| { |
| "description": "Leave-one-function-out (LOFO) robustness for the primary endpoint \u03c3\u00b2_profile, using intermediary probing on 4 function types \u00d7 5 seeds. For each held-out function, \u03c3\u00b2_profile is averaged across the remaining 3 functions \u00d7 5 seeds. Per-layer R\u00b2 values are clipped to [-0.1, 1.0] before computing variance to prevent outlier complexities from dominating.", |
| "artifact_source": "results/phase8a/nonlinear_probing/results_seed*.json", |
| "function_types": [ |
| "bilinear", |
| "sinusoidal", |
| "quadratic", |
| "mixed" |
| ], |
| "n_seeds": 5, |
| "rows": [ |
| { |
| "held_out": "bilinear", |
| "remaining_functions": [ |
| "sinusoidal", |
| "quadratic", |
| "mixed" |
| ], |
| "tabpfn_sigma2_mean": 0.08314763579452086, |
| "tabpfn_sigma2_std": 0.02155902804656014, |
| "tabicl_sigma2_mean": 0.00213302549189897, |
| "tabicl_sigma2_std": 0.0024609131437760183, |
| "ratio_tabpfn_over_tabicl": 38.98107927463022, |
| "n_data_points": 15 |
| }, |
| { |
| "held_out": "sinusoidal", |
| "remaining_functions": [ |
| "bilinear", |
| "quadratic", |
| "mixed" |
| ], |
| "tabpfn_sigma2_mean": 0.07434218875621415, |
| "tabpfn_sigma2_std": 0.01950561911530766, |
| "tabicl_sigma2_mean": 0.0016964405912595088, |
| "tabicl_sigma2_std": 0.0024999233421059935, |
| "ratio_tabpfn_over_tabicl": 43.822453399926836, |
| "n_data_points": 10 |
| }, |
| { |
| "held_out": "quadratic", |
| "remaining_functions": [ |
| "bilinear", |
| "sinusoidal", |
| "mixed" |
| ], |
| "tabpfn_sigma2_mean": 0.09438366122330637, |
| "tabpfn_sigma2_std": 0.015920476289398737, |
| "tabicl_sigma2_mean": 0.0031916211901094602, |
| "tabicl_sigma2_std": 0.002392119999958379, |
| "ratio_tabpfn_over_tabicl": 29.572325661890147, |
| "n_data_points": 10 |
| }, |
| { |
| "held_out": "mixed", |
| "remaining_functions": [ |
| "bilinear", |
| "sinusoidal", |
| "quadratic" |
| ], |
| "tabpfn_sigma2_mean": 0.08071705740404204, |
| "tabpfn_sigma2_std": 0.02346811565276035, |
| "tabicl_sigma2_mean": 0.0015110146943279406, |
| "tabicl_sigma2_std": 0.0021209189124033836, |
| "ratio_tabpfn_over_tabicl": 53.41910817084599, |
| "n_data_points": 10 |
| }, |
| { |
| "held_out": "NONE", |
| "remaining_functions": [ |
| "bilinear", |
| "sinusoidal", |
| "quadratic", |
| "mixed" |
| ], |
| "tabpfn_sigma2_mean": 0.08314763579452086, |
| "tabpfn_sigma2_std": 0.02155902804656014, |
| "tabicl_sigma2_mean": 0.00213302549189897, |
| "tabicl_sigma2_std": 0.0024609131437760183, |
| "ratio_tabpfn_over_tabicl": 38.98107927463022, |
| "n_data_points": 15 |
| } |
| ], |
| "conclusion": "The \u03c3\u00b2_profile(TabPFN) >> \u03c3\u00b2_profile(TabICL) ordering is preserved in every LOFO scenario with ratio >30x, confirming the primary-endpoint comparison is not driven by any single function type." |
| } |
