@inproceedings{morozov-etal-2025-fast,
    title = "Fast Thinking with Structured Prompts: Enabling {LLM} Reasoning without Chain-of-Thought Generation",
    author = "Morozov, Kirill  and
      Chubarova, Liubov  and
      Piontkovskaya, Irina",
    editor = "Angelova, Galia  and
      Kunilovskaya, Maria  and
      Escribe, Marie  and
      Mitkov, Ruslan",
    booktitle = "Proceedings of the 15th International Conference on Recent Advances in Natural Language Processing - Natural Language Processing in the Generative AI Era",
    month = sep,
    year = "2025",
    address = "Varna, Bulgaria",
    publisher = "INCOMA Ltd., Shoumen, Bulgaria",
    url = "https://preview.aclanthology.org/corrections-2026-01/2025.ranlp-1.87/",
    pages = "758--766",
    abstract = "The emergence of complex reasoning abilities in large language models (LLMs) has sparked great interest, and a variety of prompting techniques was proposed to coax them into emulating human thought processes. In this work, we introduce Think Node-by-Node, a graph-based reasoning framework inspired by mind maps, flowcharts, and other visual aids that help humans tackle complex problems. Rather than generating images directly, our approach leverages standard graph-building and rendering libraries, and requires no fine-tuning, only the model{'}s native coding capabilities. We further explore a ``Fast Thinking'' regime, in which a graph-reasoning example provided in the prompt, but the model generates the answers directly, without the full thought process reconstruction. Surprisingly, this approach leads to significant improvement upon baseline in general-knowledge tasks. Remarkably, Think Node-by-Node maintains strong performance even under a strict 25-token budget for answer generation. Across two instruction-tuned LLMs (0.5B and 7B parameters), our FastTNbN strategy outperforms baseline prompting techniques, improving accuracy by up to 10{\%}, and exceeds the capabilities of other structured prompting methods under equivalent generation constraints."
}