conceptEntrenamiento y Adaptación~1 min de lecturaActualizado 2026-06-07#alignment#dpo#preference-learning#rlhf
Esta nota todavía no está traducida, así que se muestra la fuente en inglés.

Direct Preference Optimization

Direct Preference Optimization (DPO) is a simpler preference-alignment method: given a prompt, a chosen response, and a rejected response, train the model to prefer the chosen one directly.

What DPO replaces

Classic RLHF trains a reward model and then optimizes a policy with PPO. DPO folds that preference objective into supervised-style training, using pairs like:

Field Meaning
Prompt The user request or task context
Chosen The preferred answer
Rejected The worse answer
Reference model Anchor for how far the model should move

The data still comes from human or high-quality evaluator preferences. DPO just removes the explicit reward-model and RL loop.

  • It is easier to implement and debug than PPO-based RLHF.
  • It uses ordinary training infrastructure.
  • It is stable enough for many instruction and style-alignment jobs.
  • It makes preference data directly useful.

What it is good for

DPO is best when you can express the improvement as "answer A is better than answer B": tone, helpfulness, refusal quality, concision, formatting preference, or policy behavior. It is less useful when you lack reliable preference pairs.

Pitfall

DPO quality is capped by pair quality. If "chosen" responses are only slightly better, or if annotators reward superficial polish over correctness, the model learns that bias.

Connects to: RLHF · data quality · evaluation