Exact doubly stochastic low-rank attention / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "sequence",
4 "model": "transformer_tiny",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.001,
10 "rank": 8
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "rank": 8
17 },
18 "mean": 0.4039061442017555
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "rank": 8
24 },
25 "mean": 0.5307872667908669
26 },
27 {
28 "cfg": {
29 "lr": 0.006,
30 "rank": 8
31 },
32 "mean": 0.8058461993932724
33 }
34 ],
35 "full": {
36 "mean": 0.4480689875781536,
37 "std": 0.08669399607049459,
38 "per_seed": [
39 0.4437996447086334,
40 0.3383534550666809,
41 0.37107908725738525,
42 0.4623923897743225,
43 0.6176453828811646,
44 0.3899054229259491,
45 0.5426207780838013,
46 0.41875573992729187
47 ],
48 "n": 8
49 },
50 "selected_cfg_full": {
51 "mean": 0.4480689875781536,
52 "std": 0.08669399607049459,
53 "per_seed": [
54 0.4437996447086334,
55 0.3383534550666809,
56 0.37107908725738525,
57 0.4623923897743225,
58 0.6176453828811646,
59 0.3899054229259491,
60 0.5426207780838013,
61 0.41875573992729187
62 ],
63 "n": 8
64 }
65 },
66 "idea": {
67 "mean": 0.4980953820049763,
68 "std": 0.0717662616842552,
69 "per_seed": [
70 0.5194478631019592,
71 0.45840978622436523,
72 0.43587931990623474,
73 0.6198630928993225,
74 0.5728811025619507,
75 0.39179709553718567,
76 0.535282552242279,
77 0.45120224356651306
78 ],
79 "n": 8
80 },
81 "comparison": {
82 "delta_mean": 0.05002639442682266,
83 "idea_wins": 2,
84 "n_pairs": 8,
85 "per_seed_diffs": [
86 0.0756482183933258,
87 0.12005633115768433,
88 0.06480023264884949,
89 0.157470703125,
90 -0.04476428031921387,
91 0.0018916726112365723,
92 -0.007338225841522217,
93 0.03244650363922119
94 ],
95 "p_value": 0.08665,
96 "mde": 0.05673288695634186,
97 "mde_rel_pct": 12.661641070717113,
98 "verdict": "no significant win",
99 "system_worked": false
100 },
101 "mechanism_signature": {
102 "prediction": "trained low-rank attention remains row-stochastic and shared-marginal, with O(nr) state",
103 "observed": {
104 "row_u": 2.086162567138672e-07,
105 "row_v": 2.2351741790771484e-07,
106 "shared_col": 2.5033950805664062e-06,
107 "factor_entries": 520,
108 "dense_entries": 1024,
109 "predicted_storage_ratio": 1.9692307692307693,
110 "observed_storage_ratio": 1.9692307692307693,
111 "confirmed": true
112 },
113 "confirmed": true
114 },
115 "protocol_notes": {
116 "matched_track": "sequence multi-token forecast",
117 "shared_architecture": "same input/position/FFN/head; only attention operator differs",
118 "n_train": 400,
119 "epochs": 8,
120 "idea_grid": [
121 [
122 {
123 "mean": 0.4980953820049763,
124 "std": 0.0717662616842552,
125 "per_seed": [
126 0.5194478631019592,
127 0.45840978622436523,
128 0.43587931990623474,
129 0.6198630928993225,
130 0.5728811025619507,
131 0.39179709553718567,
132 0.535282552242279,
133 0.45120224356651306
134 ],
135 "n": 8
136 },
137 {
138 "lr": 0.001,
139 "rank": 8
140 }
141 ],
142 [
143 {
144 "mean": 0.6111128553748131,
145 "std": 0.07747015035751928,
146 "per_seed": [
147 0.6425154209136963,
148 0.5512837767601013,
149 0.5704087018966675,
150 0.7150149941444397,
151 0.646016538143158,
152 0.5375900864601135,
153 0.7234078645706177,
154 0.5026654601097107
155 ],
156 "n": 8
157 },
158 {
159 "lr": 0.003,
160 "rank": 8
161 }
162 ],
163 [
164 {
165 "mean": 0.9795129522681236,
166 "std": 0.22563627392238808,
167 "per_seed": [
168 0.9576879739761353,
169 0.7885317802429199,
170 1.4408782720565796,
171 1.1196062564849854,
172 0.7925991415977478,
173 0.8517841100692749,
174 1.1472877264022827,
175 0.7377283573150635
176 ],
177 "n": 8
178 },
179 {
180 "lr": 0.006,
181 "rank": 8
182 }
183 ]
184 ]
185 }
186}