Sign-Reset PI Optimizer / bench_report.json
Mechanism confirmed, baseline not beaten
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": 0.002,
10 "momentum": 0.9
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.0005,
16 "momentum": 0.0
17 },
18 "mean": 15.173156976699829
19 },
20 {
21 "cfg": {
22 "lr": 0.0005,
23 "momentum": 0.9
24 },
25 "mean": 7.033971428871155
26 },
27 {
28 "cfg": {
29 "lr": 0.001,
30 "momentum": 0.0
31 },
32 "mean": 12.429494142532349
33 },
34 {
35 "cfg": {
36 "lr": 0.001,
37 "momentum": 0.9
38 },
39 "mean": 6.920896649360657
40 },
41 {
42 "cfg": {
43 "lr": 0.0015,
44 "momentum": 0.0
45 },
46 "mean": 10.261504888534546
47 },
48 {
49 "cfg": {
50 "lr": 0.0015,
51 "momentum": 0.9
52 },
53 "mean": 6.831890940666199
54 },
55 {
56 "cfg": {
57 "lr": 0.002,
58 "momentum": 0.0
59 },
60 "mean": 8.732401132583618
61 },
62 {
63 "cfg": {
64 "lr": 0.002,
65 "momentum": 0.9
66 },
67 "mean": 6.79731810092926
68 },
69 {
70 "cfg": {
71 "lr": 0.003,
72 "momentum": 0.0
73 },
74 "mean": 7.4249961376190186
75 },
76 {
77 "cfg": {
78 "lr": 0.003,
79 "momentum": 0.9
80 },
81 "mean": 6.80940568447113
82 },
83 {
84 "cfg": {
85 "lr": 0.004,
86 "momentum": 0.0
87 },
88 "mean": 7.365786075592041
89 },
90 {
91 "cfg": {
92 "lr": 0.004,
93 "momentum": 0.9
94 },
95 "mean": 6.832672119140625
96 },
97 {
98 "cfg": {
99 "lr": 0.005,
100 "momentum": 0.0
101 },
102 "mean": 8.720722794532776
103 },
104 {
105 "cfg": {
106 "lr": 0.005,
107 "momentum": 0.9
108 },
109 "mean": 6.841493129730225
110 },
111 {
112 "cfg": {
113 "lr": 0.006,
114 "momentum": 0.0
115 },
116 "mean": 12.199845552444458
117 },
118 {
119 "cfg": {
120 "lr": 0.006,
121 "momentum": 0.9
122 },
123 "mean": 6.900767803192139
124 },
125 {
126 "cfg": {
127 "lr": 0.01,
128 "momentum": 0.0
129 },
130 "mean": 15.477824091911316
131 },
132 {
133 "cfg": {
134 "lr": 0.01,
135 "momentum": 0.9
136 },
137 "mean": 14.715186953544617
138 },
139 {
140 "cfg": {
141 "lr": 0.02,
142 "momentum": 0.0
143 },
144 "mean": 11.526971578598022
145 },
146 {
147 "cfg": {
148 "lr": 0.02,
149 "momentum": 0.9
150 },
151 "mean": 21.544694423675537
152 }
153 ],
154 "full": {
155 "mean": 6.9946147203445435,
156 "std": 0.5462041094541707,
157 "per_seed": [
158 7.059720993041992,
159 6.078768253326416,
160 6.975253105163574,
161 7.075530052185059,
162 6.368229866027832,
163 6.91272497177124,
164 7.652469158172607,
165 7.834221363067627
166 ],
167 "n": 8
168 }
169 },
170 "idea": {
171 "mean": 7.51734858751297,
172 "std": 0.6549113710902053,
173 "per_seed": [
174 7.288843154907227,
175 7.805213451385498,
176 7.1009440422058105,
177 7.081621170043945,
178 6.880852222442627,
179 7.159562110900879,
180 7.777229309082031,
181 9.044523239135742
182 ],
183 "n": 8,
184 "diagnostics": [
185 {
186 "metric": 7.288843154907227,
187 "resets": 89,
188 "mean_integral_norm": 15.755404067039489,
189 "final_integral_norm": 0.0
190 },
191 {
192 "metric": 7.805213451385498,
193 "resets": 86,
194 "mean_integral_norm": 15.125472933905465,
195 "final_integral_norm": 0.0
196 },
197 {
198 "metric": 7.1009440422058105,
199 "resets": 85,
200 "mean_integral_norm": 14.83641802413123,
201 "final_integral_norm": 0.0
202 },
203 {
204 "metric": 7.081621170043945,
205 "resets": 94,
206 "mean_integral_norm": 16.25910394702639,
207 "final_integral_norm": 0.0
208 },
209 {
210 "metric": 6.880852222442627,
211 "resets": 101,
212 "mean_integral_norm": 13.210206178256444,
213 "final_integral_norm": 10.246424674987793
214 },
215 {
216 "metric": 7.159562110900879,
217 "resets": 86,
218 "mean_integral_norm": 14.10479433962277,
219 "final_integral_norm": 0.0
220 },
221 {
222 "metric": 7.777229309082031,
223 "resets": 84,
224 "mean_integral_norm": 13.188805731705257,
225 "final_integral_norm": 0.0
226 },
227 {
228 "metric": 9.044523239135742,
229 "resets": 77,
230 "mean_integral_norm": 15.760829194954463,
231 "final_integral_norm": 16.458066940307617
232 }
233 ],
234 "cfg": {
235 "lr": 0.004,
236 "kp": 1.0,
237 "ki": 0.2
238 }
239 },
240 "comparison": {
241 "delta_mean": 0.5227338671684265,
242 "idea_wins": 0,
243 "n_pairs": 8,
244 "per_seed_diffs": [
245 0.22912216186523438,
246 1.726445198059082,
247 0.12569093704223633,
248 0.006091117858886719,
249 0.5126223564147949,
250 0.24683713912963867,
251 0.12476015090942383,
252 1.2103018760681152
253 ],
254 "p_value": 0.0081,
255 "mde": 0.51612198792684,
256 "mde_rel_pct": 7.378847993237528,
257 "verdict": "idea worse (significant)",
258 "system_worked": false
259 },
260 "mechanism_signature": {
261 "predicted_reset_effect": "global gradient sign reversal resets all integral tensors to zero",
262 "observed_reset_rate": 0.6267857142857143,
263 "observed_mean_integral_norm": 14.780129302080187,
264 "confirmed": true
265 }
266}