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."
    }
  }
}