function of the difference in their latent quality scores, typically using a sigmoid function. It underpins the loss functions in RLHF, DPO, and OPD.
Preference Transfer
The process of cascading alignment through a chain of models, where a larger model teaches a smaller one via OPD, and that smaller model can then teach an even smaller one, preserving judgment capability while reducing model size and cost.
Source: juejin.cn ↗ Google Translate ↗ Backup ↗