Sequential Sampling
A candidate generation strategy for LLM inference in which the model references prior candidates when generating multiple candidates, using refinement prompts to guide progressive improvement of output quality and leveraging the LLM's in-context learning ability.
Timeline (last 90 days)
The first core component of the UvA-MT solution is sequential sampling with refined prompts.
The paper states that the sequential sampling strategy outperforms traditional parallel sampling.
In sequential sampling, poor quality of earlier candidates may cause errors to be inherited or even amplified by subsequent sampling.
This idea of sequential sampling plus pairwise voting can be transferred to other generation tasks such as summarization, polishing, and even code generation.
All Facts (4)
The first core component of the UvA-MT solution is sequential sampling with refined prompts.
50%UnverifiedThe paper states that the sequential sampling strategy outperforms traditional parallel sampling.
50%UnverifiedIn sequential sampling, poor quality of earlier candidates may cause errors to be inherited or even amplified by subsequent sampling.
50%UnverifiedThis idea of sequential sampling plus pairwise voting can be transferred to other generation tasks such as summarization, polishing, and even code generation.
50%