State-Dependent Metric Projected Optimizer / report_bench_2026-08-31T095654.md

✓✓ Beats tuned baseline

Raw ⬇ ZIP

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

(универсальный)

Ответ агента:

{ "worked": true, "confidence": 8, "verdict": "The registered tabular Friedman regression benchmark used the same mlp_tiny architecture and matched learning-rate union for tuned Adam baseline versus the state-dependent metric projected optimizer. The idea improved mean test MSE from 17.666198 to 15.638604 across all 8 paired seeds, with delta_mean=-2.027594 and permutation p_value=0.0081, giving the bench verdict idea better (significant).", "metrics": { "baseline": "Tuned Adam mean test MSE 17.666198.", "idea": "State-dependent metric projection mean test MSE 15.638604; 8/8 paired wins; zero constraint violations; mechanism correlation 0.807046 and confirmed=true." }, "bench_report": { "bench_version": 1, "track": "tabular", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.003, "beta2": 0.999 }, "sweep": [ { "cfg": { "lr": 0.0003, "beta2": 0.99 }, "mean": 219.74071502685547 }, { "cfg": { "lr": 0.001, "beta2": 0.99 }, "mean": 133.59764099121094 }, { "cfg": { "lr": 0.003, "beta2": 0.999 }, "mean": 16.845557928085327 } ], "full": { "mean": 17.666198015213013, "std": 2.17622036048186, "per_seed": [ 17.691715240478516, 14.373583793640137, 20.163808822631836, 15.15312385559082, 18.178590774536133, 15.800971031188965, 19.925329208374023, 20.042461395263672 ], "n": 8 } }, "idea": { "per_seed": [ 15.087238311767578, 12.6608247756958, 16.30777355957031, 14.145133972167969, 15.023676872253418, 15.424975395202637, 17.59527587890625, 18.863933563232422 ], "mean": 15.638604044914246, "best_cfg": { "lr": 0.003, "beta2": 0.999, "metric_beta": -0.5 }, "nearby_settings": [ { "cfg": { "lr": 0.0003, "beta2": 0.99, "metric_beta": -0.5 }, "mean": 216.299165725708 }, { "cfg": { "lr": 0.001, "beta2": 0.99, "metric_beta": -0.5 }, "mean": 81.73757028579712 }, { "cfg": { "lr": 0.003, "beta2": 0.999, "metric_beta": -0.5 }, "mean": 15.638604044914246 } ] }, "comparison": { "delta_mean": -2.027593970298767, "idea_wins": 8, "n_pairs": 8, "per_seed_diffs": [ -2.6044769287109375, -1.712759017944336, -3.8560352325439453, -1.0079898834228516, -3.154913902282715, -0.3759956359863281, -2.3300533294677734, -1.17852783203125 ], "p_value": 0.0081, "mde": 0.9805142418858956, "mde_rel_pct": 5.550227847788975, "verdict": "idea better (significant)", "system_worked": true }, "mechanism_signature": { "predicted": "m decreases as observed v increases (inverse-variance metric)", "observed_metric_mean": 3.7805249429075047, "observed_log_metric_vs_negative_log_variance_corr": 0.8070462500798509, "baseline_final_train_loss": 17.63215286254883, "idea_final_train_loss": 15.715824356079102, "max_constraint_violation": 0.0, "confirmed": true }, "protocol_note": "Baseline and idea used identical registered tabular task, mlp_tiny architecture, 12 epochs, batch 128, and the same learning-rate union; baseline Adam beta2 was swept." }, "how_to_run": "/home/maxwelhelp/main/bin/python3 stage2_metric_bench.py", "files": [ "stage2_metric_bench.py", "bench_report.json" ], "limitations": "Only the structurally matched registered tabular track was tested. General polytope projection, wall-clock/FLOP measurements, alternative architectures, and broader metric parameterizations were not tested.", "system_verdict": "worked", "practical_verdict": "helps", "mechanism_ok": 1, "system_judged": true }