Knowledge-Update Question Answering on LongMemEval (n=45 specific questions)
29Correct CountBaseline (BM25 + LLM-freshness)
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Baseline (BM25 + LLM-freshness)Pipeline=BM25 + LLM-freshness, Backbone=gpt-4o-mini, Freshness Signal=LLM-based2026.05 | 29 | 64.4 | 49.8 | |
| Deterministic pipeline (BM25 + extract + Python max(timestamp))Pipeline=BM25 + extract + Python max(timestamp), Backbone=gpt-4o-mini, Freshness Signal=max(timestamp)2026.05 | 26 | 57.8 | 43.3 |