Back to Arena

Compressive Transformer

by DeepMind (Rae et al.)

System Card

OrganizationDeepMind (Rae et al.)
Released2019-11
Architectureexternal-memory-network / Compacted past-activation memory + TransformerXL short-term
DetailsMaintains a TransformerXL-style short-term memory of past activations, but compresses old activations into a compressed memory instead of discarding them. Introduces PG-19 benchmark.
Parameters—
Domainlong-context
Open SourcePartial
PaperView Paper
WebsiteVisit
CodeRepository
iclr-2020deepmindpg-19compression

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
5/5
RULER
75.483p
NIAH
75.138p
LooGLE
77.655p
LongBench
603p
∞Bench
82.878p
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
74.142p
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:Compressive Transformer paper (arXiv:1911.05507); evaluated on BABILong: Testing the Limits of LLMs with Long-Context Reasoning-in-a-Haystack (AIRI, 2406)Compressive Transformer paper (arXiv:1911.05507); evaluated on InfiniteBench: Extending Long Context Evaluation Beyond 100K Tokens (Tsinghua / OpenBMB, 2402)Compressive Transformer paper (arXiv:1911.05507); evaluated on LongBench: A Bilingual, Multitask Benchmark for Long Context Understanding (Tsinghua KEG, 2308)Compressive Transformer paper (arXiv:1911.05507); evaluated on LooGLE: Can Long-Context Language Models Understand Long Contexts? (Peking University, 2311)Compressive Transformer paper (arXiv:1911.05507); evaluated on Needle in a Haystack (Greg Kamradt, 2024)Compressive Transformer paper (arXiv:1911.05507); evaluated on RULER: What's the Real Context Size of Your Long-Context Language Models (NVIDIA, 2404)