Nonreciprocal Brownian Optimizer / report_bench_2026-08-31T192247.md

Mechanism confirmed, baseline not beaten

Raw ⬇ ZIP

Стенд-проверка (stage-2) · промт оператора:

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 9, "verdict": "Implemented the nonreciprocal two-replica optimizer on the prescribed tabular optimizer track with the shared mlp_tiny architecture, paired seeds, baseline sweep, learning-rate parity, and 8-seed permutation testing. The trained replicas showed nonzero separation/circulation, confirming the mechanism signature, but the idea performed significantly worse: test MSE 6.8408 versus baseline 4.8192, paired delta +2.0217, permutation p=0.0081. Therefore it does not transfer as a practical win on this benchmark.", "metrics": { "baseline": "mlp_tiny tabular, Adam, best lr=0.006 and weight_decay=0.0001; full 8-seed test MSE mean 4.8191644549, std 0.1716614496", "idea": "two replicas with annealed k1=0.20, k2=0.05, best tested lr=0.003 and weight_decay=0.0001; full 8-seed test MSE mean 6.8408182263, std 0.1725772231", "paired_delta": "idea - baseline = +2.0216537714; 0/8 idea wins; permutation p=0.0081", "mechanism_signature": "observed_epoch_area_seed0=-1.9684885011e-08, mean replica separation=0.0770557180, confirmed=true" }, "bench_report": { "bench_version": 1, "track": "tabular", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006, "weight_decay": 0.0001 }, "sweep": [ { "cfg": { "lr": 0.0015, "weight_decay": 0.0 }, "mean": 6.582355260848999 }, { "cfg": { "lr": 0.0015, "weight_decay": 0.0001 }, "mean": 6.403406500816345 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 6.050842046737671 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0001 }, "mean": 5.60420548915863 }, { "cfg": { "lr": 0.006, "weight_decay": 0.0 }, "mean": 5.5363428592681885 }, { "cfg": { "lr": 0.006, "weight_decay": 0.0001 }, "mean": 4.698180794715881 } ], "full": { "mean": 4.819164454936981, "std": 0.17166144964543392, "per_seed": [ 4.596839904785156, 4.635165214538574, 4.646710395812988, 4.914007663726807, 5.076730251312256, 4.984879016876221, 4.754410743713379, 4.944572448730469 ], "n": 8 } }, "idea": { "mean": 6.840818226337433, "std": 0.17257722313102455, "per_seed": [ 6.538304328918457, 6.585258960723877, 6.972299098968506, 6.974996089935303, 6.823678970336914, 7.017618656158447, 6.853421211242676, 6.960968494415283 ], "n": 8 }, "comparison": { "delta_mean": 2.0216537714004517, "idea_wins": 0, "n_pairs": 8, "per_seed_diffs": [ 1.9414644241333008, 1.9500937461853027, 2.3255887031555176, 2.060988426208496, 1.7469487190246582, 2.0327396392822266, 2.099010467529297, 2.0163960456848145 ], "p_value": 0.0081, "mde": 0.13669161157539178, "mde_rel_pct": 2.8364172431459234, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "prediction": "nonzero replica circulation when k1!=k2", "k1": 0.2, "k2": 0.05, "observed_epoch_area_seed0": -1.9684885011277033e-08, "observed_mean_replica_separation": 0.07705571800470352, "confirmed": true } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py", "files": [ "bench_experiment.py", "bench_report.json" ], "limitations": "Only the prescribed tabular track and mlp_tiny model were tested; no vision, sequence, or dynamics track was run. The implementation used annealed deterministic coupling without explicit Brownian noise, and the mechanism signature used a simple epoch-level first-coordinate circulation proxy rather than a full stationary entropy-production estimate. The idea has approximately twice the parameter-update compute because it trains two replicas.", "system_verdict": "partial", "practical_verdict": "harms", "mechanism_ok": 1, "system_judged": true }