Needle-in-a-Haystack Retrieval on RULER 16K context length
29.4RULER-MV ScoreMemDLM (Train & Inference)
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| MemDLM (Train & Inference)Backbone=LLaDA-MoE, Setting=Train & Inference2026.03 | 29.4 | 56.84 | 57.24 | |
| MemDLM (Train-Only)Backbone=LLaDA-MoE, Setting=Train-Only2026.03 | 25.48 | 55.3 | 54.25 | |
| Standard MDLMBackbone=LLaDA-MoE2026.03 | 22.35 | 52.56 | 44.2 |