![[Poster]Composed-Program-Induction-with-Latent-Program-Lattice.pdf](/_next/image?url=%2Fapi%2Fmedia%2Ffile%2F%255BPoster%255DComposed-Program-Inuduction-with-Latent-Program-Lattice-900x1350.png&w=3840&q=75)
#Compositional Reasoning
#Program Induction
#Lattice Problem
Composed-Program Induction with Latent Program Lattice
This paper introduces PLAE, a model that learns a structured latent program lattice where complex programs are represented as integer combinations of primitive operations. By framing program induction as a Closest Vector Problem, PLAE enables both fast one-step reasoning and stepwise, verifiable reasoning, while also discovering refined primitives through lattice reduction.
NeurIPS-25 Workshop on Symmetry and Geometry in Neural Representations | Extended Abstract
