Accuracy (%) on LifeBench 2.0 and LoCoMo,
across three base LLMs. Click any column header to sort.
Paper · Dataset
SH Single-hop · MH Multi-hop · TR Temporal · ND Non-declarative · KU Knowledge update ·
CR Causal · CD Conflict detection · HI Hidden information · UA Unanswerable ·
Micro micro-average · Macro macro-average (mean of the 9 categories) · LoCoMo overall (excludes adversarial questions).
— = not reported. Gold Evidence† feeds the annotated evidence directly to the answer model (reader under perfect retrieval) and is excluded from the per-column best.