Method Drift›Long-context / context-window extension
H2O
H$_2$O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language ModelsLong-context / context-window extension · first seen Jun 24, 2023
superseded — cited as a baseline and beaten by newer methods
5 papers critique it · 3 beat it on benchmarks
What papers say
Verbatim critique sentences, each from a paper that cites H2O as a baseline.
“Although these methods generally have low additional overhead, they often lead to noticeable performance degradation.”
— A$^2$ATS: Retrieval-Based KV Cache Reduction via Windowed Rotary Position Embedding and Query-Aware Vector Quantization“While these methods differ in selecting tokens for KV cache retention, they generally apply a uniform budget size across layers, even though the optimal budget size may vary.”
— ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty“H2O maintains a fixed-size KV cache during decoding by dynamically evicting tokens.”
— CSAttention: Centroid-Scoring Attention for Accelerating LLM Inference“both rely on manually designed patterns or rules, which limits their ability to capture highly input-dependent attention sparsity”
— Long-Context Modeling with Dynamic Hierarchical Sparse Attention for On-Device LLMs“H2O~zhang2024h2o reduces memory costs and achieves better accuracy than StreamingLLM, but its reliance on attention maps makes it incompatible with the efficient attention implementation FlashAttention~dao2023flashattention, leading to slow attention computation.”
— LaCache: Ladder-Shaped KV Caching for Efficient Long-Context Modeling of Large Language Models
Beaten on benchmarks
Head-to-head results where a newer method reports beating H2O. Values are copied from the source paper's tables — verify against the cited paper.
- A$^2$ATS: Retrieval-Based KV Cache Reduction via Windowed Rotary Position Embedding and Query-Aware Vector Quantization
A2ATS beats H2O · Accuracy [Llama-3.1-8B-Instruct, Sparsity ~0.060]
86.6 vs 27.0
- A$^2$ATS: Retrieval-Based KV Cache Reduction via Windowed Rotary Position Embedding and Query-Aware Vector Quantization
A2ATS beats H2O · Accuracy [MegaBeam-Mistral-7B-512K, Sparsity ~0.062]
86.3 vs 22.3
- ZigZagkv: Dynamic KV Cache Compression for Long-context Modeling based on Layer Uncertainty
ZigZagKV beats H2O · Avg. [KV Size = 128]
43.30 vs 40.00
- CSAttention: Centroid-Scoring Attention for Accelerating LLM Inference
CSAttention beats H2O · Avg [Llama-3.1-8B-Instruct]
52.04 vs 45.11
- CSAttention: Centroid-Scoring Attention for Accelerating LLM Inference
CSAttention beats H2O · Avg [Qwen3-8B]
52.25 vs 48.04
- CSAttention: Centroid-Scoring Attention for Accelerating LLM Inference
CSAttention beats H2O · Avg [Mistral-7B-Instruct-v0.3]
49.92 vs 37.45
- CSAttention: Centroid-Scoring Attention for Accelerating LLM Inference
CSAttention beats H2O · Overall [Overall]
31.2 vs 29.9
Newer alternatives
Recent methods in the same sub-problem, not yet superseded in the knowledge base.