LegalSearchLM: Rethinking Legal Case Retrieval as Legal Elements Generation

TL;DR

This paper tackles limitations in legal case retrieval (LCR) by introducing LEGAR BENCH, the first large-scale Korean LCR benchmark spanning 411 crime types over 1.2M candidate cases, and by reframing retrieval as generation of legal elements with LegalSearchLM. LegalSearchLM uses constrained decoding grounded by an FM-index to generate key legal elements from a query and retrieve target cases, addressing information loss from fixed embeddings and noise from purely lexical matching. Experiments show state-of-the-art performance on LEGAR BENCH (Standard and Stricter) and strong generalization to unseen crime types, outperforming naive in-domain generative baselines by about 15%. Together, LEGAR BENCH and LegalSearchLM offer a scalable, domain-specific resource and method that advance practical LCR for legal professionals and support robust, element-level reasoning in retrieval.

Abstract

Legal Case Retrieval (LCR), which retrieves relevant cases from a query case, is a fundamental task for legal professionals in research and decision-making. However, existing studies on LCR face two major limitations. First, they are evaluated on relatively small-scale retrieval corpora (e.g., 100-55K cases) and use a narrow range of criminal query types, which cannot sufficiently reflect the complexity of real-world legal retrieval scenarios. Second, their reliance on embedding-based or lexical matching methods often results in limited representations and legally irrelevant matches. To address these issues, we present: (1) LEGAR BENCH, the first large-scale Korean LCR benchmark, covering 411 diverse crime types in queries over 1.2M candidate cases; and (2) LegalSearchLM, a retrieval model that performs legal element reasoning over the query case and directly generates content containing those elements, grounded in the target cases through constrained decoding. Experimental results show that LegalSearchLM outperforms baselines by 6-20% on LEGAR BENCH, achieving state-of-the-art performance. It also demonstrates strong generalization to out-of-domain cases, outperforming naive generative models trained on in-domain data by 15%.

Figures (6)

• Figure 1: Examples of Relevance Cases. (Query) is a query case on distributing false images/videos for profit. The Blue Highlight indicates profit, the Yellow Highlight represents the creation of false images/videos, and the Green Highlight denotes distribution—the three key legal elements of the crime. Both (Standard) and (Stricter) satisfy the three elements, and (Stricter) additionally meets the requirements concerning the scale of distributed images/videos (Red Highlight) and the total financial gains obtained (Purple Highlight). (A) and (B) are not target cases, as (A) distributed a false image without intending to obtain financial gains, and (B) committed the offense for financial gain through the unlawful filming of real footage, not the creation of false images.
• Figure 2: Examples of the construction process for each step in LEGAR BENCH$_{\textit{Standard}}$. Step 1 defines major crime categories based on Korean Criminal Act. Step 2 refines these categories using charge titles, and Step 3 further specifies them based on statutory provisions.
• Figure 3: Examples of the construction process of LEGAR BENCH$_{\textit{Stricter}}$. Each crime type (e.g., Insult) includes specifically defined factors (in boxes filled with sky blue) and sub-factors (in boxes outlined in black). Cases are annotated by mapping all sub-factors to corresponding predefined options.
• Figure 4: Inference process of LegalSearchLM. Given a Query doc as input, LegalSearchLM generates key legal elements expected to appear in the Target doc via core-first-token-aware constrained decoding over a prefix-indexed corpus (Generated content). Since the generated content is grounded in the corpus, it can be linked back to its source document, enabling retrieval.
• Figure 5: Performance on LEGAR BENCH$_{\textit{Stricter}}$ by four different difficulties, where $N$ represents the number of factors that should be matched. LegalSearchLM achieves the best performance across all difficulty levels, demonstrating robustness in complex retrieval settings.