Concept分类式推理 / classification-style inference / Constrained Decoding
Constrained Decoding
A language model inference approach in which the model selects outputs by directly reading and ranking the probability values of tokens corresponding to predefined candidate answers, rather than autoregressively generating full text. It is commonly used in closed action spaces such as intent classification and multiple-choice evaluation tasks.
Timeline (last 90 days)
Oct 1
Traditional autoregressive language models sample from the entire vocabulary (typically tens of thousands to hundreds of thousands of tokens) or select the highest-probability token during generation.
Unverified50%