Neural Architectures for Resolving References in Program Code via Permutation Learning

arXiv·low signalXBlueskyLinkedInCopy link

Researchers abstract code reference resolution — a fundamental task in compilers, decompilers, and code intelligence tools — into permutation indexing problems, then benchmark neural architectures on synthetic tasks designed to isolate this capability. Motivated by real-world decompilation, the work identifies which architectures (transformers, RNNs, pointer networks) handle direct vs. indirect reference rewriting most effectively. Relevant to builders of code intelligence tools and decompilation pipelines seeking to understand which neural components handle symbolic reference tracking.