Question Answering on LongMemEvalS
33.8AccuracyContent salience
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Content salienceRetention budget=20%, Evaluation protocol=LLM-judge2026.06 | 33.8 | 36,500 | |
| LRE-supervisedRetention budget=20%, Evaluation protocol=LLM-judge2026.06 | 29.6 | 38,800 | |
| LRE-self-supRetention budget=20%, Evaluation protocol=LLM-judge2026.06 | 28 | 38,900 | |
| RecencyRetention budget=20%, Evaluation protocol=LLM-judge2026.06 | 27.6 | 26,100 | |
| Full context†Policy=No eviction, Retention budget=None2026.06 | 24.2 | 50,000 |