You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Eliciting In-context Retrieval and Reasoning for Long-context Large Language Models
Published Date
2025-1-14
Source
arxiv
Head Name
Retrieval Head
Summary
Innovation: Proposed retrieval hit rate, a more direct metric to evaluate attention heads.
Tasks: In long-context retrieval tasks, identify the retrieval head and utilize it to improve the model's retrieval performance.
Significant Result: Through Joint Retrieval Head Training, a new retrieval head can be obtained to predict the Top-K context for generating final responses, enhancing the model's in-context retrieval and reasoning capabilities.
The text was updated successfully, but these errors were encountered:
Title
Eliciting In-context Retrieval and Reasoning for Long-context Large Language Models
Published Date
2025-1-14
Source
arxiv
Head Name
Retrieval Head
Summary
The text was updated successfully, but these errors were encountered: