Skip to content
Last Dictionary
Plane
Training & Post-trainingReinforcement LearningAI Engineering & MLOps

reinforcement fine-tuning

nounAlso: RFT, RL fine-tuning, reinforced fine-tuning

In plain wordsImproving a trained AI model by trial and error instead of examples to copy: it tries answers, a grader scores them, and the model shifts toward whatever earned high scores.

  1. 1

    Fine-tuning a pretrained model with reinforcement learningreinforcement learning, n.: A way of teaching a computer program by trial and error: it tries actions, earns points for good results and penalties for bad ones, and gradually learns which choices pay off over time. rather than supervised targets: the model samples its own outputs for training prompts, a reward function, verifierverifier, n.: A program or AI model that checks whether a proposed answer is correct — by running tests, comparing with a known result or judging the reasoning — so that good answers can be picked or rewarded. or grader scores them, and a policy gradientpolicy gradient, n.: A family of methods that improve an AI's strategy directly: make actions that led to good results a bit more likely, and actions that led to bad results a bit less likely. method such as proximal policy optimizationproximal policy optimization, n.: A widely used method for training AI by trial and error that improves the strategy in small, safe steps; it was the workhorse behind training chatbots like ChatGPT with human feedback. or group relative policy optimizationgroup relative policy optimization, n.: A reinforcement-learning recipe for language models that has the model answer the same question many times and rewards each answer by how much better it did than the group's average. raises the probability of high-scoring outputs. Because the model learns from its own samples, reinforcement fine-tuning can exceed the quality of available demonstrations and tends to generalize better and forget less than supervised fine-tuningsupervised fine-tuning, n.: Teaching a pretrained AI model by showing it many examples of good answers to requests, so it learns to imitate them — like learning a job by studying worked examples., but it needs a reliable reward and a starting model that already succeeds some of the time. RLHFreinforcement learning from human feedback, n.: A way of training AI using people's judgments: humans compare pairs of answers, a scoring model learns their preferences, and the AI is then trained to produce answers that score well. and RLVRreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds. are both forms of it.

    SFT got us the format; reinforcement fine-tuning against the test-based grader got us the accuracy.

  2. 2

    Commercial service

    A customization service announced by OpenAI in December 2024 and offered for its o4-mini reasoning model in 2025, in which customers supply tasks and programmable graders (string checks, code or model-based scoring) and the provider runs RL fine-tuning for them; other platforms offer similar services.

Origin

RL fine-tuning of sequence models dates at least to Sequence Tutor (2016) and early RLHF. The phrase spread with ReFT (Trung Quoc Luong and colleagues, January 2024), which applied PPO after an SFT warm-up to math problems, and especially with OpenAI's Reinforcement Fine-Tuning Research Program, announced during its December 2024 product event.

Usage note

Abbreviation clash: RFT also means rejection samplingrejection sampling, n.: Generating many candidate answers and keeping only those that pass a check — a scoring model, a test or a correct final answer — often so the AI can then be trained on its own best work. fine-tuning, and ReFT also names representation fine-tuning, a parameter-efficient fine-tuningparameter-efficient fine-tuning, n.: A family of tricks for customizing a large AI model by training only a tiny fraction of its numbers, or a small add-on, leaving the rest untouched to save memory and money. method. As of October 2026, OpenAI had announced that it was winding down its fine-tuning API and platform, which includes RFT.

Related

See also
  • reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds.
  • reinforcement learning from human feedbackreinforcement learning from human feedback, n.: A way of training AI using people's judgments: humans compare pairs of answers, a scoring model learns their preferences, and the AI is then trained to produce answers that score well.
  • supervised fine-tuningsupervised fine-tuning, n.: Teaching a pretrained AI model by showing it many examples of good answers to requests, so it learns to imitate them — like learning a job by studying worked examples.
  • verifierverifier, n.: A program or AI model that checks whether a proposed answer is correct — by running tests, comparing with a known result or judging the reasoning — so that good answers can be picked or rewarded.
  • group relative policy optimizationgroup relative policy optimization, n.: A reinforcement-learning recipe for language models that has the model answer the same question many times and rewards each answer by how much better it did than the group's average.
  • fine-tuningfine-tuning, n.: Taking an AI model that has already learned a lot and training it a little more on a smaller, focused set of examples so it gets better at one job or behaves a certain way.
Contrast with
  • supervised fine-tuningsupervised fine-tuning, n.: Teaching a pretrained AI model by showing it many examples of good answers to requests, so it learns to imitate them — like learning a job by studying worked examples.
  • rejection samplingrejection sampling, n.: Generating many candidate answers and keeping only those that pass a check — a scoring model, a test or a correct final answer — often so the AI can then be trained on its own best work.

Sources

  1. Luong, T. Q., Zhang, X., Jie, Z., Sun, P., Jin, X. & Li, H. (2024). ReFT: Reasoning with Reinforced Fine-Tuning. arXiv:2401.08967. arxiv.org/abs/2401.08967
  2. OpenAI (2025). Reinforcement fine-tuning. OpenAI API documentation.

Mentioned in

fine-tuningfine-tuning, n.: Taking an AI model that has already learned a lot and training it a little more on a smaller, focused set of examples so it gets better at one job or behaves a certain way., reinforcement learning with verifiable rewardsreinforcement learning with verifiable rewards, n.: Training an AI by letting it attempt problems whose answers can be checked automatically — a math result, a passing program — and rewarding it only when the check succeeds., rejection samplingrejection sampling, n.: Generating many candidate answers and keeping only those that pass a check — a scoring model, a test or a correct final answer — often so the AI can then be trained on its own best work., supervised fine-tuningsupervised fine-tuning, n.: Teaching a pretrained AI model by showing it many examples of good answers to requests, so it learns to imitate them — like learning a job by studying worked examples.