Work

XLOG A CUDA-Native Engine for Neurosymbolic Integration

Year: 2026
Type: article
Abstract: xlog is a CUDA-native logic programming engine that couples neural perception with deterministic Datalog, probabilistic inference, and epistemic reasoning over world views through a shared typed frontend and provider-owned CUDA runtime. The routes share device data planes but not one universal execution boundary: ordinary Datalog and exact inference are host-orchestrated, while the certified resident recursive and Monte Carlo sampled cores record zero tracked host–device transfers and then return one bounded terminal receipt. End-to-end gradients are realized through the probabilistic path. A CUDA-backed knowledge-compilation pipeline (provenance → CNF → Decision-DNNF → circuit) turns neural outputs into a differentiable circuit whose forward pass is exact weighted model counting and whose backward pass yields exact gradients. The final smoothed circuit is certified against its back-end source formula before it is cached or evaluated: bounded status and error scalars are observed by the host, while resolution-proof traces are checked on the GPU. A resident Monte Carlo alternative covers supported programs for which approximate inference is selected. Circuit caching yields a 2.74× MNIST-addition training speedup, and a worst-case-optimal join subsystem a 27.96× geometric-mean gain over xlog’s own binary-join baseline. Against external engines the picture is mixed: 2.8× faster per-epoch MNIST-addition training than Scallop at equal accuracy; on three hub-skewed triangle-counting cases, Soufflé-to-fused-xlog execution-time ratios are 2.37×, 7.95× and 9.00×, with fused peak device allocations of 85–359MB while every materializing arm completes; and exact inference is correctness-equivalent to but slower than ProbLog2. On a public video benchmark, a proximity predicate trained through symbolic credit alone, never shown a distance label, replaces hand-set geometry at no cost in held-out accuracy; the same detector inside the Event-Calculus rule search does not survive ten-fold cross-validation, and on a leak-free split the committed clause does not transfer. On a maritime corpus of thousands of gold events, weighted clauses beat crisp selection by 0.065 F1, and a single chronological training pass reproduces that result exactly.
Language: English
Open Access status: hybrid
OpenAlex ID: W7214212807
FWCI:
Cites: 32
Cited by:
Related to: 9