Spectral subspace initialization for nonlinear teachers / report_bench_2026-09-01T103143.md
Mechanism confirmed, baseline not beaten
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Implemented bounded supervised spectral-subspace initialization for the first Linear layer of the matched tabular mlp_tiny benchmark using the canonical train_model path. The trained-model mechanism signature was confirmed: projection energy along the spectral direction was 1.797x baseline. Nevertheless, tuned spectral initialization significantly worsened test MSE, so the idea does not give a benchmark win.", "metrics": { "baseline": "Friedman#1 tabular mlp_tiny, tuned lr=0.01, 20 epochs: test MSE 7.9319 +/- 0.8101 over 8 paired seeds.", "idea": "Spectral initialization, same lr grid and budget, best lr=0.01: test MSE 8.8684 +/- 0.7219 over 8 paired seeds; paired delta +0.9365, 0/8 wins, permutation p=0.0081.", "mechanism_signature": "Confirmed; mean trained-model first-layer projection-energy ratio idea/baseline=1.797, with spectral eigengaps 2.06-2.59." }, "bench_report": { "bench_version": 1, "track": "tabular", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "weight_decay": 0.0 }, "sweep": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0 }, "mean": 17.7983 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 13.1535 }, { "cfg": { "lr": 0.01, "weight_decay": 0.0 }, "mean": 7.7542 } ], "full": { "mean": 7.931941092014313, "std": 0.8101161551333251, "per_seed": [ 7.2100510597229, 8.814193725585938, 7.362153053283691, 7.630332946777344, 6.520580291748047, 8.472823143005371, 8.71557903289795, 8.729815483093262 ], "n": 8 } }, "idea": { "best_cfg": { "lr": 0.01, "weight_decay": 0.0 }, "sweep": [ { "cfg": { "lr": 0.001, "weight_decay": 0.0 }, "mean": 44.673 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 12.8833 }, { "cfg": { "lr": 0.01, "weight_decay": 0.0 }, "mean": 8.7758 } ], "full": { "mean": 8.868400931358337, "std": 0.7219192817226784, "per_seed": [ 9.187010765075684, 8.92486572265625, 8.406270980834961, 8.585136413574219, 7.358145713806152, 9.900330543518066, 9.11246395111084, 9.472983360290527 ], "n": 8 } }, "comparison": { "delta_mean": 0.9364598393440247, "idea_wins": 0, "n_pairs": 8, "per_seed_diffs": [ 1.9769597053527832, 0.1106719970703125, 1.0441179275512695, 0.954803466796875, 0.8375654220581055, 1.4275074005126953, 0.3968849182128906, 0.7431678771972656 ], "p_value": 0.0081, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "confirmed": true, "observed_energy_ratio": 1.7969543217617925, "threshold": 1.2 } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_spectral.py", "files": [ "bench_spectral.py", "bench_report.json" ], "limitations": "Only the designated tabular Friedman#1/mlp_tiny track was tested; matrix-valued lifted responses, multi-dimensional subspaces r>1, image patch embeddings, and larger data regimes were not tested.", "system_verdict": "partial", "practical_verdict": "harms", "mechanism_ok": 1, "system_judged": true }