Inverse-Square Fractional Attention / report_bench_2026-09-04T211721.md
Failed on benchmark
Стенд-проверка (stage-2) · промт оператора:
(универсальный)
Ответ агента:
{ "worked": false, "confidence": 9, "verdict": "Implemented inverse-square fractional attention in a matched transformer on the sequence forecast benchmark using train_model, a shared learning-rate grid, 8 paired seeds, and the benchmark permutation test. The math check passed, but the idea significantly worsened test MSE: 0.137701 versus 0.132134, paired delta +0.005567, p=0.0386. The trained-model mechanism signature was not confirmed because the observed attention odds ratio did not quantitatively match the kernel prediction.", "metrics": { "baseline": "Best lr=0.0015; 8-seed sequence test MSE mean 0.132134, std 0.012442", "idea": "Best shared-grid lr=0.006; 8-seed sequence test MSE mean 0.137701, std 0.012268", "comparison": "paired delta +0.005567; idea wins 2/8 seeds; permutation p=0.0386; verdict idea worse (significant)", "math_check": "Scale-homogeneity max relative error 3.40e-05; near-origin amplification 40.8057 predicted and observed; admissibility passed", "mechanism_signature": "Predicted kernel ratio 2.5716e9; observed trained attention odds ratio 4.5553e5; confirmed=false" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 bench_run.py", "files": [ "bench_run.py", "bench_report.json" ], "limitations": "Only the structurally matched sequence track was tested. Fixed s, sigma, lambda, and center were not swept; no vision or point-cloud benchmark was available. The run used 18 epochs and 2000/500 train/test samples. The temporal coordinate embedding makes the self-distance comparison numerically extreme, so the signature was conservatively marked unconfirmed.", "bench_report": { "bench_version": 1, "track": "sequence", "model": "transformer_tiny", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.0015, "weight_decay": 0.0 }, "sweep": [ { "cfg": { "lr": 0.0015, "weight_decay": 0.0 }, "mean": 0.13084225729107857 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 0.13562202639877796 }, { "cfg": { "lr": 0.006, "weight_decay": 0.0 }, "mean": 0.1526990309357643 } ], "full": { "mean": 0.1321341758593917, "std": 0.012441957039020079, "per_seed": [ 0.15123596787452698, 0.11964408308267593, 0.12323945015668869, 0.12924952805042267, 0.1238316223025322, 0.11987392604351044, 0.13868176937103271, 0.1513170599937439 ], "n": 8 } }, "idea": { "mean": 0.13770121987909079, "std": 0.012267757502007736, "per_seed": [ 0.14959165453910828, 0.1260286271572113, 0.1206536665558815, 0.13636907935142517, 0.1417829692363739, 0.12556812167167664, 0.142854243516922, 0.1587613970041275 ], "n": 8 }, "comparison": { "delta_mean": 0.005567044019699097, "idea_wins": 2, "n_pairs": 8, "per_seed_diffs": [ -0.0016443133354187012, 0.00638454407453537, -0.00258578360080719, 0.0071195513010025024, 0.017951346933841705, 0.005694195628166199, 0.004172474145889282, 0.007444337010383606 ], "p_value": 0.0386, "mde": 0.00529214319698178, "mde_rel_pct": 4.00512824374318, "verdict": "idea worse (significant)", "system_worked": false }, "idea_sweep": [ { "cfg": { "lr": 0.0015, "weight_decay": 0.0 }, "mean": 0.1338496170938015 }, { "cfg": { "lr": 0.003, "weight_decay": 0.0 }, "mean": 0.13636435382068157 }, { "cfg": { "lr": 0.006, "weight_decay": 0.0 }, "mean": 0.13316075690090656 } ], "math_check": { "homogeneity_max_relative_error": 3.397532652615887e-05, "near_origin_amplification": 40.80571546736738, "predicted_amplification": 40.80571546736738, "admissible_d3": true, "relative_amplification_error": 0.0 }, "idea_best_cfg": { "lr": 0.006, "weight_decay": 0.0 }, "mechanism_signature": { "predicted_near_vs_far_kernel_ratio": 2571602743.0040903, "observed_attention_odds_ratio_after_bias": 455527.4706971457, "baseline_endpoint_recent_attention": 0.04563657566905022, "idea_endpoint_recent_attention": 0.9985866546630859, "confirmed": false } }, "system_verdict": "failed", "practical_verdict": "harms", "mechanism_ok": 0, "system_judged": true }