Adaptive Householder Gradient Subspaces / bench_report.json

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "tabular",
  4  "model": "mlp_tiny",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.006
 10    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 128.04924488067627
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 17.483699202537537
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 15.365184903144836
 29      }
 30    ],
 31    "full": {
 32      "mean": 15.365184903144836,
 33      "std": 2.415935250017397,
 34      "per_seed": [
 35        14.475372314453125,
 36        12.608717918395996,
 37        13.87478256225586,
 38        14.577168464660645,
 39        13.041308403015137,
 40        17.592876434326172,
 41        16.508373260498047,
 42        20.24287986755371
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 15.514657378196716,
 49    "std": 1.759501310100533,
 50    "per_seed": [
 51      14.62114143371582,
 52      13.303616523742676,
 53      15.129264831542969,
 54      15.524860382080078,
 55      13.418356895446777,
 56      15.95352840423584,
 57      17.32539176940918,
 58      18.84109878540039
 59    ],
 60    "n": 8,
 61    "stats_per_seed": [
 62      {
 63        "rank_mean": 3.0,
 64        "orth_mean": 2.0764549699434394e-07,
 65        "projection_ratio": 1.0
 66      },
 67      {
 68        "rank_mean": 3.0,
 69        "orth_mean": 2.2836561377717305e-07,
 70        "projection_ratio": 1.0
 71      },
 72      {
 73        "rank_mean": 3.0,
 74        "orth_mean": 2.202748265271263e-07,
 75        "projection_ratio": 1.0
 76      },
 77      {
 78        "rank_mean": 3.0,
 79        "orth_mean": 2.2073787706331257e-07,
 80        "projection_ratio": 1.0
 81      },
 82      {
 83        "rank_mean": 3.0,
 84        "orth_mean": 2.6148559885265223e-07,
 85        "projection_ratio": 1.0
 86      },
 87      {
 88        "rank_mean": 3.0,
 89        "orth_mean": 2.169296318873156e-07,
 90        "projection_ratio": 1.0
 91      },
 92      {
 93        "rank_mean": 3.0,
 94        "orth_mean": 2.209224932375125e-07,
 95        "projection_ratio": 1.0
 96      },
 97      {
 98        "rank_mean": 3.0,
 99        "orth_mean": 2.2133620220583023e-07,
100        "projection_ratio": 1.0
101      }
102    ]
103  },
104  "comparison": {
105    "delta_mean": 0.14947247505187988,
106    "idea_wins": 2,
107    "n_pairs": 8,
108    "per_seed_diffs": [
109      0.1457691192626953,
110      0.6948986053466797,
111      1.2544822692871094,
112      0.9476919174194336,
113      0.3770484924316406,
114      -1.639348030090332,
115      0.8170185089111328,
116      -1.4017810821533203
117    ],
118    "p_value": 0.7055,
119    "mde": 0.9082613543107244,
120    "mde_rel_pct": 5.911164493209762,
121    "verdict": "no measurable effect",
122    "system_worked": false
123  },
124  "mechanism_signature": {
125    "track_rationale": "optimizer modification matches the tabular optimizer track",
126    "idea_trials": [
127      {
128        "cfg": {
129          "lr": 0.001,
130          "rel_tol": 0.3,
131          "block": 4,
132          "kmax": 16
133        },
134        "mean": 128.22543716430664,
135        "std": 14.557069548048865
136      },
137      {
138        "cfg": {
139          "lr": 0.003,
140          "rel_tol": 0.3,
141          "block": 4,
142          "kmax": 16
143        },
144        "mean": 17.355298161506653,
145        "std": 1.848799316741261
146      },
147      {
148        "cfg": {
149          "lr": 0.006,
150          "rel_tol": 0.3,
151          "block": 4,
152          "kmax": 16
153        },
154        "mean": 15.514657378196716,
155        "std": 1.759501310100533
156      }
157    ],
158    "mechanism_signature": {
159      "prediction": "adaptive Householder subspaces retain fewer than kmax dimensions while maintaining orthogonality",
160      "observed_rank_mean": 3.0,
161      "observed_orthogonality_mean": 2.247122175681583e-07,
162      "observed_projected_gradient_norm_ratio": 1.0,
163      "confirmed": true
164    },
165    "runtime_sec": 14.900958776473999
166  },
167  "custom_track": null
168}