Neural optimization models are strong solvers — and nearly impossible to audit.
A team of researchers has released Evolving Programmatic Bottlenecks (EPB), a framework that takes black-box Neural Combinatorial Optimization (NCO) models — the kind used to solve routing, scheduling, and logistics problems — and distills them into portfolios of human-readable programs. The key mechanism: an LLM iteratively rewrites a bank of programs, using both numerical gradients and text-based feedback to improve how well those programs approximate the original model's decisions at each step. A second process prunes redundant programs and expands the bank when gaps are found. The resulting portfolio, the authors say, largely matches the original model's performance.
NCO has been a frustrating case for the AI interpretability field. Standard explainability tools assume static decisions with a fixed vocabulary of concepts, but NCO models make sequential, context-sensitive choices — think dynamically re-routing a delivery fleet mid-run. EPB's finding that NCO behavior shifts across optimization stages, and can be decomposed into combinations of classic heuristics, gives practitioners something they can actually reason about and stress-test before deploying a model in a high-stakes logistics or industrial setting.
Interpretability research has made real ground in vision and language models, but sequential decision-making systems have largely been left behind. EPB is a preprint, not a shipped product, so the gap between "largely matches performance" and production-ready remains uncharted — but it is the kind of foundational tooling that auditors and regulators will eventually demand.