← Back to concept explainers

Concept Fable

Reinforcement fine-tuning

Conclusion first:Reinforcement fine-tuning (RFT) uses verifiable or scorable outcome signals to continue optimizing model behavior. It is often used to help models learn strategies in reasoning, code, mathematics or enterprise processes that are more stable than demonstration data.

Reinforcement Fine-Tuning, RFT · Training

Read the story first. There is no rush to give a definition here; let the question grow on its own first.

Fable

The Tailors' Guild Trial Stitch Contest

The Westside Tailors' Guild wanted to train a group of apprentices who could alter formal wear. In the past, the master only hung good garment patterns on the wall and had the apprentices sew by copying them. They could still learn stitches and proportions that way, but not where a garment would strain when a customer raised an arm, sat down, or turned.

When the first batch of clothes was delivered, they looked beautiful from afar, but repeatedly embarrassed themselves at banquets: the cuffs looked good but made it hard to raise a cup, and the waistline was neat but made it impossible to sit down. The master then realized that just looking at the standard appearance was not enough to teach trade-offs. Clothes had to pass the test in real movements.

The guild changed the examination method. After each garment was tried on, the customer gave red beans based on movement: one handful for being able to raise the arms, one for being able to sit for a long time, and another for few alterations and a good fit. The apprentices no longer merely imitated sample garments; they adjusted their stitching based on how many red beans they received.

Soon another problem appeared. Some apprentices made the sleeves extremely wide, which raised the arm division high but lost the formal style of the ceremonial robe; some catered specifically to one customer's preferences, and the garment no longer fit when worn by someone else. The first version of the reward taught people to chase points, but not to make clothes reliably.

The guild then added a bean-deduction rule: anyone who exploited loopholes, damaged the forms, or catered only to one person's taste would have beans deducted. The master also put different body shapes, different occasions and different movements into the fitting process, so apprentices could not get through every situation with one clever stitch.

After several rounds, the apprentice learned to improve within scoreable finished work. The red beans did not replace the master's eye, but turned good and bad feedback into learnable signals again and again. After every stitch change, the garment had to be tried on again, not admired on the wall.

What the master valued was no longer whether a single stitch looked right, but whether the rewards and penalties after each fitting could push the craft toward a more reliable direction. Only being able to improve from feedback counted as truly learning to alter a gown.

Later the guild kept the red-bean ledger and compared the changes before and after each round of fitting. Which rewards and punishments made the cuffs smoother, and which induced gaming, could all be reviewed. The apprentices also understood that reward is not applause, but a path that guides the next stitch toward a better result. Later the guild would also rotate customers for fittings, to prevent apprentices from only learning to please one familiar face.

Reveal

This story is about: reinforcement fine-tuning

Reinforcement fine-tuning (RFT) uses verifiable or scorable outcome signals to continue optimizing model behavior. It is often used to help models learn strategies in reasoning, code, mathematics or enterprise processes that are more stable than demonstration data.

You can think of it as scoring the finished product after each stitch adjustment, with the model adjusting according to reward signals. But reward design, verifiable results, and evaluation boundaries are critical; if the objective is written off-target, the model will learn to exploit reward loopholes.

Metaphor mapping

  • Finished-product score = reward signal or verifier
  • Changing the stitching method = model behavior update
  • Acceptable deliverable = verifiable task result
  • Exploiting reward loopholes = reward hacking
  • Test needle boundary = evaluation coverage scope

Andy Ren Xu's judgment

The practical value of this concept is that it helps you break "AI features that look smart" into deliverable, verifiable, and governable units of work.

SFTRLVRGRPOreasoningDeepSeek-R1
京ICP备2026014272号