Dissecting Hybrid Neuro-Symbolic Systems: An Architectural Ablation Study of the Latent Probabilistic Framework

We present a comprehensive architectural ablation study of the Latent Probabilistic Framework (LPF), a hybrid neuro-symbolic system for evidence aggregation under uncertainty. Through systematic evaluation across 15 random seeds (n = 135 test samples each), we identify the critical components driving performance and calibration. Our key findings reveal: (1) Symbolic SPN reasoning is the most critical component, with its removal causing a 6.1 percentage point accuracy drop (96.7% → 90.6%, p < 0.001) and nearly tripling calibration error (ECE: 2.3% → 6.1%); (2) Monte Carlo sampling provides meaningful gains, contributing 1.9 percentage points despite 16× computational cost; (3) Surprisingly, simple averaging in VAE latent space achieves 95.6% accuracy with perfect cross-seed consistency (σ = 0.0%), only 1.1 points below the full system, suggesting the VAE learns highly structured representations; (4) Temperature scaling and weight penalties are refinements rather than necessities, with T = 1.0 (no scaling) actually improving calibration (ECE: 2.3% → 1.7%). These findings demonstrate that while neural components learn effective encodings, symbolic reasoning remains essential for calibration and the final performance margin.

LPF is instantiated in two complementary forms—LPF-SPN, which performs structured factor-based inference, and LPF-Learned, which learns evidence aggregation end-to-end—enabling a principled comparison between explicit probabilistic reasoning and learned aggregation under a shared uncertainty representation. Across eight domains (seven synthetic and the FEVER benchmark), LPF-SPN achieves high accuracy (up to 97.8%), low calibration error (ECE 1.4%), and strong probabilistic fit as measured by negative log-likelihood, substantially outperforming evidential deep learning and graph-based baselines.

Our work provides actionable insights for designing hybrid AI systems that optimally balance neural learning with symbolic constraints, and contributes (i) the first general framework bridging latent uncertainty representations with structured probabilistic reasoning \citep{Aliyu2026LPFMain}; (ii) a comprehensive ablation across 7 configurations, 15 seeds, and 14,175 total predictions; (iii) statistical significance testing with mean ± std reporting; (iv) practical configuration guidelines for deployment; and (v) formal guarantees presented in the companion paper.

Keywords: Neural-symbolic integration, ablation studies, probabilistic reasoning, variational autoencoders, sum-product networks, evidence aggregation