File size: 5,120 Bytes
ad89351
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
{
  "description": "Controlled scaling experiment: 8M vs 46.5M parameter Diffusion-LM riddle solvers",
  "dataset": {
    "name": "Riddles \u2014 A Synthetic Riddle Dataset for NLP",
    "source": "Kaggle",
    "url": "https://www.kaggle.com/datasets/prajwaldongre/riddles-a-synthetic-riddle-dataset-for-nlp",
    "license": "CC0",
    "train_size": 232,
    "heldout_size": 66,
    "unique_total": 298,
    "duplicates_removed": 702,
    "split_method": "SHA-256 of normalized riddle text",
    "answer_max_len": 4,
    "vocab_source": "training set only"
  },
  "checkpoints": {
    "phase3": {
      "label": "Phase 3 Reconstructed",
      "architecture": {
        "params": null,
        "T": 200,
        "d_model": 256,
        "n_layers": 4
      },
      "training": {
        "steps": 500,
        "lr": 0.0003,
        "schedule": "constant",
        "batch": 64,
        "seed": 42,
        "T": 200
      },
      "eval_heldout": {
        "exact_match_K1_pct": 46.96969696969697,
        "exact_match_K10_pct": 46.96969696969697,
        "token_f1_K1": 0.7124098124098122,
        "n": 66
      },
      "eval_train": {
        "exact_match_K1_pct": 87.5,
        "exact_match_K10_pct": 86.45833333333333,
        "token_f1_K1": 0.9597222222222221,
        "n": 192
      },
      "decoding_modes": {
        "A_full_chain_em_pct": 0.5,
        "B_pure_noise_1shot_em_pct": 0.494949494949495,
        "C_gold_corrupted_1shot_em_pct": 0.5303030303030303
      },
      "signal_degradation": {
        "raw_xt_tok_acc_at_t2_pct": 100.0,
        "model_tok_acc_at_t2_pct": 76.6,
        "degradation_ppt": 23.4
      },
      "state_drift": {
        "t2_oracle_tok_acc_pct": 75.9,
        "t2_free_tok_acc_pct": 72.0,
        "drift_ppt": 3.91,
        "oracle_flat_across_noise": false
      }
    },
    "phase4_original": {
      "label": "Phase 4 Original",
      "architecture": {
        "params": null,
        "T": 500,
        "d_model": 512,
        "n_layers": 6
      },
      "training": {
        "steps": 20000,
        "lr": 0.0003,
        "schedule": "constant",
        "batch": 64,
        "seed": 42,
        "T": 500
      },
      "eval_heldout": {
        "exact_match_K1_pct": 15.151515151515152,
        "exact_match_K10_pct": 15.151515151515152,
        "token_f1_K1": 0.5692279942279943,
        "n": 66
      },
      "eval_train": {
        "exact_match_K1_pct": 51.5625,
        "exact_match_K10_pct": 50.520833333333336,
        "token_f1_K1": 0.9034226190476198,
        "n": 192
      },
      "decoding_modes": {
        "A_full_chain_em_pct": 0.15656565656565657,
        "B_pure_noise_1shot_em_pct": 0.11616161616161616,
        "C_gold_corrupted_1shot_em_pct": 0.13636363636363635
      },
      "signal_degradation": {
        "raw_xt_tok_acc_at_t2_pct": 100.0,
        "model_tok_acc_at_t2_pct": 68.3,
        "degradation_ppt": 31.7
      },
      "state_drift": {
        "t2_oracle_tok_acc_pct": 68.3,
        "t2_free_tok_acc_pct": 68.3,
        "drift_ppt": 0.0,
        "oracle_flat_across_noise": true
      }
    },
    "phase4_scheduled": {
      "label": "Phase 4 Scheduled",
      "architecture": {
        "params": null,
        "T": 500,
        "d_model": 512,
        "n_layers": 6
      },
      "training": {
        "steps": 20000,
        "lr": 0.0001,
        "schedule": "warmup+cosine",
        "batch": 64,
        "seed": 42,
        "T": 500
      },
      "eval_heldout": {
        "exact_match_K1_pct": 13.636363636363637,
        "exact_match_K10_pct": 13.636363636363637,
        "token_f1_K1": 0.6034992784992784,
        "n": 66
      },
      "eval_train": {
        "exact_match_K1_pct": 40.625,
        "exact_match_K10_pct": 39.583333333333336,
        "token_f1_K1": 0.8827380952380958,
        "n": 192
      },
      "decoding_modes": {
        "A_full_chain_em_pct": 0.13636363636363635,
        "B_pure_noise_1shot_em_pct": 0.19696969696969696,
        "C_gold_corrupted_1shot_em_pct": 0.21717171717171718
      },
      "signal_degradation": {
        "raw_xt_tok_acc_at_t2_pct": 100.0,
        "model_tok_acc_at_t2_pct": 76.6,
        "degradation_ppt": 23.4
      },
      "state_drift": {
        "t2_oracle_tok_acc_pct": 75.1,
        "t2_free_tok_acc_pct": 74.6,
        "drift_ppt": 0.51,
        "oracle_flat_across_noise": false
      }
    }
  },
  "key_findings": [
    "Diffusion corruption is NOT the primary bottleneck at near-clean inputs. Raw-x_t achieves 100% token accuracy at t=2; the learned denoiser degrades this to 68-77%.",
    "Scaling from 8M to 46.5M parameters reduces held-out exact match from 47% to 15%, despite 40x more training steps.",
    "Phase 4 Original has 0.0% state drift (oracle = free-running) and is timestep-agnostic (flat 68.3% accuracy at every noise level).",
    "Phase 4 Scheduled has ~0.5 ppt drift and one-shot decoding (B/C = 20-22%) outperforms full chain (A = 14%).",
    "K=10 best-of-K never improves EM: only 1.08-1.26 unique candidates out of 10 (pairwise Jaccard > 0.97).",
    "The failure is denoiser/model-behavior-limited, not diffusion-corruption-limited."
  ]
}