Add TreeSHAP references to documentation (#12207)

Author: RAMitchell

doc/contrib/featuremap.rst

@@ -45,7 +45,8 @@ XGBoost includes features designed to improve understanding of the model. Here's
 - Tree visualization.
 - Tree as dataframe.
 
-For GPU support, the SHAP value uses XGBoost's in-tree ``QuadratureTreeSHAP`` implementation. It supports categorical features, while vector-leaf is still in progress.
+SHAP values use XGBoost's in-tree ``QuadratureTreeSHAP`` implementation on both CPU and
+GPU. It supports categorical features, while vector-leaf is still in progress.
 
 ----------
 Evaluation

doc/gpu/index.rst

@@ -35,7 +35,10 @@ The GPU algorithms currently work with CLI, Python, R, and JVM packages. See :do
 
 GPU-Accelerated SHAP values
 =============================
-XGBoost provides an in-tree GPU implementation of ``QuadratureTreeSHAP`` for computing SHAP values when the GPU is used.
+XGBoost uses its in-tree ``QuadratureTreeSHAP`` implementation for computing SHAP values
+on both CPU and GPU. The GPU path uses the same Quadrature-TreeSHAP formulation described
+by Wettenstein et al. (2026) for exact TreeSHAP feature attributions when the GPU is
+selected.
 
 .. code-block:: python
 
@@ -92,6 +95,10 @@ References
 **********
 `Mitchell R, Frank E. (2017) Accelerating the XGBoost algorithm using GPU computing. PeerJ Computer Science 3:e127 https://doi.org/10.7717/peerj-cs.127 <https://peerj.com/articles/cs-127/>`_
 
+`Lundberg SM, Erion GG, Lee S-I. (2018) Consistent Individualized Feature Attribution for Tree Ensembles. arXiv:1802.03888 <https://arxiv.org/abs/1802.03888>`_
+
+`Wettenstein R, Mitchell R, Yu P. (2026) Quadrature-TreeSHAP: Depth-Independent TreeSHAP and Shapley Interactions. arXiv:2605.04497 <https://arxiv.org/abs/2605.04497>`_
+
 `NVIDIA Parallel Forall: Gradient Boosting, Decision Trees and XGBoost with CUDA <https://devblogs.nvidia.com/parallelforall/gradient-boosting-decision-trees-xgboost-cuda/>`_
 
 `Out-of-Core GPU Gradient Boosting <https://arxiv.org/abs/2005.09148>`_

doc/prediction.rst

@@ -39,8 +39,13 @@ After 1.4 release, we added a new parameter called ``strict_shape``, one can set
 
   Output is a 3-dim array, with ``(rows, groups, columns + 1)`` as shape.  Whether
   ``approx_contribs`` is used does not change the output shape. If the strict shape
-  parameter is not set, it can be a 2 or 3 dimension array depending on whether
-  multi-class model is being used.
+  parameter is not set, it can be a 2 or 3 dimension array depending on whether multi-class
+  model is being used. See `Lundberg et al. (2018)`_ for the original TreeSHAP feature
+  attribution method. When ``approx_contribs`` is ``False``, XGBoost uses the
+  ``QuadratureTreeSHAP`` implementation described by `Wettenstein et al. (2026)`_ for exact
+  TreeSHAP feature attributions on both CPU and GPU. When ``approx_contribs`` is ``True``,
+  XGBoost uses an approximate contribution method on CPU; the GPU predictor does not
+  implement approximated contributions.
 
 - When using ``pred_interactions`` with ``strict_shape`` set to ``True``:
 
@@ -191,3 +196,5 @@ More information and examples are given in the `Concrete ML documentation`_.
 .. _Zama: https://www.zama.ai/
 .. _Concrete ML: https://github.com/zama-ai/concrete-ml
 .. _Concrete ML documentation: https://docs.zama.ai/concrete-ml
+.. _Lundberg et al. (2018): https://arxiv.org/abs/1802.03888
+.. _Wettenstein et al. (2026): https://arxiv.org/abs/2605.04497