Online Taylor Residual World Model / bench_report.json
Beats tuned baseline
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": 0.006,
10 "wd": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "wd": 0.0
17 },
18 "mean": 0.41046395525336266
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "wd": 0.0
24 },
25 "mean": 0.05568788480013609
26 },
27 {
28 "cfg": {
29 "lr": 0.006,
30 "wd": 0.0
31 },
32 "mean": 0.04543822933919728
33 }
34 ],
35 "full": {
36 "mean": 0.04543822933919728,
37 "std": 0.014322681353099428,
38 "per_seed": [
39 0.04790250211954117,
40 0.028566891327500343,
41 0.057277221232652664,
42 0.07493516057729721,
43 0.03164242208003998,
44 0.03424108773469925,
45 0.04783138260245323,
46 0.04110916703939438
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.0008183727071746211,
53 "std": 0.0005857487704347297,
54 "per_seed": [
55 0.0005037843780949299,
56 0.0009486444474386546,
57 0.00047346246859643554,
58 0.0017819429777846984,
59 0.00027305720189630295,
60 0.001774419258314151,
61 0.00031528659531089673,
62 0.00047638432996090045
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": -0.044619856632022656,
68 "idea_wins": 8,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 -0.04739871774144624,
72 -0.02761824688006169,
73 -0.05680375876405623,
74 -0.07315321759951252,
75 -0.03136936487814367,
76 -0.032466668476385095,
77 -0.047516096007142336,
78 -0.040632782709433476
79 ],
80 "p_value": 0.00475,
81 "mde": 0.010499328941172536,
82 "mde_rel_pct": 23.106817967739982,
83 "verdict": "idea better (significant)",
84 "system_worked": true
85 },
86 "mechanism_signature": {
87 "trained_model_test_mse_before_rls": 0.04790250211954117,
88 "trained_model_test_mse_after_rls": 0.0005037843780949299,
89 "relative_correction": 0.9894831301956267,
90 "effective_memory_approx": 199.99999999999983,
91 "confirmed": true
92 },
93 "idea_sweep": [
94 {
95 "cfg": {
96 "lr": 0.006,
97 "wd": 0.0,
98 "lam": 0.9,
99 "p0": 10.0
100 },
101 "result": {
102 "mean": 0.0009322937952897222,
103 "std": 0.000574016816315803,
104 "per_seed": [
105 0.000917718187136525,
106 0.001109695949367615,
107 0.00047612724871288754,
108 0.001871022990995478,
109 0.000309179974464434,
110 0.0017833768326602597,
111 0.0003931757740019105,
112 0.0005980534049786682
113 ],
114 "n": 8
115 }
116 },
117 {
118 "cfg": {
119 "lr": 0.006,
120 "wd": 0.0,
121 "lam": 0.98,
122 "p0": 10.0
123 },
124 "result": {
125 "mean": 0.0008583902984778593,
126 "std": 0.0006199359208421757,
127 "per_seed": [
128 0.0005780594359315019,
129 0.0009298460194156407,
130 0.00048077141540815444,
131 0.0018578044183666753,
132 0.00028224137960249416,
133 0.0019118917096093264,
134 0.0003379446056156663,
135 0.0004885634038734143
136 ],
137 "n": 8
138 }
139 },
140 {
141 "cfg": {
142 "lr": 0.006,
143 "wd": 0.0,
144 "lam": 0.995,
145 "p0": 10.0
146 },
147 "result": {
148 "mean": 0.0008183727071746211,
149 "std": 0.0005857487704347297,
150 "per_seed": [
151 0.0005037843780949299,
152 0.0009486444474386546,
153 0.00047346246859643554,
154 0.0017819429777846984,
155 0.00027305720189630295,
156 0.001774419258314151,
157 0.00031528659531089673,
158 0.00047638432996090045
159 ],
160 "n": 8
161 }
162 }
163 ],
164 "runtime_sec": 22.165063619613647
165}