Back to Arena

Self-RAG

by University of Washington / Allen AI (Asai et al.)

System Card

OrganizationUniversity of Washington / Allen AI (Asai et al.)
Released2023-10
Architectureagentic-workflow / Self-reflective on-demand retrieval with reflection tokens
DetailsTrains a single LM that adaptively decides when to retrieve, then emits reflection tokens to critique retrieved passages and its own generations. Reflection tokens make the LM controllable at inference time.
Parameters—
Domainrag-retrievalagent-memory
Open SourceYes
PaperView Paper
WebsiteVisit
CodeRepository
iclr-2024-oralreflection-tokensadaptive-retrievalfactuality

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Data Transparency:6 estimated
Long-Context Retrieval
1/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
603pEstimated
∞Bench
no data
Multi-Turn Recall
1/2
LoCoMo
77.673pEstimated
MemoryBank
no data
Cross-Session Memory
1/1
LongMemEval
79.171pEstimated
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
73.967pEstimated
HotpotQA
77.890pEstimated
Agent Task Memory
1/1
AgentBench-Mem
7226pEstimated
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:Arena estimate — derived from capability profile, not independently verified