Latency on S-NIAH (Long-context retrieval)
10.5Latency (s)Direct
Evaluation Results
| Method | Links | |
|---|---|---|
| DirectParadigm=P1: Direct, Backbone=Mistral, Model Size=7B2026.03 | 10.5 | |
| DirectParadigm=P1: Direct, Backbone=Llama, Model Size=8B2026.03 | 11.8 | |
| DirectParadigm=P1: Direct, Backbone=Qwen3, Model Size=8B2026.03 | 12.3 | |
| DirectParadigm=P1: Direct, Backbone=Mistral, Model Size=Cdsrl2026.03 | 16.8 | |
| DirectParadigm=P1: Direct, Backbone=Qwen3, Model Size=32B2026.03 | 18.7 | |
| DirectParadigm=P1: Direct, Backbone=Mistral, Model Size=8x22B2026.03 | 20.2 | |
| DirectParadigm=P1: Direct, Backbone=Llama, Model Size=70B2026.03 | 21.4 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Llama, Model Size=405B2026.03 | 28.1 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Qwen3, Model Size=235B2026.03 | 31.4 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Llama, Model Size=70B2026.03 | 35.6 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Mistral, Model Size=Cdsrl2026.03 | 36.5 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Qwen3, Model Size=32B2026.03 | 38.2 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Mistral, Model Size=8x22B2026.03 | 41.8 | |
| DirectParadigm=P1: Direct, Backbone=Qwen3, Model Size=235B2026.03 | 42.1 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Qwen3, Model Size=8B2026.03 | 45.9 | |
| DirectParadigm=P1: Direct, Backbone=Llama, Model Size=405B2026.03 | 48.6 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Llama, Model Size=8B2026.03 | 48.7 | |
| λ-RLMParadigm=P3: λ-RLM, Backbone=Mistral, Model Size=7B2026.03 | 52.3 | |
| RLMParadigm=P2: RLM, Backbone=Llama, Model Size=405B2026.03 | 86.3 | |
| RLMParadigm=P2: RLM, Backbone=Qwen3, Model Size=235B2026.03 | 98.6 | |
| RLMParadigm=P2: RLM, Backbone=Mistral, Model Size=Cdsrl2026.03 | 120.4 | |
| RLMParadigm=P2: RLM, Backbone=Llama, Model Size=70B2026.03 | 128.4 | |
| RLMParadigm=P2: RLM, Backbone=Qwen3, Model Size=32B2026.03 | 142.8 | |
| RLMParadigm=P2: RLM, Backbone=Mistral, Model Size=8x22B2026.03 | 155.1 | |
| RLMParadigm=P2: RLM, Backbone=Qwen3, Model Size=8B2026.03 | 164.3 | |
| RLMParadigm=P2: RLM, Backbone=Llama, Model Size=8B2026.03 | 178.2 | |
| RLMParadigm=P2: RLM, Backbone=Mistral, Model Size=7B2026.03 | 195.7 |