Financial product recommendation requires not only predictive accuracy but also regulatory-compliant explainability. Existing multi-task learning (MTL) approaches (MMoE, PLE) are validated on 2-4 homogeneous tasks, leaving open whether they scale to production-scale heterogeneous workloads where binary, multiclass, and regression tasks compete for shared capacity. We present three contributions. First, we propose Heterogeneous Expert PLE, where seven architecturally distinct experts (DeepFM, Temporal Ensemble, Hyperbolic GCN, PersLay, Causal, LightGCN, Optimal Transport) share a common basket with a FeatureRouter assigning each expert its designated feature groups, providing a structural guarantee against expert collapse and inherent explainability through business-interpretable gate weights. Second, ablation on a 13-task benchmark (7 binary + 3 multiclass + 3 regression, 1M customers) shows that loss-level inter-task transfer (adaTT) degrades performance in this heterogeneous setting because of instability in estimating 156 task-pair affinities. The single largest improvement instead comes from correcting a subtle uncertainty-weighting implementation gap where per-task loss weights were silently ignored. Third, we experiment with GradSurgery, a task-type gradient projection method that reduces the 156 task-pair problem to 3 task-type-group projections. GradSurgery shows no meaningful advantage over the PLE-only baseline while requiring significantly more VRAM, and is not adopted for production. In the ablation study, PLE with softmax gating achieves the best NDCG@3 (0.714) among gate variants. Softmax outperforms sigmoid in this heterogeneous setting, reversing findings from the homogeneous-task literature. The operational motivation - consolidating 13 individual models into a single MTL model - is validated: the shared-bottom baseline already exceeds per-task XGBoost ceilings. Companion paper: "From Prediction to Persuasion: Agentic Recommendation Reason Generation for Regulatory-Compliant Financial AI" (forthcoming).
Jeong et al. (Fri,) studied this question.