feat(ara): remove tie_to_original_matrix term

This term was found experimentally to be 3-4 orders of magnitude smaller than the others in most runs, and have no meaningful effect on the result of the optimization.

src/heretic/main.py

@@ -484,11 +484,6 @@ def run():
                 0.0,
                 1.0,
             )
-            tie_to_original_matrix_weight = trial.suggest_float(
-                "tie_to_original_matrix_weight",
-                0.2,  # Minimum to prevent "optimizing" away the regularization term.
-                1.0,
-            )
         else:
             direction_scope = trial.suggest_categorical(
                 "direction_scope",
@@ -576,7 +571,6 @@ def run():
                 end_layer_index,
                 preserve_good_behavior_weight,
                 steer_bad_behavior_weight,
-                tie_to_original_matrix_weight,
             )
         else:
             print("* Resetting model...")
@@ -772,7 +766,6 @@ def run():
                     trial.params["end_layer_index"],
                     trial.params["preserve_good_behavior_weight"],
                     trial.params["steer_bad_behavior_weight"],
-                    trial.params["tie_to_original_matrix_weight"],
                 )
             else:
                 print("* Resetting model...")

src/heretic/model.py

@@ -545,7 +545,6 @@ class Model:
         end_layer_index: int,
         preserve_good_behavior_weight: float,
         steer_bad_behavior_weight: float,
-        tie_to_original_matrix_weight: float,
     ):
         for layer_index in range(start_layer_index, end_layer_index):
             for component, modules in self.get_layer_modules(layer_index).items():
@@ -589,17 +588,9 @@ class Model:
                             ** 2
                         ).mean()
 
-                        # The matrix itself should change as little as possible overall.
-                        # This prevents overfitting due to underdetermination of the
-                        # optimization problem from a relatively small number of I/O pairs.
-                        tie_to_original_matrix = (
-                            (matrix - original_matrix) ** 2
-                        ).mean()
-
                         return (
                             preserve_good_behavior_weight * preserve_good_behavior
                             + steer_bad_behavior_weight * steer_bad_behavior
-                            + tie_to_original_matrix_weight * tie_to_original_matrix
                         )
 
                     optimizer = LBFGS(

src/heretic/utils.py

@@ -257,7 +257,6 @@ def get_trial_parameters(settings: Settings, trial: Trial) -> dict[str, str]:
             "end_layer_index": f"{trial.params['end_layer_index']}",
             "preserve_good_behavior_weight": f"{trial.params['preserve_good_behavior_weight']:.4f}",
             "steer_bad_behavior_weight": f"{trial.params['steer_bad_behavior_weight']:.4f}",
-            "tie_to_original_matrix_weight": f"{trial.params['tie_to_original_matrix_weight']:.4f}",
         }
     else:
         params = {}