Luis Martí

Recent publications

• Karmim, Y., Pino, R., Contreras, H., Lira, H., Cifuentes, S., Escoffier, S., Martí, L., Seddah, D., & Barrière, V. (2026). Leveraging Wikidata for Geographically Informed Sociocultural Bias Dataset Creation: Application to Latin America. In P. Chen, V. Zouhar, H. Hu, S. Khanuja, W. Zhu, B. Haddow, A. Birch, A. F. Aji, R. Sennrich, & S. Hooker (Eds.), Proceedings of the Workshop on Multilingual Multicultural Evaluation of the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL’2026). https://inria.hal.science/hal-05510068

  @inproceedings{karmimleveraging2026,
    title = {Leveraging Wikidata for Geographically Informed Sociocultural Bias Dataset Creation: {A}pplication to {L}atin {A}merica},
    author = {Karmim, Yannis and Pino, Renato and Contreras, Hernan and Lira, Hernan and Cifuentes, Sebastien and Escoffier, Simon and Mart\'{i}, Luis and Seddah, Djam{\'e} and Barri{\`e}re, Valentin},
    year = {2026},
    month = mar,
    booktitle = {Proceedings of the Workshop on Multilingual Multicultural Evaluation of the 19th Conference of the European Chapter of the Association for Computational Linguistics (EACL'2026)},
    address = {Rabbat, Morocco},
    url = {https://inria.hal.science/hal-05510068},
    editor = {Chen, Pinzhen and Zouhar, Vil\'{e}m and Hu, Hanxu and Khanuja, Simran and Zhu, Wenhao and Haddow, Barry and Birch, Alexandra and Aji, Alham Fikri and Sennrich, Rico and Hooker, Sara},
    hal_id = {hal-05510068},
    hal_version = {v1},
    eprint = {hal-05510068},
    eprinttype = {hal}
  }
  

• Vinayagame, V., Senay, G., & Martí, L. (2025). MATATA: Weakly Supervised End-to-End MAthematical Tool-Augmented Reasoning for Tabular Applications. Document Analysis and Recognition – ICDAR 2025: 19th International Conference, Wuhan, China, September 16–21, 2025, Proceedings, Part I, 428–467. https://doi.org/10.1007/978-3-032-04614-7_24
  Business documents often contain substantial tabular and textual information with numerical values, requiring mathematical reasoning for effective document understanding. While small language models (SLMs) still struggle at this task, tool-augmented multi-step agents perform better, at the cost of relying on closed-source or larger models, external data, or extensive prompt engineering. This work introduces MATATA, a novel weakly supervised end-to-end approach to training multi-step reasoning language agents for document tabular applications. MATATA presents an annotation-free paradigm for each agent to enhance 3.8B/8B SLMs. During its two-stage training, MATATA uses the outcome of the multi-step reasoning chain as weak supervision. This approach avoids having to individually supervise each intermediate agent in the reasoning chain. By employing an adaptive planner and shared tools across different datasets, MATATA shows robust performance. Experiments demonstrate that MATATA achieves state-of-the-art on FinQA, and on TAT-QA among reasoning methods based on open-source SLMs. Although being SLM-based, MATATA closely matches GPT-4-based frameworks on TabMWP. This novel weakly supervised approach enables training an end-to-end multi-step reasoning agent without intermediate supervision, supporting future developments of cost-effective powerful agentic systems.

  @inproceedings{10.1007/978-3-032-04614-7_24,
    author = {Vinayagame, Vishnou and Senay, Gregory and Mart\'{i}, Luis},
    title = {MATATA: {W}eakly Supervised End-to-End MAthematical Tool-Augmented Reasoning for Tabular Applications},
    year = {2025},
    isbn = {978-3-032-04613-0},
    publisher = {Springer-Verlag},
    address = {Berlin, Heidelberg},
    url = {https://doi.org/10.1007/978-3-032-04614-7_24},
    doi = {10.1007/978-3-032-04614-7_24},
    booktitle = {Document Analysis and Recognition – ICDAR 2025: 19th International Conference, Wuhan, China, September 16–21, 2025, Proceedings, Part I},
    pages = {428--467},
    keywords = {Document Understanding, Tabular Mathematical Reasoning, Small Language Models, Weak Supervision},
    location = {Wuhan, China}
  }
  

• Vinayagame, V., Senay, G., & Martí, L. (2024). Tool-Augmented Compositional Reasoning LLMs with Weak Supervision: A Scalable Approach to Reduce Human Efforts in Agent Customization. BayLearn 2024: Bay Area Machine Learning Symposimum.
  Mathematical reasoning capabilities are increasing with tool-augmented language agents, but methods often rely on proprietary models to generate trajectories for training or human efforts for prompt engineering. This work introduces a progressive refinement learning paradigm through self-annotation and weak-supervision. By updating the model’s beliefs and evolving from human inputs, the reliance on human supervision and stronger teacher models is minimized, jointly reducing the need to adapt prompting to the models.

  @inproceedings{vishnou-2024,
    author = {Vinayagame, Vishnou and Senay, Gregory and Mart\'{i}, Luis},
    booktitle = {BayLearn 2024: Bay Area Machine Learning Symposimum},
    title = {Tool-Augmented Compositional Reasoning LLMs with Weak Supervision: A Scalable Approach to Reduce Human Efforts in Agent Customization},
    year = {2024},
    location = {San Francisco, CA, USA}
  }
  

See all publications →