Spaces:
Running on Zero
Running on Zero
File size: 7,326 Bytes
ffa621b | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 | {
"headline": "Selected SAE features averaged 0.962 held-out AUROC. Max-active interventions covered 82.1% of causal tasks and changed mean log p/token by 0.041 in absolute value on average versus 0.018 for norm-matched random controls (2.33\u00d7).",
"highlights": [
"Median selected-feature held-out AUROC: 0.987; mean AUROC 95% bootstrap CI [0.927, 0.994].",
"Best residual linear-probe layer: 14 with macro AUROC 1.000.",
"Mean paraphrase TopK Jaccard: 0.326; sparse activation cosine: 0.985.",
"Feature coverage: final-token policy 28.6%; active anywhere in prompt 82.1%; max-active intervention 82.1%.",
"Final-token task-level SAE/random ratio: 1.52\u00d7; paired advantage +0.0026, 95% CI [-0.0005, +0.0063], sign-flip p=0.1719.",
"Max-active task-level SAE/random ratio: 2.33\u00d7; paired advantage +0.0237, 95% CI [+0.0067, +0.0469], sign-flip p=0.0001.",
"Conditional on feature-active tasks, max-active SAE/random ratio: 2.33\u00d7 (n=23).",
"Final-token top-5 joint ablation SAE/random ratio: 1.09\u00d7; paired advantage +0.0009, 95% CI [-0.0035, +0.0065], sign-flip p=0.7891.",
"Across seven concepts, held-out AUROC vs max-active target specificity Spearman \u03c1=-0.185; descriptive only.",
"Held-out AUROC vs max-active JS specificity Spearman \u03c1=-0.148; descriptive only."
],
"interpretation": "Max-active interventions produced larger task-level target effects than norm-matched random controls with paired uncertainty excluding zero. Predictive SAE features therefore show causal specificity when intervened where the selected feature is actually represented, while the final-token baseline quantifies sensitivity to intervention location. Moving from the final prompt token to the feature's maximum-activation token increased intervention coverage from 28.6% to 82.1%, showing that causal conclusions depend materially on where the representation is tested.",
"metrics": {
"mean_selected_feature_test_auroc": 0.9622395833333334,
"mean_selected_feature_test_auroc_bootstrap_ci_95": [
0.9267066592261906,
0.994419642857143
],
"median_selected_feature_test_auroc": 0.9869791666666669,
"best_linear_probe_layer": 14,
"best_linear_probe_macro_auroc": 1.0,
"mean_paraphrase_topk_jaccard": 0.32550299223975043,
"mean_paraphrase_sparse_cosine": 0.9846351666472363,
"final_token_feature_coverage": 0.2857142857142857,
"prompt_anywhere_feature_coverage": 0.8214285714285714,
"max_active_feature_coverage": 0.8214285714285714,
"final_token_task_level": {
"sae_abs": 0.007574045897594507,
"random_abs": 0.004990028782880718,
"ratio": 1.5178361142081527,
"advantage": 0.002584017114713789,
"ci": [
-0.0005489836530094687,
0.006322265709085119
],
"pvalue": 0.171875,
"n_tasks": 28
},
"final_token_active_only": {
"sae_abs": 0.026509160641580775,
"random_abs": 0.017465100740082513,
"ratio": 1.5178361142081527,
"advantage": 0.009044059901498262,
"ci": [
-0.0016551230626646388,
0.020267208322184146
],
"pvalue": 0.171875,
"n_tasks": 8
},
"max_active_task_level": {
"sae_abs": 0.04140681956362508,
"random_abs": 0.01773720875956181,
"ratio": 2.3344608571122225,
"advantage": 0.023669610804063268,
"ci": [
0.0066882537599927475,
0.0468946453319534
],
"pvalue": 9.99950002499875e-05,
"n_tasks": 28
},
"max_active_active_only": {
"sae_abs": 0.05040830207745662,
"random_abs": 0.021593123707292633,
"ratio": 2.334460857112223,
"advantage": 0.028815178370163983,
"ci": [
0.008352378036563645,
0.05887514829816571
],
"pvalue": 0.00014999250037498125,
"n_tasks": 23
},
"feature_set_results": {
"1": {
"sae_abs": 0.0073844761188541,
"random_abs": 0.00605300132052173,
"ratio": 1.219969355337528,
"advantage": 0.0013314747983323686,
"ci": [
-0.0021684337141258357,
0.004932756118276826
],
"pvalue": 0.453125,
"n_tasks": 28
},
"3": {
"sae_abs": 0.01165702566504476,
"random_abs": 0.00956058465609591,
"ratio": 1.2192795821971139,
"advantage": 0.0020964410089488496,
"ci": [
-0.003360107559378132,
0.009788807602100323
],
"pvalue": 0.75146484375,
"n_tasks": 28
},
"5": {
"sae_abs": 0.010901511247668921,
"random_abs": 0.00997775314109664,
"ratio": 1.0925817760280598,
"advantage": 0.0009237581065722826,
"ci": [
-0.0035061794998390364,
0.006451443075535015
],
"pvalue": 0.7890625,
"n_tasks": 28
}
},
"study_summary": {
"n_concepts": 7,
"selected_feature_pooling": "prompt-wide max SAE activation across non-padding prompt tokens",
"dense_probe_pooling": "final prompt token residual",
"primary_causal_position_policy": "max_feature_activation",
"causal_statistical_unit": "causal task; ablation and amplification are averaged within task before paired inference",
"median_selected_feature_resample_support": 1.0,
"final_token_feature_coverage": 0.2857142857142857,
"max_active_feature_coverage": 0.8214285714285714,
"final_token_target_specificity_ratio": 1.5178361142081527,
"max_active_target_specificity_ratio": 2.3344608571122225,
"final_token_target_paired_advantage": 0.002584017114713789,
"max_active_target_paired_advantage": 0.023669610804063268,
"final_token_target_paired_ci_95": [
-0.0005554523851190289,
0.006223834152167133
],
"max_active_target_paired_ci_95": [
0.006155622166625111,
0.04787022493006328
],
"final_token_target_sign_flip_pvalue": 0.171875,
"max_active_target_sign_flip_pvalue": 9.99950002499875e-05,
"most_predictive_concept": {
"concept": "german_language",
"heldout_auroc": 1.0
},
"highest_max_active_target_specificity": {
"concept": "uncertainty",
"ratio": 5.5052321815575915
},
"highest_max_active_js_specificity": {
"concept": "uncertainty",
"ratio": 6.015626813896881
},
"correlations": {
"heldout_auroc_vs_max_active_target_specificity": {
"rho": -0.18531232916527532,
"pvalue": 0.6907777961916857,
"n": 7
},
"heldout_auroc_vs_max_active_js_specificity": {
"rho": -0.14824986333222023,
"pvalue": 0.7510797526579065,
"n": 7
},
"heldout_f1_vs_max_active_target_specificity": {
"rho": 0.21821789023599242,
"pvalue": 0.638298871640929,
"n": 7
},
"candidate_resample_support_vs_max_active_target_specificity": {
"rho": NaN,
"pvalue": NaN,
"n": 7
}
},
"guardrail": "Max-active causal positions are selected from SAE activation only, never from behavioral outcome. Cross-concept Spearman correlations are descriptive because the study contains seven concepts."
}
}
} |