In a set of “24-point” experiments, reinforcement learning (RL), which adjusts a model’s behavior through trial, error, and outcome feedback, improved accuracy by 3.5% after the rules changed. Supervised fine-tuning (SFT), which trains the model using standard-answer demonstrations, declined by 8.1%. When only the card colors were replaced, RL improved by 17.6%, while SFT fell by 9.9%. These results matter because whether the training objective points toward “correct outcomes” is now shaping whether models can transfer to new situations.
What this is
The study set the J, Q, and K values to 10 during training, then changed them to 11, 12, and 13 respectively during testing. Another experiment replaced spades and clubs with hearts and diamonds. SFT was more prone to memorize “treat these cards as 10,” while RL, because it continuously adjusts its strategy based on whether answers succeed, was more likely to retain transferable problem-solving habits.
Common RL algorithms include GRPO, PPO, and DPO. GRPO compares multiple responses to the same question; PPO adds a step-by-step evaluation mechanism; DPO directly learns preferences between good and bad answers. The experiment demonstrates a tendency, not proof that RL can generalize across every task.
Industry view
We believe this gives AI application training a clear criterion: if a product must handle open-ended tasks such as changing rules, code debugging, or other tasks with variable inputs, relying only on standard-answer demonstrations makes it difficult to avoid dependence on surface patterns. Training loops built around verifiable outcomes are much closer to real-world requirements.
The risks are equally clear. RL training costs more, the process can fluctuate sharply, and flawed reward design can push the model away from the true objective. Engineering teams also need to control numerical discrepancies between the problem-generation model and the scoring model; otherwise, tiny probability differences can accumulate throughout long responses. RL and SFT are therefore not substitutes. In practice, RL usually still requires SFT first to establish basic comprehension and response formatting.
Impact on regular people
For enterprise IT: When enterprises deploy AI, training and model selection cannot rely on demo performance alone. Teams should also test stability after rules, inputs, and workflows change, while accounting for the cost of reward validation and continuous training.
For workplace professionals: Work that repeatedly applies fixed templates is easier to automate. Tasks requiring outcome judgment, experience transfer, and anomaly diagnosis will create new collaboration opportunities as models become more capable.
For the consumer market: Consumer-facing assistants will gradually move from answering similar questions toward handling personalized rules and multi-step tasks, but reliability and consistency still need validation in real-world settings.