Needle-In-A-Haystack (NIAH)

The Needle-In-A-Haystack (NIAH) dataset/benchmark is used to evaluate retrieval performance in hybrid linear-attention models, particularly assessing their long-context recall capabilities.

Papers using Needle-In-A-Haystack (NIAH) (1)

Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It2026