XLOG is a CUDA‑native logic programming engine that integrates neural perception with deterministic Datalog, probabilistic inference, and epistemic world views through a typed frontend backed by a provider‑owned CUDA runtime. All reasoning modes share the device data plane, but their execution boundaries differ: ordinary Datalog and exact inference are orchestrated on the host, while certified resident recursive and Monte Carlo sampled cores operate without any host‑device transfers and emit a bounded terminal receipt. The probabilistic path supports end‑to‑end gradients via GPU knowledge compilation from provenance to CNF to Decision‑DNNF, enabling exact weighted model counting and backward gradients. Before caching or evaluation, the final smoothed circuit is certified against its source formula. Experiments show that circuit caching yields a 2.74× speed‑up for MNIST‑addition training; a worst‑case‑optimal join subsystem delivers a 27.96× geometric‑mean gain over XLOG’s binary‑join baseline. MNIST‑addition accuracy matches Scallop’s (0.9561 vs 0.9468), though no per‑epoch speed claim is made because baseline epoch time varies with CPU quota. In five hub‑skewed triangle‑counting cases, the Souffle‑to‑fused‑XLOG execution‑time ratio rises from 0.88× at 150k edges (Souffle faster) to 5.54× at 1.2M edges; fused peak device allocations are 85‑1,033 MB versus 3,287‑44,979 MB for the materializing arm. Exact inference is correctness‑equivalent to ProbLog2 but slower. On a public video benchmark, a proximity predicate trained solely via symbolic credit replaces hand‑set geometry while preserving held‑out accuracy; within Event‑Calculus rule search it fails ten‑fold cross‑validation and does not transfer on a leak‑free split. On a maritime corpus, weighted clauses beat crisp selection by 0.065 F1, a result reproduced after a single chronological training pass.
Review: XLOG demonstrates a practical pathway to fuse symbolic reasoning and neural computation on a single GPU, achieving notable speedups especially in zero‑transfer probabilistic inference and optimized join processing, and thus opens new avenues for scaling neurosymbolic workloads.