Adaptive Proximal Quasi-Newton Training / report_bench_2026-09-01T162004.md

Failed on benchmark

Raw ⬇ ZIP

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

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 10, "verdict": "Implemented and ran the full stage-2 tabular benchmark using the matched Friedman#1 regression task and shared mlp_tiny architecture. The adaptive proximal quasi-Newton system was significantly worse than the AdamW baseline: test MSE 7.772281 vs 6.920556, paired delta +0.851725, permutation p=0.0081. The trained-model mechanism signature was not confirmed because observed adaptive stepsizes were far above the secant-derived stability estimate.", "metrics": { "baseline": "AdamW sweep, best lr=0.01 and weight_decay=0.0; 8-seed test MSE mean=6.920556, std=0.571610", "idea": "Adaptive diagonal proximal quasi-Newton, best lr=0.003, lambda=1e-05, eta_up=1.5; 8-seed test MSE mean=7.772281, std=0.664117", "paired_delta_mean": 0.8517247438430786, "permutation_p_value": 0.0081, "idea_wins": 0, "n_pairs": 8 }, "bench_report": { "bench_version": 1, "track": "tabular", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.01, "wd": 0.0 }, "full": { "mean": 6.9205562472343445, "std": 0.5716103719461259, "per_seed": [ 7.096155643463135, 6.974558353424072, 6.666646480560303, 6.943192005157471, 5.7007622718811035, 7.199983596801758, 6.8888444900512695, 7.8943071365356445 ], "n": 8 } }, "idea": { "mean": 7.772280991077423, "std": 0.6641172020856825, "chosen_cfg": { "lr": 0.003, "lam": 1e-05, "eta_up": 1.5 } }, "comparison": { "delta_mean": 0.8517247438430786, "idea_wins": 0, "n_pairs": 8, "p_value": 0.0081, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "prediction": "adaptive eta should remain below local secant stability boundary 2/kappa", "predicted_eta_c_from_trained_secants": 6.806255525998862e-09, "observed_max_final_eta": 0.00043937891010204985, "observed_median_eta": 0.0004584609905860991, "accepted_mean": 124.25, "rejected_mean": 25.75, "confirmed": false } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_stage2.py", "files": [ "bench_stage2.py", "bench_report.json", "bench_stdout.txt", "bench_stderr.txt" ], "limitations": "Only the structurally matched tabular track was tested; vision, sequence, and dynamics were not tested. The implementation uses diagonal secant scaling rather than full limited-memory BFGS, and wall-clock/FLOP speedup was not measured.", "system_verdict": "failed", "practical_verdict": "harms", "mechanism_ok": 0, "system_judged": true }