EOS token
A token serving as the sequence terminator in autoregressive and diffusion language models; once it is committed, the decoding process ends and no further tokens are generated.
Timeline (last 90 days)
An early end-of-sequence (EOS) token directly truncates the answer and is the worst error case in parallel decoding.
During the aggregation stage, spatial relation information is lost in the global representation regardless of whether it is aggregated via the EOS token or mean pooling
EOS token aggregation and mean pooling are currently the two most mainstream methods for generating a global sentence vector from sequence representations
The EOS token aggregates global contextual information through the self-attention mechanism and is widely used in contrastive learning (e.g., CLIP) and retrieval tasks
All Facts (4)
An early end-of-sequence (EOS) token directly truncates the answer and is the worst error case in parallel decoding.
50%UnverifiedDuring the aggregation stage, spatial relation information is lost in the global representation regardless of whether it is aggregated via the EOS token or mean pooling
50%UnverifiedEOS token aggregation and mean pooling are currently the two most mainstream methods for generating a global sentence vector from sequence representations
50%UnverifiedThe EOS token aggregates global contextual information through the self-attention mechanism and is widely used in contrastive learning (e.g., CLIP) and retrieval tasks
50%