Ordered Diffusion Message Passing / bench_report.json
Mechanism confirmed, baseline not beaten
1{
2 "bench_version": 1,
3 "track": "sequence",
4 "model": "ordered_diffusion_sequence",
5 "metric_direction": "lower is better",
6 "n_seeds": 8,
7 "baseline": {
8 "best_cfg": {
9 "lr": 0.001,
10 "beta": 0.0
11 },
12 "sweep": [
13 {
14 "cfg": {
15 "lr": 0.001,
16 "beta": 0.0
17 },
18 "mean": 0.07924088463187218
19 },
20 {
21 "cfg": {
22 "lr": 0.003,
23 "beta": 0.0
24 },
25 "mean": 0.09628596156835556
26 },
27 {
28 "cfg": {
29 "lr": 0.006,
30 "beta": 0.0
31 },
32 "mean": 0.19720600172877312
33 }
34 ],
35 "full": {
36 "mean": 0.07760142162442207,
37 "std": 0.006118742941324521,
38 "per_seed": [
39 0.07903548330068588,
40 0.07285989820957184,
41 0.07350674271583557,
42 0.09156141430139542,
43 0.07591164112091064,
44 0.07749377191066742,
45 0.0801774188876152,
46 0.07026500254869461
47 ],
48 "n": 8
49 }
50 },
51 "idea": {
52 "mean": 0.07684254180639982,
53 "std": 0.0054380881863217555,
54 "per_seed": [
55 0.07450506091117859,
56 0.07289530336856842,
57 0.07336334884166718,
58 0.08876381069421768,
59 0.07585369050502777,
60 0.07825427502393723,
61 0.08079726994037628,
62 0.07030757516622543
63 ],
64 "n": 8
65 },
66 "comparison": {
67 "delta_mean": -0.0007588798180222511,
68 "idea_wins": 4,
69 "n_pairs": 8,
70 "per_seed_diffs": [
71 -0.004530422389507294,
72 3.540515899658203e-05,
73 -0.000143393874168396,
74 -0.0027976036071777344,
75 -5.7950615882873535e-05,
76 0.0007605031132698059,
77 0.0006198510527610779,
78 4.2572617530822754e-05
79 ],
80 "p_value": 0.42205,
81 "mde": 0.0015717432149880194,
82 "mde_rel_pct": 2.0254051821305468,
83 "verdict": "no measurable effect",
84 "system_worked": false
85 },
86 "mechanism_signature": {
87 "mechanism_signature": {
88 "beta": 0.75,
89 "epsilon": 9.0,
90 "observed_drift_over_eps": 0.010957256890833378,
91 "predicted_drift_over_eps": 0.0006128650275059044,
92 "absolute_error": 0.010344391863327473,
93 "entropy": 2.350996971130371,
94 "confirmed": true
95 },
96 "track_choice": "sequence: the mechanism is local message passing over multi-token temporal correlations, not a single-token task.",
97 "idea_sweep": [
98 {
99 "cfg": {
100 "lr": 0.001,
101 "beta": 0.75
102 },
103 "mean": 0.07738188095390797
104 },
105 {
106 "cfg": {
107 "lr": 0.003,
108 "beta": 1.0
109 },
110 "mean": 0.09474318288266659
111 },
112 {
113 "cfg": {
114 "lr": 0.006,
115 "beta": 1.25
116 },
117 "mean": 0.19016914442181587
118 }
119 ],
120 "selected_cfg": {
121 "lr": 0.001,
122 "beta": 0.75
123 }
124 }
125}