Targeted retrieval on Needle-in-a-Haystack (NIAH)
99.5Mean RecallQwen3-14B
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen3-14BTraining Source=Off-the-shelf reference models2026.06 | 99.5 | |
| Qwen3-32BTraining Source=Off-the-shelf reference models2026.06 | 99.3 | |
| CONTEXTRLTraining Source=Qwen3-8B, Training Strategy=CONTEXTRL2026.06 | 99 | |
| Base modelTraining Source=Qwen3-8B, Training Strategy=Base model2026.06 | 98.8 | |
| RL baselineTraining Source=Qwen3-8B, Training Strategy=RL baseline2026.06 | 98.5 | |
| Qwen3-Coder-30BTraining Source=Off-the-shelf reference models2026.06 | 85.7 | |
| CONTEXTRLTraining Source=Klear-AgentForge-8B, Training Strategy=CONTEXTRL2026.06 | 71.3 | |
| Base modelTraining Source=Klear-AgentForge-8B, Training Strategy=Base model2026.06 | 68.3 | |
| RL baselineTraining Source=Klear-AgentForge-8B, Training Strategy=RL baseline2026.06 | 65.5 |