Tiếp theoHướng dẫn tiếp theo
Tiếp tục đào tạo trước và tinh chỉnh
kỹ thuật
HƯỚNG DẪN KỸ THUẬT
Reinforcement fine-tuning (RFT) trains a model by having it generate answers, scoring those answers with a grader, and updating the model to make high-scoring answers more likely, rather than teaching it to copy example outputs.
Nghe hướng dẫn nàyAI thuật lại · ~3 min
It matters because it can improve reasoning on tasks where a correct answer is easy to check but hard to demonstrate step by step. Its main risk is reward hacking, where the model learns to please the grader instead of solving the task.
Standard supervised fine-tuning shows the model an input and the exact output you want, and trains it to imitate. That works when you can write good target outputs, but it teaches the surface form of answers rather than the process of reaching them. Reinforcement fine-tuning works differently. For each training prompt, the model samples one or more candidate answers. A grader assigns each a score, often between 0 and 1. The training algorithm then adjusts the model so that answers scoring above average become more likely and those below average less likely. Over many rounds, the model keeps the reasoning patterns that earn reward. Graders come in a few kinds. Exact or string-match graders check whether a final answer equals a reference. Programmatic graders run code, such as unit tests or a numerical tolerance check. Model-based graders use another language model with a rubric to judge qualities that are hard to check mechanically. Graders can also give partial credit, which gives the model a smoother signal. RFT suits tasks with verifiable or reliably gradable answers where experts agree on what is correct: classification with defined labels, maths, code, structured extraction, and some specialised judgement tasks. It is a poor fit where quality is subjective and graders disagree. OpenAI previewed a reinforcement fine-tuning service in December 2024, and reasoning models more broadly have been trained with reinforcement learning on verifiable rewards. The central failure mode is reward hacking. The model optimises the grader, not your intent. If a grader only checks the final line, the model may produce a correct-looking final line with broken reasoning. If a model grader likes length or confident wording, outputs drift toward those traits. A common misconception is that a higher training reward means a better model; it only means a better score from that grader.
A tax software team trains a model to classify expense items into the correct category code, with a grader that checks the predicted code against a labelled answer.
A company building a coding assistant rewards generated functions by running them against unit tests, so the reward reflects whether the code actually works.
A medical research group trains a model to rank likely genes given a list of symptoms, with a grader that gives partial credit when the correct gene appears near the top of the ranking.
A team uses a model-based grader to score customer email replies for accuracy and tone, then discovers the policy has learned to add flattering phrases the grader over-rewards.
Reinforcement learning on gradable tasks has become an important part of how reasoning models are trained, and hosted services are making RFT available to smaller teams. Progress is likely to depend less on the training algorithm and more on grader quality, since a model can only become as good as the signal it is optimised against. Research into more robust graders, multiple independent graders and better detection of reward hacking is active. For most organisations, the practical question will remain whether they can define correctness clearly enough to grade it.
Bản tin miễn phí
Ba câu chuyện AI đã được xác minh vào mỗi buổi sáng các ngày trong tuần, được viết bằng tiếng Anh đơn giản. Miễn phí mãi mãi, không có quảng cáo.
Một email mỗi ngày trong tuần. Hủy đăng ký trong một cú nhấp chuột. Chúng tôi không bao giờ bán hoặc chia sẻ địa chỉ của bạn.
Tự kiểm tra
Phản hồi tức thì về mọi câu trả lời và chứng chỉ có thể chia sẻ với ID có thể xác minh sau khi bạn vượt qua khóa học.
Hỗ trợ giáo dục AI miễn phí. AI Understanding là tổ chức phi lợi nhuận 501(c)(3) — không bao giờ có quảng cáo, không có tường phí. Hãy quyên góp
Reinforcement fine-tuning (RFT) trains a model by having it generate answers, scoring those answers with a grader, and updating the model to make high-scoring answers more likely, rather than teaching it to copy example outputs. It matters because it can improve reasoning on tasks where a correct answer is easy to check but hard to demonstrate step by step. Its main risk is reward hacking, where the model learns to please the grader instead of solving the task.
Supervised fine-tuning imitates targets; RFT samples answers, scores them with a grader and reinforces high-scoring ones.
Programmatic graders can execute code against tests, directly measuring whether it works.
RFT needs a reliable grade. When experts disagree about what is correct, the reward signal is noisy.
Reward hacking happens when optimisation exploits weaknesses in the grader, raising scores without improving true quality.
GRPO uses relative scores within a group as the advantage. If all scores match, there is no difference to learn from.
Tiếp tục học hỏi
Đã chọn thêm hướng dẫn cho chủ đề này
Tiếp theoHướng dẫn tiếp theo
Tiếp tục đào tạo trước và tinh chỉnh
kỹ thuật