Tiny Local Recurrence with Adaptive Computation / bench_report.json

Unverified

Raw ⬇ ZIP
  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    },
 11    "sweep": [
 12      {
 13        "cfg": {
 14          "lr": 0.001
 15        },
 16        "mean": 0.0002686133266252
 17      },
 18      {
 19        "cfg": {
 20          "lr": 0.003
 21        },
 22        "mean": 2.2429956516134553e-05
 23      },
 24      {
 25        "cfg": {
 26          "lr": 0.006
 27        },
 28        "mean": 1.2719097412627889e-05
 29      }
 30    ],
 31    "full": {
 32      "mean": 1.3684802866009704e-05,
 33      "std": 2.4297331089709535e-06,
 34      "per_seed": [
 35        1.0965275578200817e-05,
 36        1.3063688129477669e-05,
 37        1.363205592497252e-05,
 38        1.321537001786055e-05,
 39        1.694267484708689e-05,
 40        1.790331407391932e-05,
 41        1.3341849808057304e-05,
 42        1.0414194548502564e-05
 43      ],
 44      "n": 8
 45    }
 46  },
 47  "idea": {
 48    "mean": 1.1164625242088277e-06,
 49    "std": 6.736045713192092e-07,
 50    "per_seed": [
 51      2.8550934416671225e-07,
 52      2.1125936200405704e-06,
 53      1.6401930906795315e-06,
 54      3.9009708530102216e-07,
 55      1.0302179589416482e-06,
 56      1.806109253266186e-06,
 57      1.334532043983927e-06,
 58      3.3244779729102447e-07
 59    ],
 60    "n": 8,
 61    "chosen_cfg": {
 62      "lr": 0.006,
 63      "tmax": 8,
 64      "alpha": 0.25,
 65      "halt_bias": 0.0
 66    },
 67    "sweep": [
 68      {
 69        "cfg": {
 70          "lr": 0.006,
 71          "tmax": 8,
 72          "alpha": 0.25,
 73          "halt_bias": 0.0
 74        },
 75        "mean": 1.107098285046959e-06,
 76        "per_seed": [
 77          2.8550934416671225e-07,
 78          2.1125936200405704e-06,
 79          1.6401930906795315e-06,
 80          3.9009708530102216e-07
 81        ]
 82      },
 83      {
 84        "cfg": {
 85          "lr": 0.001,
 86          "tmax": 8,
 87          "alpha": 0.25,
 88          "halt_bias": 0.0
 89        },
 90        "mean": 5.165409334040305e-05,
 91        "per_seed": [
 92          1.0778670002764557e-05,
 93          8.488429011777043e-05,
 94          8.795395115157589e-05,
 95          2.299946208950132e-05
 96        ]
 97      },
 98      {
 99        "cfg": {
100          "lr": 0.006,
101          "tmax": 8,
102          "alpha": 0.25,
103          "halt_bias": 0.0
104        },
105        "mean": 1.107098285046959e-06,
106        "per_seed": [
107          2.8550934416671225e-07,
108          2.1125936200405704e-06,
109          1.6401930906795315e-06,
110          3.9009708530102216e-07
111        ]
112      }
113    ]
114  },
115  "comparison": {
116    "delta_mean": -1.2568340341800877e-05,
117    "idea_wins": 8,
118    "n_pairs": 8,
119    "per_seed_diffs": [
120      -1.0679766234034105e-05,
121      -1.0951094509437098e-05,
122      -1.1991862834292988e-05,
123      -1.2825272932559528e-05,
124      -1.5912456888145243e-05,
125      -1.6097204820653133e-05,
126      -1.2007317764073377e-05,
127      -1.008174675121154e-05
128    ],
129    "p_value": 0.0081,
130    "mde": 1.9149415921520303e-06,
131    "mde_rel_pct": 13.993198228001951,
132    "verdict": "idea better (significant)",
133    "system_worked": true
134  },
135  "mechanism_signature": {
136    "prediction": "adaptive recurrence reduces average executed microsteps as halting bias increases, while sharing one transition rule",
137    "predicted": {
138      "higher_bias": "lower_steps"
139    },
140    "observed": {
141      "idea_mean_weighted_steps": 1.287009134888649,
142      "idea_mean_halt_mass": 1.0,
143      "all_steps": [
144        1.235519289970398,
145        1.3300005197525024,
146        1.2526575326919556,
147        1.3078808784484863,
148        1.3313579559326172,
149        1.2955666780471802,
150        1.2435346841812134,
151        1.2995555400848389
152      ]
153    },
154    "confirmed": true,
155    "parameter_counts": {
156      "baseline": 13313,
157      "idea": 17090
158    }
159  },
160  "custom_track": null
161}