Scissorhands

by Rice / Stanford / Meta (Liu et al.)

System Card

OrganizationRice / Stanford / Meta (Liu et al.)

Released2023-05

Architecturekv-cache-extension / Persistence-of-importance KV pruning

DetailsBased on the "persistence of importance" hypothesis: tokens with high past attention impact remain pivotal for future generations. Maintains a fixed KV budget by storing pivotal tokens with higher probability.

Parameters—

Domainlong-context

Open SourceNo

PaperView Paper

neurips-2023pruningkvquantization-compatible

Capability Profile

Benchmark Scores

6 of 14 benchmarks

Long-Context Retrieval

5/5

73.266p

72.223p

72.918p

603p

84.584p

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

75.652p

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:Scissorhands paper (arXiv:2305.17118); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)Scissorhands paper (arXiv:2305.17118); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)Scissorhands paper (arXiv:2305.17118); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)Scissorhands paper (arXiv:2305.17118); evaluated on LooGLE: Can Long-Context Language Models Understand Long Contexts? (Peking University, 2311)Scissorhands paper (arXiv:2305.17118); evaluated on Needle in a Haystack (Greg Kamradt, 2024)Scissorhands paper (arXiv:2305.17118); evaluated on RULER: What's the Real Context Size of Your Long-Context Language Models (NVIDIA, 2404)