Question Answering on LoCoMo (CSM, RM, Delta Metrics)
0.55CSMEPC
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| EPCProtocol=Cost-Matched (CM), Write Budget (B)=2K, Readers=2-reader average (GPT-5.2 and Claude Sonnet 4)2026.05 | 0.55 | 0.46 | 0.06 | 0.09 | |
| Summ. (L)Protocol=Cost-Matched (CM), Write Budget (B)=2K, Readers=2-reader average (GPT-5.2 and Claude Sonnet 4)2026.05 | 0.51 | 0.44 | 0.1 | 0.07 | |
| ReadAg.Protocol=Cost-Matched (CM), Write Budget (B)=2K, Readers=2-reader average (GPT-5.2 and Claude Sonnet 4)2026.05 | 0.49 | 0.41 | 0.12 | 0.08 | |
| MemWlk.Protocol=Cost-Matched (CM), Write Budget (B)=2K, Readers=2-reader average (GPT-5.2 and Claude Sonnet 4)2026.05 | 0.39 | 0.33 | 0.22 | 0.06 | |
| MemBankProtocol=Cost-Matched (CM), Write Budget (B)=2K, Readers=2-reader average (GPT-5.2 and Claude Sonnet 4)2026.05 | 0.37 | 0.31 | 0.24 | 0.06 | |
| Summ. (E)Protocol=Cost-Matched (CM), Write Budget (B)=2K, Readers=2-reader average (GPT-5.2 and Claude Sonnet 4)2026.05 | 0.33 | 0.27 | 0.28 | 0.06 | |
| VerbatimProtocol=Cost-Matched (CM), Write Budget (B)=2K, Readers=2-reader average (GPT-5.2 and Claude Sonnet 4)2026.05 | 0.29 | 0.24 | 0.32 | 0.05 |