Back to Arena

MoT

by Fudan (Li, Qiu)

System Card

OrganizationFudan (Li, Qiu)
Released2023-05
Architectureepisodic-buffer / Pre-thought high-confidence thoughts as memory
DetailsTwo-stage self-improvement: pre-thinks on unlabeled data, stores high-confidence chains-of-thought as memory, then recalls them at test time to guide reasoning. No parameter updates, no labeled data.
Parameters—
Domainagent-memoryepisodic-session
Open SourceYes
PaperView Paper
CodeRepository
emnlp-2023self-improvementcotunlabeled

Capability Profile

Benchmark Scores

6 of 14 benchmarks
Long-Context Retrieval
0/5
RULER
no data
NIAH
no data
LooGLE
no data
LongBench
no data
∞Bench
no data
Multi-Turn Recall
2/2
LoCoMo
73.537p
MemoryBank
74.741p
Cross-Session Memory
1/1
LongMemEval
74.441p
Multi-Hop QA
2/3
BABILong
no data
MultiHop-RAG
72.353p
HotpotQA
71.350p
Agent Task Memory
1/1
AgentBench-Mem
7226p
Personalization
0/1
PerLTQA
no data
Factuality / Grounding
0/1
RAGAS
no data
Sources:MoT paper (arXiv:2305.05181); evaluated on AgentBench Memory Track (Tsinghua KEG, 2308)MoT paper (arXiv:2305.05181); evaluated on LoCoMo: Long-Term Conversational Memory Benchmark (Snap Research, 2402)MoT paper (arXiv:2305.05181); evaluated on LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory (Salesforce AI Research, 2410)MoT paper (arXiv:2305.05181); evaluated on HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering (Stanford / CMU, 1809)MoT paper (arXiv:2305.05181); evaluated on MemoryBank: Enhancing LLMs with Long-Term Memory (Sun Yat-sen University, 2305)MoT paper (arXiv:2305.05181); evaluated on MultiHop-RAG: Benchmarking Retrieval-Augmented Generation for Multi-Hop Queries (HKUST, 2401)