{
  "metadata": {
    "train_tasks": [
      "aime",
      "gpqa",
      "hmmt"
    ],
    "test_tasks": [
      "supergpqa",
      "mmlupro",
      "usaco"
    ],
    "models": [
      "Ministral-3-14B-Reasoning-2512",
      "Ministral-3-3B-Reasoning-2512",
      "Ministral-3-8B-Reasoning-2512",
      "Nemotron-Cascade-14B-Thinking",
      "Nemotron-Cascade-8B-Thinking",
      "Olmo-3-32B-Think",
      "Qwen3-0.6B",
      "Qwen3-1.7B",
      "Qwen3-14B",
      "Qwen3-30B-A3B-Thinking-2507",
      "Qwen3-32B",
      "R1-Distill-Llama-70B",
      "R1-Distill-Llama-8B",
      "R1-Distill-Qwen-14B",
      "R1-Distill-Qwen-32B",
      "R1-Distill-Qwen-7B",
      "SmolLM3-3B"
    ],
    "n_features": 80,
    "feature_names": [
      "ce_loss:entropy",
      "ce_loss:freq",
      "ce_loss:gauss_nll",
      "ce_loss:inv_entropy",
      "ce_loss:inv_freq",
      "ce_loss:prob",
      "ce_loss:surprise",
      "ce_loss:uniform",
      "entropy:entropy",
      "entropy:freq",
      "entropy:gauss_nll",
      "entropy:inv_entropy",
      "entropy:inv_freq",
      "entropy:prob",
      "entropy:surprise",
      "entropy:uniform",
      "margin:entropy",
      "margin:freq",
      "margin:gauss_nll",
      "margin:inv_entropy",
      "margin:inv_freq",
      "margin:prob",
      "margin:surprise",
      "margin:uniform",
      "rank:entropy",
      "rank:freq",
      "rank:gauss_nll",
      "rank:inv_entropy",
      "rank:inv_freq",
      "rank:prob",
      "rank:surprise",
      "rank:uniform",
      "reciprocal_rank:entropy",
      "reciprocal_rank:freq",
      "reciprocal_rank:gauss_nll",
      "reciprocal_rank:inv_entropy",
      "reciprocal_rank:inv_freq",
      "reciprocal_rank:prob",
      "reciprocal_rank:surprise",
      "reciprocal_rank:uniform",
      "top_1_correct:entropy",
      "top_1_correct:freq",
      "top_1_correct:gauss_nll",
      "top_1_correct:inv_entropy",
      "top_1_correct:inv_freq",
      "top_1_correct:prob",
      "top_1_correct:surprise",
      "top_1_correct:uniform",
      "top_2_correct:entropy",
      "top_2_correct:freq",
      "top_2_correct:gauss_nll",
      "top_2_correct:inv_entropy",
      "top_2_correct:inv_freq",
      "top_2_correct:prob",
      "top_2_correct:surprise",
      "top_2_correct:uniform",
      "top_3_correct:entropy",
      "top_3_correct:freq",
      "top_3_correct:gauss_nll",
      "top_3_correct:inv_entropy",
      "top_3_correct:inv_freq",
      "top_3_correct:prob",
      "top_3_correct:surprise",
      "top_3_correct:uniform",
      "top_5_correct:entropy",
      "top_5_correct:freq",
      "top_5_correct:gauss_nll",
      "top_5_correct:inv_entropy",
      "top_5_correct:inv_freq",
      "top_5_correct:prob",
      "top_5_correct:surprise",
      "top_5_correct:uniform",
      "wrong_confidence:entropy",
      "wrong_confidence:freq",
      "wrong_confidence:gauss_nll",
      "wrong_confidence:inv_entropy",
      "wrong_confidence:inv_freq",
      "wrong_confidence:prob",
      "wrong_confidence:surprise",
      "wrong_confidence:uniform"
    ],
    "max_sparsity": 2,
    "objectives_run": [
      "nnls",
      "spearman_direct",
      "pairwise_hinge",
      "soft_spearman",
      "exhaustive_grid"
    ],
    "expert_overrides": {
      "aime": [
        "kimi-k2.5-suffix-1000",
        "minimax-m2.5-suffix-1000",
        "qwen3next-suffix-1000"
      ],
      "gpqa": [
        "kimi-k2.5-suffix-1000",
        "minimax-m2.5-suffix-1000",
        "qwen3next-suffix-1000"
      ],
      "hmmt": [
        "kimi-k2.5-suffix-1000",
        "minimax-m2.5-suffix-1000",
        "qwen3next-suffix-1000"
      ],
      "supergpqa": [
        "kimi-k2.5-suffix-1000",
        "minimax-m2.5-suffix-1000",
        "qwen3next-suffix-1000"
      ],
      "mmlupro": [
        "kimi-k2.5-suffix-1000",
        "minimax-m2.5-suffix-1000",
        "qwen3next-suffix-1000"
      ],
      "usaco": [
        "kimi-k2.5-suffix-1000",
        "minimax-m2.5-suffix-1000",
        "qwen3next-suffix-1000"
      ]
    },
    "expert_exclusions": "none"
  },
  "results": {
    "nnls": {
      "objective": "nnls",
      "best_K": 2,
      "best_mean_spearman": 0.5802569442658317,
      "k_scores": {
        "1": 0.5038555171054306,
        "2": 0.5802569442658317
      },
      "k_details": {
        "1": {
          "selected_indices": [
            44
          ],
          "selected_features": [
            "top_1_correct:inv_freq"
          ],
          "mean_spearman": 0.5038555171054306,
          "coefficients": [
            1.0
          ],
          "per_task_spearman": {
            "aime": 0.43803713955158596,
            "gpqa": 0.6348039215686275,
            "hmmt": 0.4387254901960785
          },
          "test_per_task_spearman": {
            "supergpqa": 0.8382352941176472,
            "mmlupro": 0.7916666666666667,
            "usaco": 0.6544117647058824
          },
          "test_mean_spearman": 0.7614379084967321
        },
        "2": {
          "selected_indices": [
            44,
            8
          ],
          "selected_features": [
            "top_1_correct:inv_freq",
            "entropy:entropy"
          ],
          "mean_spearman": 0.5802569442658317,
          "coefficients": [
            0.28986906093212395,
            1.0
          ],
          "per_task_spearman": {
            "aime": 0.46871200926808365,
            "gpqa": 0.6519607843137255,
            "hmmt": 0.6200980392156864
          },
          "test_per_task_spearman": {
            "supergpqa": 0.6151960784313726,
            "mmlupro": 0.661764705882353,
            "usaco": 0.607843137254902
          },
          "test_mean_spearman": 0.6282679738562092
        }
      },
      "final_model": {
        "selected_features": [
          "top_1_correct:inv_freq",
          "entropy:entropy"
        ],
        "coefficients": [
          0.28986906093212395,
          1.0
        ],
        "per_task_spearman": {
          "aime": 0.46871200926808365,
          "gpqa": 0.6519607843137255,
          "hmmt": 0.6200980392156864
        },
        "mean_spearman": 0.5802569442658317,
        "test_per_task_spearman": {
          "supergpqa": 0.6151960784313726,
          "mmlupro": 0.661764705882353,
          "usaco": 0.607843137254902
        },
        "test_mean_spearman": 0.6282679738562092
      }
    },
    "spearman_direct": {
      "objective": "spearman_direct",
      "best_K": 2,
      "best_mean_spearman": 0.722080038377117,
      "k_scores": {
        "1": 0.5038555171054306,
        "2": 0.722080038377117
      },
      "k_details": {
        "1": {
          "selected_indices": [
            44
          ],
          "selected_features": [
            "top_1_correct:inv_freq"
          ],
          "mean_spearman": 0.5038555171054306,
          "coefficients": [
            1.0
          ],
          "per_task_spearman": {
            "aime": 0.43803713955158596,
            "gpqa": 0.6348039215686275,
            "hmmt": 0.4387254901960785
          },
          "test_per_task_spearman": {
            "supergpqa": 0.8382352941176472,
            "mmlupro": 0.7916666666666667,
            "usaco": 0.6544117647058824
          },
          "test_mean_spearman": 0.7614379084967321
        },
        "2": {
          "selected_indices": [
            44,
            12
          ],
          "selected_features": [
            "top_1_correct:inv_freq",
            "entropy:inv_freq"
          ],
          "mean_spearman": 0.722080038377117,
          "coefficients": [
            1.0,
            0.8615150407299678
          ],
          "per_task_spearman": {
            "aime": 0.6515342327784095,
            "gpqa": 0.8284313725490197,
            "hmmt": 0.6862745098039217
          },
          "test_per_task_spearman": {
            "supergpqa": 0.8946078431372549,
            "mmlupro": 0.8627450980392157,
            "usaco": 0.7303921568627452
          },
          "test_mean_spearman": 0.829248366013072
        }
      },
      "final_model": {
        "selected_features": [
          "top_1_correct:inv_freq",
          "entropy:inv_freq"
        ],
        "coefficients": [
          1.0,
          0.8615150407299678
        ],
        "per_task_spearman": {
          "aime": 0.6515342327784095,
          "gpqa": 0.8284313725490197,
          "hmmt": 0.6862745098039217
        },
        "mean_spearman": 0.722080038377117,
        "test_per_task_spearman": {
          "supergpqa": 0.8946078431372549,
          "mmlupro": 0.8627450980392157,
          "usaco": 0.7303921568627452
        },
        "test_mean_spearman": 0.829248366013072
      }
    },
    "pairwise_hinge": {
      "objective": "pairwise_hinge",
      "best_K": 2,
      "best_mean_spearman": 0.70041817636994,
      "k_scores": {
        "1": 0.5038555171054306,
        "2": 0.70041817636994
      },
      "k_details": {
        "1": {
          "selected_indices": [
            44
          ],
          "selected_features": [
            "top_1_correct:inv_freq"
          ],
          "mean_spearman": 0.5038555171054306,
          "coefficients": [
            1.0
          ],
          "per_task_spearman": {
            "aime": 0.43803713955158596,
            "gpqa": 0.6348039215686275,
            "hmmt": 0.4387254901960785
          },
          "test_per_task_spearman": {
            "supergpqa": 0.8382352941176472,
            "mmlupro": 0.7916666666666667,
            "usaco": 0.6544117647058824
          },
          "test_mean_spearman": 0.7614379084967321
        },
        "2": {
          "selected_indices": [
            44,
            8
          ],
          "selected_features": [
            "top_1_correct:inv_freq",
            "entropy:entropy"
          ],
          "mean_spearman": 0.70041817636994,
          "coefficients": [
            1.0,
            0.8160187134197638
          ],
          "per_task_spearman": {
            "aime": 0.6233133526392316,
            "gpqa": 0.823529411764706,
            "hmmt": 0.6544117647058824
          },
          "test_per_task_spearman": {
            "supergpqa": 0.8431372549019609,
            "mmlupro": 0.8333333333333335,
            "usaco": 0.7794117647058825
          },
          "test_mean_spearman": 0.8186274509803922
        }
      },
      "final_model": {
        "selected_features": [
          "top_1_correct:inv_freq",
          "entropy:entropy"
        ],
        "coefficients": [
          1.0,
          0.8160187134197638
        ],
        "per_task_spearman": {
          "aime": 0.6233133526392316,
          "gpqa": 0.823529411764706,
          "hmmt": 0.6544117647058824
        },
        "mean_spearman": 0.70041817636994,
        "test_per_task_spearman": {
          "supergpqa": 0.8431372549019609,
          "mmlupro": 0.8333333333333335,
          "usaco": 0.7794117647058825
        },
        "test_mean_spearman": 0.8186274509803922
      }
    },
    "soft_spearman": {
      "objective": "soft_spearman",
      "best_K": 2,
      "best_mean_spearman": 0.7143130836291549,
      "k_scores": {
        "1": 0.5038555171054306,
        "2": 0.7143130836291549
      },
      "k_details": {
        "1": {
          "selected_indices": [
            44
          ],
          "selected_features": [
            "top_1_correct:inv_freq"
          ],
          "mean_spearman": 0.5038555171054306,
          "coefficients": [
            1.0
          ],
          "per_task_spearman": {
            "aime": 0.43803713955158596,
            "gpqa": 0.6348039215686275,
            "hmmt": 0.4387254901960785
          },
          "test_per_task_spearman": {
            "supergpqa": 0.8382352941176472,
            "mmlupro": 0.7916666666666667,
            "usaco": 0.6544117647058824
          },
          "test_mean_spearman": 0.7614379084967321
        },
        "2": {
          "selected_indices": [
            44,
            12
          ],
          "selected_features": [
            "top_1_correct:inv_freq",
            "entropy:inv_freq"
          ],
          "mean_spearman": 0.7143130836291549,
          "coefficients": [
            0.9973068140739522,
            1.0
          ],
          "per_task_spearman": {
            "aime": 0.6380372901031505,
            "gpqa": 0.8333333333333335,
            "hmmt": 0.6715686274509804
          },
          "test_per_task_spearman": {
            "supergpqa": 0.875,
            "mmlupro": 0.8651960784313727,
            "usaco": 0.7843137254901962
          },
          "test_mean_spearman": 0.8415032679738562
        }
      },
      "final_model": {
        "selected_features": [
          "top_1_correct:inv_freq",
          "entropy:inv_freq"
        ],
        "coefficients": [
          0.9973068140739522,
          1.0
        ],
        "per_task_spearman": {
          "aime": 0.6380372901031505,
          "gpqa": 0.8333333333333335,
          "hmmt": 0.6715686274509804
        },
        "mean_spearman": 0.7143130836291549,
        "test_per_task_spearman": {
          "supergpqa": 0.875,
          "mmlupro": 0.8651960784313727,
          "usaco": 0.7843137254901962
        },
        "test_mean_spearman": 0.8415032679738562
      }
    },
    "exhaustive_grid": {
      "objective": "exhaustive_grid",
      "best_K": 2,
      "best_mean_spearman": 0.7736928104575164,
      "k_scores": {
        "1": 0.5106209150326797,
        "2": 0.7736928104575164
      },
      "k_details": {
        "1": {
          "selected_indices": [
            44
          ],
          "selected_features": [
            "top_1_correct:inv_freq"
          ],
          "mean_spearman": 0.5106209150326797,
          "coefficients": [
            1.0
          ],
          "per_task_spearman": {
            "aime": 0.4583333333333333,
            "gpqa": 0.6348039215686274,
            "hmmt": 0.4387254901960784
          },
          "test_per_task_spearman": {
            "supergpqa": 0.8382352941176472,
            "mmlupro": 0.7916666666666667,
            "usaco": 0.6544117647058824
          },
          "test_mean_spearman": 0.7614379084967321
        },
        "2": {
          "selected_indices": [
            12,
            76
          ],
          "selected_features": [
            "entropy:inv_freq",
            "wrong_confidence:inv_freq"
          ],
          "mean_spearman": 0.7736928104575164,
          "coefficients": [
            2.381685551976158,
            1.0
          ],
          "per_task_spearman": {
            "aime": 0.7230392156862745,
            "gpqa": 0.8455882352941176,
            "hmmt": 0.7524509803921569
          },
          "test_per_task_spearman": {
            "supergpqa": 0.8406862745098039,
            "mmlupro": 0.8578431372549021,
            "usaco": 0.8014705882352942
          },
          "test_mean_spearman": 0.8333333333333334
        }
      },
      "final_model": {
        "selected_features": [
          "entropy:inv_freq",
          "wrong_confidence:inv_freq"
        ],
        "coefficients": [
          2.381685551976158,
          1.0
        ],
        "per_task_spearman": {
          "aime": 0.7230392156862745,
          "gpqa": 0.8455882352941176,
          "hmmt": 0.7524509803921569
        },
        "mean_spearman": 0.7736928104575164,
        "test_per_task_spearman": {
          "supergpqa": 0.8406862745098039,
          "mmlupro": 0.8578431372549021,
          "usaco": 0.8014705882352942
        },
        "test_mean_spearman": 0.8333333333333334
      }
    }
  },
  "best_objective": "exhaustive_grid",
  "best_result": {
    "objective": "exhaustive_grid",
    "best_K": 2,
    "best_mean_spearman": 0.7736928104575164,
    "k_scores": {
      "1": 0.5106209150326797,
      "2": 0.7736928104575164
    },
    "k_details": {
      "1": {
        "selected_indices": [
          44
        ],
        "selected_features": [
          "top_1_correct:inv_freq"
        ],
        "mean_spearman": 0.5106209150326797,
        "coefficients": [
          1.0
        ],
        "per_task_spearman": {
          "aime": 0.4583333333333333,
          "gpqa": 0.6348039215686274,
          "hmmt": 0.4387254901960784
        },
        "test_per_task_spearman": {
          "supergpqa": 0.8382352941176472,
          "mmlupro": 0.7916666666666667,
          "usaco": 0.6544117647058824
        },
        "test_mean_spearman": 0.7614379084967321
      },
      "2": {
        "selected_indices": [
          12,
          76
        ],
        "selected_features": [
          "entropy:inv_freq",
          "wrong_confidence:inv_freq"
        ],
        "mean_spearman": 0.7736928104575164,
        "coefficients": [
          2.381685551976158,
          1.0
        ],
        "per_task_spearman": {
          "aime": 0.7230392156862745,
          "gpqa": 0.8455882352941176,
          "hmmt": 0.7524509803921569
        },
        "test_per_task_spearman": {
          "supergpqa": 0.8406862745098039,
          "mmlupro": 0.8578431372549021,
          "usaco": 0.8014705882352942
        },
        "test_mean_spearman": 0.8333333333333334
      }
    },
    "final_model": {
      "selected_features": [
        "entropy:inv_freq",
        "wrong_confidence:inv_freq"
      ],
      "coefficients": [
        2.381685551976158,
        1.0
      ],
      "per_task_spearman": {
        "aime": 0.7230392156862745,
        "gpqa": 0.8455882352941176,
        "hmmt": 0.7524509803921569
      },
      "mean_spearman": 0.7736928104575164,
      "test_per_task_spearman": {
        "supergpqa": 0.8406862745098039,
        "mmlupro": 0.8578431372549021,
        "usaco": 0.8014705882352942
      },
      "test_mean_spearman": 0.8333333333333334
    }
  }
}