Centered Heavy-Tail Clipping Optimizer / report_bench_2026-09-04T163450.md

Failed on benchmark

Raw ⬇ ZIP

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

(универсальный)

Ответ агента:

{ "worked": false, "confidence": 10, "verdict": "The centered heavy-tail clipping optimizer was tested on the registered tabular Friedman#1 benchmark with the shared mlp_tiny architecture and 8 paired seeds. The tuned baseline achieved mean test MSE 11.4363, while the best centered-clipping configuration achieved 15.0967; the idea was significantly worse, with paired permutation p=0.03245. The trained-model mechanism signature was not confirmed because the observed mean directional cosine was 0.5657, so there is no benchmark win.", "metrics": { "baseline": "Adam with global norm clipping; best lr=0.024, tau=1.0; mean MSE 11.4363187551.", "idea": "Adam with coordinate-median centered residual clipping; best lr=0.024, tau=1.0; mean MSE 15.0966548920.", "mechanism_signature": "Residual bound matched tau numerically, but observed directional cosine mean was 0.5656541213; confirmed=false." }, "bench_report": { "bench_version": 1, "track": "tabular", "model": "mlp_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.024, "tau": 1.0 }, "sweep": [ { "cfg": { "lr": 0.0015, "tau": 0.5 }, "mean": 149.6931171417 }, { "cfg": { "lr": 0.0015, "tau": 1.0 }, "mean": 149.6915607452 }, { "cfg": { "lr": 0.0015, "tau": 2.0 }, "mean": 149.6908798218 }, { "cfg": { "lr": 0.003, "tau": 0.5 }, "mean": 20.2301805019 }, { "cfg": { "lr": 0.003, "tau": 1.0 }, "mean": 20.2300088406 }, { "cfg": { "lr": 0.003, "tau": 2.0 }, "mean": 20.2283802032 }, { "cfg": { "lr": 0.006, "tau": 0.5 }, "mean": 17.8110207319 }, { "cfg": { "lr": 0.006, "tau": 1.0 }, "mean": 17.8141782284 }, { "cfg": { "lr": 0.006, "tau": 2.0 }, "mean": 17.8153961897 }, { "cfg": { "lr": 0.012, "tau": 0.5 }, "mean": 13.6178820133 }, { "cfg": { "lr": 0.012, "tau": 1.0 }, "mean": 13.623129487 }, { "cfg": { "lr": 0.012, "tau": 2.0 }, "mean": 13.627903223 }, { "cfg": { "lr": 0.024, "tau": 0.5 }, "mean": 11.4565466046 }, { "cfg": { "lr": 0.024, "tau": 1.0 }, "mean": 11.4363187551 }, { "cfg": { "lr": 0.024, "tau": 2.0 }, "mean": 11.4457081556 } ], "full": { "mean": 11.4363187551, "std": 2.8020072672, "per_seed": [ 13.5391349792, 14.802740097, 14.9247140884, 9.8293781281, 7.3404846191, 7.8995065689, 10.3567380905, 12.7978534698 ], "n": 8 } }, "idea": { "mean": 15.096654892, "std": 1.5912454315, "per_seed": [ 14.8751220703, 15.7590332031, 13.4888353348, 14.0564727783, 13.7840528488, 13.7981147766, 16.7694339752, 18.2421741486 ], "n": 8, "cfg": { "lr": 0.024, "tau": 1.0 } }, "comparison": { "delta_mean": 3.6603361368, "idea_wins": 1, "n_pairs": 8, "per_seed_diffs": [ 1.3359870911, 0.9562931061, -1.4358787537, 4.2270946503, 6.4435682297, 5.8986082077, 6.4126958847, 5.4443206787 ], "p_value": 0.03245, "verdict": "idea worse (significant)", "system_worked": false }, "mechanism_signature": { "prediction": "Residual norm is bounded by tau and common-direction cosine remains near 1.", "predicted_residual_bound": 1.0, "observed_max_residual_mean": 1.0000003725, "observed_direction_cosine_mean": 0.5656541213, "confirmed": false } }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_experiment.py", "files": [ "bench_experiment.py", "bench_report.json" ], "limitations": "Only the registered tabular optimizer track was tested. No vision, sequence, dynamics, adaptive tau, geometric median, SGD comparison, wall-clock/FLOP accounting, or broader tail/control sweep was run. Explicit per-example gradients make the implementation computationally expensive.", "system_verdict": "failed", "practical_verdict": "harms", "mechanism_ok": 0, "system_judged": true }