A Tree-of-Thoughts Inspired Hybrid Approach for Legal Case Judgement Summarization using LLMs

Researchers have proposed a new method for summarizing legal case judgments that combines extractive and abstractive techniques, guided by a tree-of-thoughts prompting strategy for large language models, according to a preprint published this week [1]. The approach, detailed in a paper submitted to arXiv on June 26, 2026, uses two large language models, DeepSeek and LLama, to test the hybrid summarization technique [1]. The authors note that while LLMs are increasingly applied to legal judgment summarization, most prior work has relied on either purely extractive methods, which pull key sentences directly from the text, or purely abstractive methods, which generate new phrasing [1]. Hybrid extractive-abstractive techniques, they write, "have not been explored much" in this domain [1]. The proposed method draws inspiration from tree-of-thoughts reasoning, a framework that structures a model's decision-making into branching pathways to improve complex problem-solving [1]. The paper's experiments indicate that the extractive-abstractive prompt produces better summaries than prompts designed for either extractive or abstractive summarization alone [1]. The research arrives as the legal technology sector continues to evaluate how LLMs can handle domain-specific documents where precision and fidelity to source text are critical [1]. The study does not report specific quantitative metrics in its abstract, and the full paper's evaluation benchmarks were not detailed in the preprint metadata [1]. The authors conducted comparisons across the three summarization types—extractive, abstractive, and the proposed hybrid—using the two models [1]. The findings suggest that combining extraction and abstraction within a structured reasoning framework can improve summary quality for legal judgments, a document type characterized by dense argumentation and formal citation structures [1]. The preprint has been posted on arXiv under the Computation and Language category and is available for public review [1].