Published September 29, 2026 | Version v1

Two-Part MDL for Inverse Program Synthesis from Short Tables with Discrete Cell Noise

Description

We study inverse program synthesis from short numerical tables (six rows) when intermediate registers are unobserved and a fraction of cells are corrupted by discrete jumps (scale drift, dropped carry, arithmetic slip) rather than Gaussian noise. Candidate programs from a five-operator arithmetic DSL are enumerated exhaustively up to depth 2 and scored by a Two-Part MDL criterion with a frozen compositional prior and a five-channel discrete-jump likelihood. A program is accepted only if it compresses the table by 3 bits against a structureless Gaussian null. On a frozen benchmark of 5,000 tables from 20 generator families, the gate reaches table-level precision 97.2% (95% CI [96.0%, 98.1%]) and recall 45.7% ([43.5%, 47.8%]), with 0/500 false positives on white noise. Raising the enumeration depth to 3 lifts recall to 88.4% at precision 98.0%, showing that bounded recall is a search-depth ceiling. The study is synthetic and limited to the stated regime; it makes no claim of universal induction. Code, seeds and JSON artefacts are provided in the linked repository.

Files

main.pdf

Files (1.4 MB)

Name Size
md5:0a0221e157177f61cb71bd825a5ff70a
1.4 MB Preview Download

Additional details

Related works

Is supplement to
Computational notebook: https://github.com/guillaumelegras-source/mdl-inverse-program-synthesis (Other)