Back to Arena

H2O

by UT Austin / Rice / CMU / Stanford / Meta (Zhang et al.)

System Card

OrganizationUT Austin / Rice / CMU / Stanford / Meta (Zhang et al.)
Released2023-06
Architecturekv-cache-extension / Dynamic KV eviction of non-Heavy-Hitter tokens
DetailsFrames KV cache eviction as a dynamic submodular problem with theoretical guarantees. Retains a balance of recent tokens and "heavy hitters" (high attention-score accumulators), evicting the rest to shrink the cache.
Parameters—
Domainlong-context
Open SourceYes
PaperView Paper
CodeRepository
neurips-2023kv-evictionheavy-hitterthroughput

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
5/5
RULER
74.477p
NIAH
76.169p
LooGLE
75.941p
LongBench
603p
∞Bench
76.225p
Multi-Turn Recall
0/2
LoCoMo
no data
MemoryBank
no data
Cross-Session Memory
0/1
LongMemEval
no data
Multi-Hop QA
1/3
BABILong
73.941p
MultiHop-RAG
no data
HotpotQA
no data
Agent Task Memory
0/1
AgentBench-Mem
no data
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:H2O paper (arXiv:2306.14048); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)H2O paper (arXiv:2306.14048); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)H2O paper (arXiv:2306.14048); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)H2O paper (arXiv:2306.14048); evaluated on LooGLE: Can Long-Context Language Models Understand Long Contexts? (Peking University, 2311)H2O paper (arXiv:2306.14048); evaluated on Needle in a Haystack (Greg Kamradt, 2024)H2O paper (arXiv:2306.14048); evaluated on RULER: What's the Real Context Size of Your Long-Context Language Models (NVIDIA, 2404)