一句话：首次系统刻画神经符号概率组合式AI在CPU/GPU上的性能瓶颈并提出软硬件优化。
针对LLM可持续计算、鲁棒性与可解释性不足，组合式AI将LLM与符号推理、概率推断结合，但其在现有硬件上的系统影响尚不清楚。本文建立神经符号概率模型分类法并分析扩展效率，在CPU和GPU上表征运行时、内存、计算算子、数据流、可扩展性、通信与系统行为。研究发现，组合式模型受异构计算内核、非规则访存、复杂流控、数据依赖和高数据移动制约，在通用硬件上效率低下。基于这些观察，作者提出若干软件与系统级优化，展示提升性能、效率与可扩展性的潜力，并从系统与体系结构视角首次整体评估组合式神经符号AI。
Large Language Models (LLMs) have driven remarkable progress in artificial intelligence (AI), but their rapid growth faces challenges of unsustainable computation, limited robustness, and poor explainability. Compositional AI, which integrates LLMs with symbolic reasoning and probabilistic inference, has emerged as a promising paradigm to enable interpretability, robustness, trustworthiness, and data-efficient learning. Recent neuro-symbolic-probabilistic systems demonstrate strong potential in agentic applications, advancing reasoning and cognitive capabilities toward human-like intelligence. In this paper, we aim to understand the system implications of neuro-symbolic-probabilistic architectures. We first develop a taxonomy and analyze the scaling efficiency of neuro-symbolic-probabilistic models. We then perform a detailed characterization across CPUs and GPUs, evaluating their runtime, memory, compute operators, dataflow, scalability, communication, and system behaviors. Our studies reveal that compositional models suffer from inefficiencies on off-the-shelf hardware, constrained by heterogeneous compute kernels, irregular memory access, complex flow control, data dependencies, and high data movement. Guided by these insights, we present several software- and system-level optimizations to demonstrate the potential for improving the performance, efficiency, and scalability of neuro-symbolic-probabilistic computing. To our knowledge, this is among the first works to holistically assess compositional neuro-symbolic AI from system and architectural perspectives.
@inproceedings{WanYQRPWRK26,
title = {Compositional AI Beyond LLMs: System Implications of Neuro-Symbolic-Probabilistic Architectures},
author = {Zishen Wan and Hanchen Yang and Jiayi Qian and Ritik Raj and Joongun Park and Chenyu Wang and Arijit Raychowdhury and Tushar Krishna},
booktitle = {ASPLOS},
year = {2025},
doi = {10.1145/3760250.3762235}
}