Needle-in-a-Haystack

The 'Needle-in-a-Haystack' dataset is a benchmark used to evaluate the retrieval accuracy of models in long-horizon inference tasks, specifically assessing their ability to locate relevant information within extensive contexts.

Papers using Needle-in-a-Haystack (9)

DepthWeave-KV: Token-Adaptive Cross-Layer Residual Factorization for Long-Context KV Cache Compression2026CompressKV: Semantic-Retrieval-Guided KV-Cache Compression for Resource-Efficient Long-Context LLM Inference2026CONF-KV: Confidence-Aware KV Cache Eviction with Mixed-Precision Storage for Long-Horizon LLM2026HISA: Efficient Hierarchical Indexing for Fine-Grained Sparse Attention2026MTraining: Distributed Dynamic Sparse Attention for Efficient Ultra-Long Context Training2025ChunkKV: Semantic-Preserving KV Cache Compression for Efficient Long-Context LLM Inference2025Can Compressed LLMs Truly Act? An Empirical Evaluation of Agentic Capabilities in LLM Compression2025PromptDistill: Query-based Selective Token Retention in Intermediate Layers for Efficient Large Language Model Inference2025Pause-Tuning for Long-Context Comprehension: A Lightweight Approach to LLM Attention Recalibration2025