Observable-Reduced Neural World Model / report_bench_2026-09-01T103350.md

✓✓ Beats tuned baseline

Raw ⬇ ZIP

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

(универсальный)

Ответ агента:

{ "worked": true, "confidence": 7, "verdict": "The observable-reduced second-order GRU significantly outperformed the tuned generic GRU on the registered dynamics benchmark. Test MSE was 5.279662751433989e-08 versus 0.0002669833438631031, all 8 paired seeds improved, and the permutation p-value was 0.0081. The exact two-compartment invariant was not confirmed because the built-in pendulum task has no aggregate two-compartment latent ground truth.", "metrics": { "baseline": "Generic GRU, lr=0.006, 12 epochs: test MSE 0.0002669833438631031", "idea": "Structured second-order GRU, lr=0.006, 12 epochs: test MSE 5.279662751433989e-08; paired delta -0.00026693054723558873; permutation p=0.0081; 8/8 wins" }, "how_to_run": "CUDA_VISIBLE_DEVICES='' /home/maxwelhelp/main/bin/python3 bench_observable_reduced.py", "files": [ "bench_observable_reduced.py", "bench_report.json", "bench_stdout.txt" ], "limitations": "The built-in dynamics task is an actuated-pendulum forecast rather than the proposed aggregate two-compartment system, so cy=betar and cv=r-alpha-beta could not be retested. The idea model has additional coefficient and residual heads, so parameter-count parity was not separately controlled. Long-horizon rollout and baseline weight-decay sweeps were not tested.", "bench_report": { "bench_version": 1, "track": "dynamics", "model": "rnn_small", "metric_direction": "lower is better", "n_seeds": 8, "baseline": { "best_cfg": { "lr": 0.006, "epochs": 12 }, "sweep": [ { "cfg": { "lr": 0.001, "epochs": 12 }, "mean": 0.0010860307957045734 }, { "cfg": { "lr": 0.003, "epochs": 12 }, "mean": 0.0005809691647300497 }, { "cfg": { "lr": 0.006, "epochs": 12 }, "mean": 0.0002912725685746409 } ], "full": { "mean": 0.0002669833438631031, "std": 8.681528355373122e-05, "per_seed": [ 0.00031876337016001344, 0.0002849859884008765, 0.00017326572560705245, 0.0003880751901306212, 0.00033014125074259937, 0.0003085587522946298, 0.00010380324238212779, 0.000228273231186904 ], "n": 8 } }, "idea": { "best_cfg": { "lr": 0.006, "epochs": 12, "penalty": 0.0 }, "sweep": [ { "cfg": { "lr": 0.001, "epochs": 12, "penalty": 0.0 }, "mean": 1.0584739413843636e-06, "per_seed": [ 1.0828274525920278e-06, 1.0643407222232781e-06, 1.100200734072132e-06, 1.1011246670022956e-06, 1.0659464351192582e-06, 1.0385464292994584e-06, 1.0149252602786873e-06, 9.99879830487771e-07 ] }, { "cfg": { "lr": 0.003, "epochs": 12, "penalty": 0.0 }, "mean": 7.993562256558562e-08, "per_seed": [ 6.861665013957463e-08, 7.364582188529312e-08, 9.95038575979379e-08, 8.163866027643962e-08, 8.44757082063552e-08, 8.488292024821931e-08, 7.652040068251154e-08, 7.020096148835364e-08 ] }, { "cfg": { "lr": 0.006, "epochs": 12, "penalty": 0.0 }, "mean": 5.279662751433989e-08, "per_seed": [ 4.7656406820806296e-08, 5.733669539154107e-08, 5.804359304306672e-08, 4.7830777560875504e-08, 5.6010112103876963e-08, 5.760812271660143e-08, 4.9883055908139795e-08, 4.800425656981133e-08 ] } ], "mean": 5.279662751433989e-08, "std": 4.5298604585404984e-09, "per_seed": [ 4.7656406820806296e-08, 5.733669539154107e-08, 5.804359304306672e-08, 4.7830777560875504e-08, 5.6010112103876963e-08, 5.760812271660143e-08, 4.9883055908139795e-08 ], "n": 8 }, "comparison": { "delta_mean": -0.00026693054723558873, "idea_wins": 8, "n_pairs": 8, "per_seed_diffs": [ -0.00031871571375319263, -0.000284928651705485, -0.00017320768201400938, -0.0003880273593530603, -0.0003300852406304955, -0.0003085011441719132, -0.00010375335932621965, -0.0002282252269303342 ], "p_value": 0.0081, "mde": 7.760323964174408e-05, "mde_rel_pct": 29.066697015202376, "verdict": "idea better (significant)", "system_worked": true }, "mechanism_signature": { "prediction": "second-order structured dynamics improves extrapolative neural prediction", "trained_model_signature": { "output_target_corr_mean": 0.9999999671294927, "output_rmse_mean": 0.00022956307111599017, "confirmed": false, "note": "This built-in task has no aggregate two-compartment ground truth, so the exact cy=betar invariant cannot be quantitatively retested." }, "structural_match": "dynamics stability/control" } }, "system_verdict": "worked", "practical_verdict": "helps", "mechanism_ok": 0, "system_judged": true }