The Shortest Vector Problem (SVP) is a foundational computational challenge in lattice theory and a core basis for Post-Quantum Cryptography (PQC) security. With Artificial Intelligence (AI) advancement, researchers have started exploring new approaches to solving SVP via Machine Learning (ML) and Deep Learning (DL). Though promising, the emerging field of “AI for SVP” (AI4SVP) faces challenges in its early stages, especially the absence of generic methodological taxonomy and high-quality benchmark datasets, with inconsistencies in experimental setup and evaluation protocols. To address this, we introduce the first comprehensive benchmarking framework for AI4SVP. We systematically categorize 20 SVP solvers into three classical paradigms: sieving, enumeration, and lattice reduction. Building on these, we design AI4SVP-Bench (this repo), a modular framework with three AI-Augmented task interfaces: AI4Enum, AI4BKZ, and AI4Sieve. Evaluated on lattice instances spanning 80 dimensions, our AI components show significant gains: AI4Enum reduces node visits by 33.6% and time by 28.6% on 60-dimensional instances; AI4BKZ cuts SVP oracle calls by 25.3% and runtime by 58.4%; AI4Sieve reduces list operations by 18.0%. Further, We conduct extensive hyperparameter optimization on existing SVP solvers across multiple test scenarios in different dimensions, empirically verifying whether such observed optimal parameterization exhibits instance-agnostic universality. Overall, our study demonstrates empirical feasibility, draws desired principles, and aims to promote coordinated innovations for future development in the novel interdisciplinary AI4SVP realm.
Note: The paper is currently under single-blind review for KDD'26 AI for Sciences Track.