Datasets:
File size: 2,277 Bytes
17d50a7 5e9e5aa 17d50a7 5e9e5aa 17d50a7 5e9e5aa 17d50a7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 | {
"experiment": "PGAC Phase 2 \u2014 Probe-Gated Adaptive Compute on Qwen3.6-27B",
"corpus": {
"gsm8k": 30,
"simpleqa": 20,
"wikipedia": 40,
"total_residuals_per_layer": 39375
},
"primary_finding": {
"headline": "3.57x compound inference speedup at 8pp quality loss",
"config": "linear probe + top-(4k) prediction breadth + early-exit (1.67x avg)",
"best_pair": "L11\u2192L31 (gap 20, early-mid)"
},
"tradeoff_curve_L11_L31": [
{
"m": 128,
"recall": 0.4005,
"layer_speedup": 2.96,
"compound_w_exit": 4.94
},
{
"m": 256,
"recall": 0.5067,
"layer_speedup": 2.93,
"compound_w_exit": 4.89
},
{
"m": 512,
"recall": 0.6097,
"layer_speedup": 2.86,
"compound_w_exit": 4.78
},
{
"m": 1024,
"recall": 0.7002,
"layer_speedup": 2.73,
"compound_w_exit": 4.56
},
{
"m": 2048,
"recall": 0.7754,
"layer_speedup": 2.5,
"compound_w_exit": 4.18
},
{
"m": 4096,
"recall": 0.8375,
"layer_speedup": 2.14,
"compound_w_exit": 3.57,
"sweet_spot": true
},
{
"m": 8192,
"recall": 0.8972,
"layer_speedup": 1.67,
"compound_w_exit": 2.79
}
],
"sweet_spots_by_quality_budget": {
"8pp_loss_within_sae_floor": {
"m": 4096,
"compound_speedup": 3.57
},
"5pp_loss_tighter": {
"m": 8192,
"compound_speedup": 2.79
},
"12pp_loss_relaxed": {
"m": 2048,
"compound_speedup": 4.18
}
},
"layer_pair_3_experiment": {
"L11_L31": {
"gap": 20,
"recall_at_k_128": 0.4005,
"recall_at_2k": 0.5067
},
"L31_L55": {
"gap": 24,
"recall_at_k_128": 0.2796,
"recall_at_2k": 0.3671
},
"L11_L55": {
"gap": 44,
"recall_at_k_128": 0.2057,
"recall_at_2k": 0.2813
}
},
"verdict": "\ud83d\udfe2 STRONG with proper m selection \u2014 paper-grade positive contribution",
"paper_status": "ship",
"comparison_prior_art": {
"MxD_May_2025": "1.5-2x layer speedup",
"RouteSAE_2025": "training only, not inference",
"MoD_2024": "1.5x token-level skip",
"PGAC_this_work": "3.57x compound at iso-SAE-floor"
}
} |