Jumyung Park

Jumyung Park, Jiwon Park, Jinseo Shim, Sejin Kim, Paulina Vennemann, Sundong Kim. Composed-Program Induction with Latent Program Lattice.
NeurIPS 2025 Workshop on Symmetry and Geometry in Neural Representations.

WebsiteVideoPoster

Research Intern; Advised by Prof. Sundong Kim

• Led the development of Program Lattice Auto Encoder (PLAE), introducing a latent program space that models discrete program composition as an integer lattice in continuous embedding space.
• Formulated compositional program induction as a Closest Vector Problem (CVP) over a learned latent program lattice, enforcing structured geometric constraints on neural program representations.
• Designed a dual-mode inference procedure: fast one-step lattice projection (System-1) and stepwise lattice walks with intermediate decoding and re-encoding for compositional verification (System-2).
• Proposed lattice reduction (LLL) as a principled abstraction mechanism to refine primitive program bases while preserving the underlying lattice structure.

Program Induction, Structured Representation Learning, Geometric Deep Learning