Fejer reflection accelerator for fixed-point layers / bench_report.json

Failed on benchmark

Raw ⬇ ZIP
  1{
  2  "bench_version": 1,
  3  "track": "dynamics",
  4  "model": "rnn_small",
  5  "metric_direction": "lower is better",
  6  "n_seeds": 8,
  7  "baseline": {
  8    "best_cfg": {
  9      "lr": 0.005,
 10      "K": 15
 11    },
 12    "sweep": [
 13      {
 14        "cfg": {
 15          "lr": 0.001,
 16          "K": 3
 17        },
 18        "mean": 0.016172005562111735
 19      },
 20      {
 21        "cfg": {
 22          "lr": 0.003,
 23          "K": 3
 24        },
 25        "mean": 0.003774263139348477
 26      },
 27      {
 28        "cfg": {
 29          "lr": 0.005,
 30          "K": 3
 31        },
 32        "mean": 0.0017951192276086658
 33      },
 34      {
 35        "cfg": {
 36          "lr": 0.001,
 37          "K": 7
 38        },
 39        "mean": 0.015029519563540816
 40      },
 41      {
 42        "cfg": {
 43          "lr": 0.003,
 44          "K": 7
 45        },
 46        "mean": 0.002934314659796655
 47      },
 48      {
 49        "cfg": {
 50          "lr": 0.005,
 51          "K": 7
 52        },
 53        "mean": 0.0014960496919229627
 54      },
 55      {
 56        "cfg": {
 57          "lr": 0.001,
 58          "K": 15
 59        },
 60        "mean": 0.014895078958943486
 61      },
 62      {
 63        "cfg": {
 64          "lr": 0.003,
 65          "K": 15
 66        },
 67        "mean": 0.0028665231075137854
 68      },
 69      {
 70        "cfg": {
 71          "lr": 0.005,
 72          "K": 15
 73        },
 74        "mean": 0.0013684573350474238
 75      }
 76    ],
 77    "full": {
 78      "mean": 0.0011932104389416054,
 79      "std": 0.000203301503237927,
 80      "per_seed": [
 81        0.0015044162282720208,
 82        0.0011529525509104133,
 83        0.0014893548795953393,
 84        0.001327105681411922,
 85        0.001068409881554544,
 86        0.0009947591461241245,
 87        0.0010155801428481936,
 88        0.0009931050008162856
 89      ],
 90      "n": 8
 91    }
 92  },
 93  "idea": {
 94    "mean": 0.0023188923223642632,
 95    "std": 0.0018174311282403578,
 96    "per_seed": [
 97      0.006926196161657572,
 98      0.0018025230383500457,
 99      0.0029087376315146685,
100      0.0010936673497781157,
101      0.0015447858022525907,
102      0.0014355253661051393,
103      0.0016264201840385795,
104      0.0012132830452173948
105    ],
106    "n": 8
107  },
108  "comparison": {
109    "delta_mean": 0.0011256818834226578,
110    "idea_wins": 1,
111    "n_pairs": 8,
112    "per_seed_diffs": [
113      0.005421779933385551,
114      0.0006495704874396324,
115      0.0014193827519193292,
116      -0.00023343833163380623,
117      0.0004763759206980467,
118      0.00044076621998101473,
119      0.0006108400411903858,
120      0.00022017804440110922
121    ],
122    "p_value": 0.02285,
123    "mde": 0.0015021645326744831,
124    "mde_rel_pct": 125.89267438918188,
125    "verdict": "idea worse (significant)",
126    "system_worked": false
127  },
128  "mechanism_signature": {
129    "structural_match": "controlled pendulum rollout with recurrent fixed-point state updates",
130    "shared_architecture": "same inp/state/head parameters and K evaluations; only baseline vs Fejer update differs",
131    "idea_grid": [
132      {
133        "cfg": {
134          "lr": 0.001,
135          "K": 3
136        },
137        "mean": 0.013947936473414302,
138        "per_seed": [
139          0.02009131945669651,
140          0.008569814264774323,
141          0.04837782308459282,
142          0.007685891818255186,
143          0.005474410951137543,
144          0.006128633860498667,
145          0.0073469290509819984,
146          0.007908669300377369
147        ]
148      },
149      {
150        "cfg": {
151          "lr": 0.003,
152          "K": 3
153        },
154        "mean": 0.004136359391850419,
155        "per_seed": [
156          0.002831853926181793,
157          0.0031103442888706923,
158          0.016510576009750366,
159          0.0014188818167895079,
160          0.0029145670123398304,
161          0.0021216198801994324,
162          0.0022474690340459347,
163          0.0019355631666257977
164        ]
165      },
166      {
167        "cfg": {
168          "lr": 0.005,
169          "K": 3
170        },
171        "mean": 0.0023188923223642632,
172        "per_seed": [
173          0.006926196161657572,
174          0.0018025230383500457,
175          0.0029087376315146685,
176          0.0010936673497781157,
177          0.0015447858022525907,
178          0.0014355253661051393,
179          0.0016264201840385795,
180          0.0012132830452173948
181        ]
182      },
183      {
184        "cfg": {
185          "lr": 0.001,
186          "K": 7
187        },
188        "mean": 0.06781748728826642,
189        "per_seed": [
190          0.09825821965932846,
191          0.060682933777570724,
192          0.1300727277994156,
193          0.10404244810342789,
194          0.02876441925764084,
195          0.03181176632642746,
196          0.04144725203514099,
197          0.04746013134717941
198        ]
199      },
200      {
201        "cfg": {
202          "lr": 0.003,
203          "K": 7
204        },
205        "mean": 0.022045168676413596,
206        "per_seed": [
207          0.03464265167713165,
208          0.02418467029929161,
209          0.060446545481681824,
210          0.0186628308147192,
211          0.011261790990829468,
212          0.007364522200077772,
213          0.007728805299848318,
214          0.01206953264772892
215        ]
216      },
217      {
218        "cfg": {
219          "lr": 0.005,
220          "K": 7
221        },
222        "mean": 0.028678134316578507,
223        "per_seed": [
224          0.027619337663054466,
225          0.009325508028268814,
226          0.04537428915500641,
227          0.10992163419723511,
228          0.011131452396512032,
229          0.00906284898519516,
230          0.009421450085937977,
231          0.007568554021418095
232        ]
233      },
234      {
235        "cfg": {
236          "lr": 0.001,
237          "K": 15
238        },
239        "mean": 0.45379000529646873,
240        "per_seed": [
241          0.5679315328598022,
242          0.6005325317382812,
243          0.31230661273002625,
244          0.5303172469139099,
245          0.5320272445678711,
246          0.35211363434791565,
247          0.2027447521686554,
248          0.5323464870452881
249        ]
250      },
251      {
252        "cfg": {
253          "lr": 0.003,
254          "K": 15
255        },
256        "mean": 0.31799950916320086,
257        "per_seed": [
258          0.25116634368896484,
259          0.3567662239074707,
260          0.34843429923057556,
261          0.7776817083358765,
262          0.16537804901599884,
263          0.31844574213027954,
264          0.09809694439172745,
265          0.22802676260471344
266        ]
267      },
268      {
269        "cfg": {
270          "lr": 0.005,
271          "K": 15
272        },
273        "mean": 0.2506284995470196,
274        "per_seed": [
275          0.42000678181648254,
276          0.17470629513263702,
277          0.2908547818660736,
278          0.1604318916797638,
279          0.34779635071754456,
280          0.018573837354779243,
281          0.14079605042934418,
282          0.45186200737953186
283        ]
284      }
285    ],
286    "mechanism_signature": {
287      "prediction": "Fejer residual contracts approximately as 1/(K+1) under resolvent/nonexpansive assumptions",
288      "K": 7,
289      "predicted_ratio": 0.125,
290      "observed_trained_ratio": 1.12704236805439,
291      "identity_error_on_trained_model": 13.323629379272461,
292      "confirmed": false
293    }
294  }
295}