LLM Copilots for
Text-to-Model Translation

A comprehensive suite of LLM-based modeling copilots for text-to-model translation. Includes multiple strategies of varying complexity, from zero-shot and chain-of-thought prompting to compositional and agentic workflows, along with an online leaderboard for benchmarking and an interactive MiniZinc editor with a built-in AI assistant.

Serdar Kadıoğlu, Karthik Uppuluri, Akash Singirikonda

@misc{text2model,
  title         = {Text2Model: Modeling Copilots for Text-to-Model Translation},
  author        = {Serdar Kadıoğlu and Karthik Uppuluri and Akash Singirikonda},
  year          = {2026},
  eprint        = {2604.12955},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2604.12955}
}

Fine-tuning open-weight small-language models (SLMs) of various sizes from 0.6B to 20B parameters to generate MiniZinc models from text. We share our fine-tuning pipeline, datasets, models.

Serdar Kadıoğlu, Karthik Uppuluri

@misc{text2model,
  title         = {Text2Model: Modeling Copilots for Text-to-Model Translation},
  author        = {Serdar Kadıoğlu and Karthik Uppuluri and Akash Singirikonda},
  year          = {2026},
  eprint        = {2604.12955},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2604.12955}
}

The first cross-domain dataset to integrate both optimization and satisfaction problems within a unified data schema. Leverages MiniZinc's paradigm- and solver-agnostic modeling capabilities. Accompanied by an online interactive data editor to enable ongoing data curation.

Akash Singirikonda, Serdar Kadıoğlu, Karthik Uppuluri

@misc{text2zinc,
  title         = {Text2Zinc: A Cross-Domain Dataset for Modeling Optimization and Satisfaction Problems in MiniZinc},
  author        = {Akash Singirikonda and Serdar Kadıoğlu and Karthik Uppuluri},
  year          = {2025},
  eprint        = {2503.10642},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2503.10642}
}

Global LLM Agents for text-to-model translation. Introduces collaborative multi-agent workflows where specialized LLM agents work together to decompose, formulate, and validate models. Achieves strong performance through agent collaboration.

Junyang Cai, Serdar Kadıoğlu, Bistra Dilkina

@misc{gala,
  title         = {Gala: Global LLM Agents for Text-to-Model Translation},
  author        = {Junyang Cai and Serdar Kadıoğlu and Bistra Dilkina},
  year          = {2025},
  eprint        = {2509.08970},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI},
  url           = {https://arxiv.org/abs/2509.08970}
}

Named entity recognition for optimization modeling from natural language. Extracts optimization-related entities such as objectives, decision variables, constraints, and parameters from free-form problem descriptions. Combines classical NLP techniques with transformers architecture fine-tuned on optimization corpora and domain-specific data augmentation.

Serdar Kadıoğlu, Parag Pravin Dakle, Karthik Uppuluri, Regina Politi, Preethi Raghavan, SaiKrishna Rallabandi, Ravisutha Srinivasamurthy

@article{ner4opt,
  author    = {Serdar Kadıoğlu and Parag Pravin Dakle and Karthik Uppuluri and Regina Politi and Preethi Raghavan and SaiKrishna Rallabandi and Ravisutha Srinivasamurthy},
  title     = {Ner4Opt: named entity recognition for optimization modelling from natural language},
  journal   = {Constraints},
  volume    = {29},
  number    = {3-4},
  pages     = {261--299},
  year      = {2024},
  url       = {https://doi.org/10.1007/s10601-024-09376-5},
  doi       = {10.1007/S10601-024-09376-5}
}

A foundational perspective on translating natural language into constraint models, laying out a blueprint for text-to-model systems and motivating structured, model-aware copilot workflows for formulation, validation, and reasoning.

Dimos Tsouros, Hélène Verhaeghe, Serdar Kadıoğlu, Tias Guns

@misc{tsouros2023holygrail20natural,
      title={Holy Grail 2.0: From Natural Language to Constraint Models},
      author={Dimos Tsouros and Hélène Verhaeghe and Serdar Kadıoğlu and Tias Guns},
      year={2023},
      eprint={2308.01589},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2308.01589},
}

Pruning large-language and vision models to reduce inference and storage cost while supporting interactivity workload for dialogue agents.

Gili Rosenberg, Kyle Brubaker, Martin Schuetz, Elton Zhu, Serdar Kadıoğlu

@misc{icbs,
      title={Scalable iterative pruning of large language and vision models using block coordinate descent},
      author={Gili Rosenberg and J. Kyle Brubaker and Martin J. A. Schuetz and Elton Yechao Zhu and Serdar Kadıoğlu and Sima E. Borujeni and Helmut G. Katzgraber},
      year={2024},
      eprint={2411.17796},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2411.17796},
}