Adaptive Retrieval on 50,000-candidate bank 2,000 query prefixes mid-network layer
7,918GEucDeepSeek-R1-14B
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| DeepSeek-R1-14Bd (Hidden-state dimension)=5120, l (Mid-network layer index)=24, sigma (Noise level)=5, k (Subspace dimension)=128, n_cal (Calibration count)={300, 500}2026.05 | 7,918 | 7.59 | 100 | 100 | 0 | |
| Qwen3-14Bd (Hidden-state dimension)=5120, l (Mid-network layer index)=20, sigma (Noise level)=5, k (Subspace dimension)=128, n_cal (Calibration count)={300, 500}2026.05 | 7,806 | 8.3 | 100 | 100 | 0 | |
| Mistral-7Bd (Hidden-state dimension)=4096, l (Mid-network layer index)=16, sigma (Noise level)=5, k (Subspace dimension)=128, n_cal (Calibration count)={300, 500}2026.05 | 1,043 | 3.81 | 0.3 | 0.9 | 0.1 | |
| Phi-2d (Hidden-state dimension)=2560, l (Mid-network layer index)=20, sigma (Noise level)=5, k (Subspace dimension)=128, n_cal (Calibration count)={300, 500}2026.05 | 144 | 2.76 | 100 | 100 | 0 | |
| GPT-2 Smalld (Hidden-state dimension)=768, l (Mid-network layer index)=6, sigma (Noise level)=5, k (Subspace dimension)=128, n_cal (Calibration count)={300, 500}2026.05 | 24 | 1.71 | 99.4 | 98.4 | 0.1 |