OT-Sufficient Bottleneck Flow Matching / results.json
Mechanism confirmed, baseline not beaten
1{
2 "device": "cuda",
3 "math_check": {
4 "sorted_ot_cost": 0.14749999344348907,
5 "random_coupling_cost": 2.112499952316284,
6 "derivative_max_error": 0.00067901611328125
7 },
8 "locality_sweep": [
9 {
10 "epsilon": 0.06,
11 "lambda": 0.0,
12 "paired_R": 0.7111561298370361,
13 "marginal_err": 4.0413811802864075e-05
14 },
15 {
16 "epsilon": 0.06,
17 "lambda": 0.5,
18 "paired_R": 0.0283959973603487,
19 "marginal_err": 1.2326985597610474e-05
20 },
21 {
22 "epsilon": 0.06,
23 "lambda": 2.0,
24 "paired_R": 0.005961012560874224,
25 "marginal_err": 6.226822733879089e-06
26 },
27 {
28 "epsilon": 0.06,
29 "lambda": 8.0,
30 "paired_R": 0.0008196450653485954,
31 "marginal_err": 2.3618340492248535e-06
32 },
33 {
34 "epsilon": 0.16,
35 "lambda": 0.0,
36 "paired_R": 0.8314868211746216,
37 "marginal_err": 1.30385160446167e-08
38 },
39 {
40 "epsilon": 0.16,
41 "lambda": 0.5,
42 "paired_R": 0.10292951762676239,
43 "marginal_err": 3.345310688018799e-06
44 },
45 {
46 "epsilon": 0.16,
47 "lambda": 2.0,
48 "paired_R": 0.02393825724720955,
49 "marginal_err": 7.588416337966919e-06
50 },
51 {
52 "epsilon": 0.16,
53 "lambda": 8.0,
54 "paired_R": 0.004799625836312771,
55 "marginal_err": 1.0259449481964111e-05
56 },
57 {
58 "epsilon": 0.4,
59 "lambda": 0.0,
60 "paired_R": 0.9091153144836426,
61 "marginal_err": 1.30385160446167e-08
62 },
63 {
64 "epsilon": 0.4,
65 "lambda": 0.5,
66 "paired_R": 0.2552563548088074,
67 "marginal_err": 9.313225746154785e-09
68 },
69 {
70 "epsilon": 0.4,
71 "lambda": 2.0,
72 "paired_R": 0.07225342839956284,
73 "marginal_err": 1.1175870895385742e-08
74 },
75 {
76 "epsilon": 0.4,
77 "lambda": 8.0,
78 "paired_R": 0.01732504926621914,
79 "marginal_err": 2.0343810319900513e-05
80 }
81 ],
82 "comparison_lambda0": {
83 "flow_w1": 0.4429539889097214,
84 "mse_w1": 1.006102442741394,
85 "flow_mode_coverage": 0.85107421875,
86 "per_s_w1": [
87 0.45495861768722534,
88 0.29935240745544434,
89 0.3903495669364929,
90 0.6271553635597229
91 ]
92 },
93 "comparison_lambda8": {
94 "flow_w1": 0.17159544862806797,
95 "mse_w1": 1.0032900869846344,
96 "flow_mode_coverage": 0.908203125,
97 "per_s_w1": [
98 0.09672722965478897,
99 0.20712798833847046,
100 0.19947239756584167,
101 0.18305417895317078
102 ]
103 },
104 "predictions": {
105 "lambda_effect_at_eps_0.16": "paired_R should decrease monotonically with lambda",
106 "epsilon_effect_at_lambda_8": "paired_R should increase with epsilon",
107 "flow_distributional_effect": "flow W1 should be lower than deterministic MSE W1 and mode coverage near 1"
108 }
109}