incentivizing

Definition

Getting a model to adopt a behaviour by rewarding it rather than by demonstrating it. The word carries a deliberate connotation in RL papers: the ability is treated as already latent in the pre-trained model, and reward merely draws it out — reward correct final answers, and longer reasoning traces appear on their own. Contrast supervised fine-tuning, where the behaviour is shown and copied.

Explainers using this term

• Designing Distillation Data — Deciding What to Ask the TeacherDeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning