[KongchangAI]
ConceptFew-shot推理 / Few-shot prompting

Few-Shot Reasoning

A prompting-based reasoning approach in which a small number of input-output examples are provided in the prompt to guide a large language model in performing a new task, without updating model parameters; it can substantially improve adherence to output formats and reduce hallucination rates.

Timeline (last 90 days)

Nov 6

Few-shot reasoning significantly reduced the hallucination rate: the strongest model produced unsupported QoS requirements in about 1% of responses, whereas performance under the zero-shot setting was markedly worse.

Unverified50%

All Facts (1)

Source Articles