Neural combinatorial optimization (NCO) models are fast and effective on problem structures seen during training, but real-world requirements often shift after deployment. The new paper introduces LangNCO, a framework that compiles natural-language constraints into interventions applied at inference time. This allows operators to specify new rules in plain language rather than retraining the model.
According to the abstract, the approach targets situations where operational requirements change and are communicated in natural language. By converting those requirements into targeted interventions, LangNCO aims to preserve the speed and quality of NCO while adapting to new constraints. The paper is available on arXiv.