diff --git a/deeptab/configs/core.py b/deeptab/configs/core.py
index c4f11a29..6709c87f 100644
--- a/deeptab/configs/core.py
+++ b/deeptab/configs/core.py
@@ -2,7 +2,7 @@
 
 import warnings
 from collections.abc import Callable
-from dataclasses import dataclass, field
+from dataclasses import dataclass, field, fields
 from typing import Any
 
 import torch.nn as nn
@@ -513,6 +513,14 @@ def __post_init__(self) -> None:  # type: ignore[override]
                 "placement instead.",
             )
 
+        # Remember which field values the resolved fields above were derived
+        # from, so `to_preprocessor_kwargs` can tell when they went stale.
+        self._resolved_from = self._field_values()
+
+    def _field_values(self) -> tuple:
+        """Return the current values of all public fields, in declaration order."""
+        return tuple(getattr(self, f.name) for f in fields(self))
+
     def _resolve_output_dim(self) -> int | None:
         """Resolve `n_bins`/`n_knots` into `output_dim`, defaulting to 7 when unset."""
         legacy_values = {v for v in (self.n_bins, self.n_knots) if v is not None}
@@ -587,12 +595,19 @@ def to_preprocessor_kwargs(self) -> dict:
         equivalents in ``__post_init__``; only canonical names are ever
         returned here, since PreTab 1.0 no longer accepts the legacy ones.
 
+        Fields changed after construction, through ``set_params`` (e.g. the
+        ``preprocessing_config__*`` parameters of a scikit-learn search) or by
+        direct attribute assignment, are validated and resolved again first,
+        so the returned kwargs always reflect the current field values.
+
         Returns
         -------
         dict
             Mapping of canonical PreTab field name → value for every resolved
             field that is not ``None``.
         """
+        if self._field_values() != getattr(self, "_resolved_from", None):
+            self.__post_init__()
         kwargs = {
             "numerical_method": self._resolved_numerical_method,
             "categorical_method": self._resolved_categorical_method,
diff --git a/tests/configs/test_config_contracts.py b/tests/configs/test_config_contracts.py
index e0106133..674a13c7 100644
--- a/tests/configs/test_config_contracts.py
+++ b/tests/configs/test_config_contracts.py
@@ -31,6 +31,7 @@
     TabulaRNNConfig,
     TrainerConfig,
 )
+from deeptab.core.exceptions import InvalidParamError
 
 
 class TestTrainerConfig:
@@ -220,6 +221,28 @@ def test_sklearn_clone_independence(self):
         cloned.set_params(n_bins=999)
         assert cfg.n_bins == 32
 
+    def test_set_params_is_reflected_in_preprocessor_kwargs(self):
+        cfg = PreprocessingConfig(numerical_method="ple", output_dim=7)
+        cfg.set_params(numerical_method="standardization", output_dim=12, scaling="minmax")
+        kwargs = cfg.to_preprocessor_kwargs()
+        assert kwargs["numerical_method"] == "standardization"
+        assert kwargs["output_dim"] == 12
+        assert kwargs["scaling"] == "minmax"
+
+    def test_attribute_assignment_is_reflected_in_preprocessor_kwargs(self):
+        cfg = PreprocessingConfig(numerical_method="ple", output_dim=7)
+        cfg.output_dim = 3
+        cfg.categorical_method = "one-hot"
+        kwargs = cfg.to_preprocessor_kwargs()
+        assert kwargs["output_dim"] == 3
+        assert kwargs["categorical_method"] == "one-hot"
+
+    def test_set_params_values_are_validated_on_use(self):
+        cfg = PreprocessingConfig(output_dim=7)
+        cfg.set_params(output_dim=0)
+        with pytest.raises(InvalidParamError, match="output_dim"):
+            cfg.to_preprocessor_kwargs()
+
 
 N = 120
 
diff --git a/tests/models/test_estimator_configuration.py b/tests/models/test_estimator_configuration.py
index fe2f8aad..cd1ec34e 100644
--- a/tests/models/test_estimator_configuration.py
+++ b/tests/models/test_estimator_configuration.py
@@ -142,6 +142,32 @@ def test_set_params_nested_preprocessing_config(self):
         model.set_params(preprocessing_config__numerical_preprocessing="quantile")
         assert model.preprocessing_config.numerical_preprocessing == "quantile"
 
+    def test_set_params_nested_preprocessing_config_reaches_fit(self):
+        model = MLPRegressor(
+            model_config=MLPConfig(layer_sizes=[8]),
+            preprocessing_config=PreprocessingConfig(numerical_method="ple", output_dim=5),
+            trainer_config=_FAST_TRAINER,
+            random_state=0,
+        )
+        model.set_params(preprocessing_config__numerical_method="quantile", preprocessing_config__output_dim=3)
+        assert model._preprocessor_kwargs["numerical_method"] == "quantile"
+        assert model._preprocessor_kwargs["output_dim"] == 3
+
+        model.fit(X_reg, y_reg, accelerator="cpu", logger=False, enable_progress_bar=False)
+        feature_info = model._data_module.num_feature_info
+        assert all(info["preprocessing"].endswith("quantile") for info in feature_info.values())
+
+    def test_preprocessing_config_attribute_mutation_reaches_fit(self):
+        model = MLPRegressor(
+            model_config=MLPConfig(layer_sizes=[8]),
+            preprocessing_config=PreprocessingConfig(numerical_method="ple", output_dim=5),
+            trainer_config=_FAST_TRAINER,
+            random_state=0,
+        )
+        model.preprocessing_config.output_dim = 3
+        model.fit(X_reg, y_reg, accelerator="cpu", logger=False, enable_progress_bar=False)
+        assert all(info["dimension"] == 3 for info in model._data_module.num_feature_info.values())
+
     def test_set_params_replace_whole_config(self):
         model = MLPClassifier(
             model_config=MLPConfig(),