Needle-in-a-Haystack (NIAH) Retrieval on RULER 4K
100MV AccuracyMemDLM (Train & Inference)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| MemDLM (Train & Inference)Backbone=LLaDA2.12026.03 | 100 | 90.16 | 80.96 | |
| MemDLM (Train-Only)Backbone=LLaDA2.12026.03 | 99.65 | 88.56 | 78.96 | |
| Standard MDLMBackbone=LLaDA2.12026.03 | 98.65 | 59.32 | 72.34 | |
| MemDLM (Train & Inference)Backbone=LLaDA-MoE2026.03 | 96.8 | 99.6 | 83.7 | |
| MemDLM (Train-Only)Backbone=LLaDA-MoE2026.03 | 95.85 | 99.52 | 75.72 | |
| Standard MDLMBackbone=LLaDA-MoE2026.03 | 87.9 | 99.4 | 69.1 |