FORGE: Foundational Optimization Representations from Graph Embeddings

A research team has introduced Forge, a framework that pre-trains a vector-quantized graph autoencoder on a large collection of mixed-integer programming instances to create general-purpose representations for combinatorial optimization problems, according to a paper posted on arXiv [1][2]. The framework, detailed in a submission last revised on 16 June 2026, aims to address a core limitation of current learning-based optimization methods: the need to train dedicated models for each specific problem distribution and downstream task [1][2]. The authors, including Zohair Shafi, argue this requirement severely restricts scalability and generalization [1]. Forge instead learns from a diverse set of mixed-integer programming (MIP) instances in an unsupervised manner, without relying on optimization solvers or optimal solutions [2]. The system uses vector quantization to produce discrete code assignments that function as a vocabulary for representing optimization instances [2]. In an unsupervised setting, the resulting Forge embeddings effectively cluster unseen instances across different problem domains and sizes [1][2]. When fine-tuned in a supervised setting, a single pre-trained Forge model assists in predicting the integrality gap for cut-generation and variable hints for search guidance across multiple problem and size distributions [1][2]. The paper reports that this approach improved the performance of a commercial optimization solver and outperformed state-of-the-art learning-based methods in both tasks [2]. The authors have open-sourced the training code, pre-trained Forge weights, and embeddings for multiple MIP distributions to support further research [2]. The paper was posted on arXiv, an open-access repository for electronic preprints that, as of November 2024, receives about 24,000 submissions per month and hosts over two million articles [8]. The work is associated with several arXivLabs tools, including the Bibliographic Explorer and Connected Papers, which are community-developed features that provide citation navigation and literature mapping directly on the article's abstract page [6][7]. The submission history shows the paper was first uploaded on 28 August 2025 and underwent five revisions, with the file size growing from 3,149 KB to 5,657 KB in the final version [1]. The research falls within the broader effort to apply machine learning techniques, such as large language models trained with self-supervised learning on vast datasets, to structured problem domains [10].