LegalSearch-R1: Next-Gen Legal Case Retrieval

• LegalSearch-R1 is a next-generation legal retrieval system that generates legally meaningful text identifiers for retrieving case decisions.
• It leverages self-supervised fine-tuning and corpus-grounded constrained decoding via an FM-index to ensure precise, issue-sensitive matching.
• Empirical results on LEGAR BENCH show that LegalSearch-R1 outperforms traditional methods, effectively addressing both broad and fine-grained legal similarity.

Searching arXiv for the primary and related papers to ground the article with current citations. LegalSearch-R1 denotes a next-generation legal retrieval design centered on the idea that legal search should be driven by legally determinative content rather than surface overlap alone. In the formulation most directly associated with "LegalSearchLM: Rethinking Legal Case Retrieval as Legal Elements Generation" (Kim et al., 28 May 2025), the core task is Legal Case Retrieval (LCR): given a query case and a large corpus of decisions, the system returns target cases that are relevant under lawyer-defined notions of legal similarity. The defining shift is to treat retrieval as legal element generation: instead of relying only on lexical matching or fixed-size dense embeddings, the model reasons over a query case, generates legally meaningful textual identifiers grounded in target cases through constrained decoding, and maps those generated substrings back to documents. This design is paired with LEGAR BENCH, a large-scale Korean benchmark spanning 411 crime types over 1,226,814 criminal cases and a stricter factor-based view of relevance for fine-grained similarity (Kim et al., 28 May 2025).

1. Definition and problem formulation

LegalSearch-R1 is best understood as a legal retrieval architecture for similar-case search rather than citation retrieval. In the underlying task definition, a query case $q$ and a large corpus $\mathcal{D} = \{d_1, \dots, d_N\}$ are given, and retrieval aims to return a ranked subset whose top results are legally relevant under criteria defined by lawyers (Kim et al., 28 May 2025). In the standard relevance setting, relevance is tied to the same charge title and statutory provision; in stricter settings, it also depends on finer factual and legal factors that affect judgment and sentence (Kim et al., 28 May 2025).

The conceptual basis is the notion of legal elements, defined as “atomic facts that can influence the final judgment” (Kim et al., 28 May 2025). These include statutory elements, such as conditions required by a statute, and legal factors that shape liability or sentencing, such as injury severity, profit amount, victim relationship, or whether a self-defense claim was accepted (Kim et al., 28 May 2025). This framing shifts retrieval away from mere textual resemblance and toward issue- and factor-sensitive legal similarity.

This design responds to two deficiencies identified in prior LCR work. First, many existing evaluations use relatively small retrieval corpora, such as candidate pools of 100 to 55,192 cases, which do not reflect national-scale legal research (Kim et al., 28 May 2025). Second, lexical matching and dense retrieval each have domain-specific weaknesses: BM25 and related methods may treat legally critical elements and incidental details alike, while dense encoders compress long, complex legal documents into fixed-size vectors that may lose fine-grained legal distinctions (Kim et al., 28 May 2025). Related work on formulaic legal language in the CJEU similarly finds that BM25 remains a strong baseline and can outperform off-the-shelf dense models in several retrieval metrics when repetition and verbatim quotation structures are present (Mori et al., 15 Jun 2025).

2. Legal elements generation as retrieval mechanism

The defining mechanism of LegalSearch-R1 is generative retrieval over legal elements. The model type is a multilingual encoder-decoder Transformer, specifically mT5-base, used not to generate answers but to generate identifiers that function as document keys (Kim et al., 28 May 2025). The central idea is to train the model to generate legal-element-level content that appears in relevant cases and then constrain decoding so that generated sequences are exact substrings of the corpus (Kim et al., 28 May 2025).

At training time, each case is paired with legal elements extracted from that same case, yielding a self-supervised fine-tuning objective rather than relevance-labeled query-document supervision (Kim et al., 28 May 2025). For a training pair $(q, e_q)$, the loss is standard token-level cross-entropy:

$\mathcal{L}_{\text{SSFT}(q, e_q) = - \sum_{t=1}^{|e_q|} \log p_\theta(e_{q,t} \mid e_{q,<t}, q)$

The extracted elements are further transformed into first-token-aware synthetic spans so that generation begins with legally informative tokens rather than dates or locations (Kim et al., 28 May 2025). This matters because early decoding choices determine which substrings remain reachable under constrained search.

At inference time, retrieval is grounded in an FM-index over the full corpus. Constrained beam search restricts each next token to those that continue some valid corpus substring. For partial hypothesis $x_{<t}^{(i)}$, the allowed continuation set is:

$$\mathcal{C}(x_{<t}^{(i)}) = \{ x \mid \text{some document in corpus contains the prefix } x_{<t}^{(i)}x \}$$

The first token is selected from the vocabulary, and later tokens are restricted to the top-$k$ candidates within $\mathcal{C}(x_{<t}^{(i)})$ (Kim et al., 28 May 2025). Because each generated sequence is guaranteed to occur in the corpus, the system can map generated legal-element substrings back to document identifiers without hallucinated document references (Kim et al., 28 May 2025).

This suggests that LegalSearch-R1 should be viewed not as a conventional embedding retriever but as a retrieval pipeline of the form “query $\rightarrow$ legal elements $\rightarrow$ grounded substrings $\mathcal{D} = \{d_1, \dots, d_N\}$0 documents.” A plausible implication is that interpretability improves because the retrieved documents are mediated by human-readable legal-element spans rather than opaque vector similarity alone.

3. Benchmarking and relevance regimes

The benchmark most directly associated with this design is LEGAR BENCH, described as the first large-scale Korean LCR benchmark (Kim et al., 28 May 2025). It consists of two complementary relevance regimes.

LEGAR BENCH$\mathcal{D} = \{d_1, \dots, d_N\}$1 is the large-scale setting. It covers 411 crime types across 33 categories, with a retrieval pool of 1,226,814 criminal cases, of which 1,052,506 are mapped to standard groups, corresponding to 85.79% of the criminal corpus (Kim et al., 28 May 2025). Relevance is defined by membership in the same charge-title-plus-statutory-provision group, termed a “standard similar group” (Kim et al., 28 May 2025).

LEGAR BENCH$\mathcal{D} = \{d_1, \dots, d_N\}$2 evaluates fine-grained case similarity beyond the nominal offense. It starts from 160 standard groups in 8 crime categories and annotates 102 crime-specific factors with 443 options, using GPT-4o plus lawyer-designed schemas (Kim et al., 28 May 2025). The stricter benchmark uses an index of 169,230 cases, includes about 15,777 queries, and has about 14.69 targets per query (Kim et al., 28 May 2025). Relevance requires membership in the same stricter factor group, enforcing similarity in factual and legal dimensions that affect judgment and sentencing (Kim et al., 28 May 2025).

A concise summary of these two views is helpful.

Benchmark view	Core purpose	Key scale
LEGAR BENCH$\mathcal{D} = \{d_1, \dots, d_N\}$3	Broad large-scale similar-case retrieval	411 groups, 1,226,814 cases
LEGAR BENCH$\mathcal{D} = \{d_1, \dots, d_N\}$4	Fine-grained factor-based similarity	160 stricter groups, 169,230 index, ~15,777 queries

The broader significance of this benchmarking move is that it pushes legal retrieval closer to real institutional search conditions. Other recent work arrives at similar realism arguments from different directions. "Segment First, Retrieve Better: Realistic Legal Search via Rhetorical Role-Based Queries" (Nigam et al., 1 Aug 2025) argues that precedent retrieval should use rhetorically salient segments such as Facts, Issue, Arguments, and Reasoning rather than full judgments, because full-document queries overestimate what is actually available to practitioners at query time. "Legal Retrieval for Public Defenders" (Stammbach et al., 20 Jan 2026) likewise shows that realistic defender queries differ substantially from conventional legal IR benchmarks and require domain-specific corpora, query taxonomies, and evaluation.

4. Architecture, indexing, and training design

The concrete architecture underlying LegalSearch-R1 combines self-supervised legal-element generation with corpus-grounded constrained decoding (Kim et al., 28 May 2025). Training uses approximately 170K cases for LegalSearchLM fine-tuning, with up to 15 query-case–element pairs and 5 element–element pairs per case, on 8 × NVIDIA A100 80GB GPUs (Kim et al., 28 May 2025). The same 170K-case scale is used to train or adapt baselines such as Contriever and SAILER, while SAILER also receives structure-aware pretraining on 1.2M cases (Kim et al., 28 May 2025).

The retrieval layer is built on an FM-index with Burrows–Wheeler Transform, which supports substring search and prefix extension with low memory overhead on million-document corpora (Kim et al., 28 May 2025). The paper does not fully specify the downstream document scoring rule once multiple generated substrings are mapped back to cases, but it implies that several legal elements are generated per query and jointly guide retrieval (Kim et al., 28 May 2025).

This design sits within a broader shift in legal retrieval toward more structured and hybrid pipelines. "Section-Weighted Hybrid Approach for Legal Case Retrieval" (Arulanandam et al., 2 Jun 2026) proposes an alternative section-aware two-stage architecture: offline segmentation into facts, issues, decision, and reasoning; Stage 1 hybrid BM25 plus dense ANN retrieval fused with Reciprocal Rank Fusion; and Stage 2 section-level scoring with query-wise Z-score normalization and learned section weights. "Assessing the Performance Gap Between Lexical and Semantic Models for Information Retrieval With Formulaic Legal Language" (Mori et al., 15 Jun 2025) further shows that lexical and semantic methods behave differently depending on repetitiveness, quotation structure, and query length, reinforcing the need for task-specific retrieval design rather than assuming dense methods uniformly dominate.

A related misconception is that dense retrieval alone is sufficient once enough legal data are available. The results summarized for LEGAR BENCH$\mathcal{D} = \{d_1, \dots, d_N\}$5 argue against that view: fine-grained factor-sensitive relevance remains difficult for fixed-vector encoders, and BM25 remains competitive in such settings (Kim et al., 28 May 2025). Another misconception is that generative retrieval necessarily hallucinates identifiers. In this framework, constrained decoding is specifically introduced to ensure that generated sequences are exact substrings of existing corpus documents (Kim et al., 28 May 2025).

5. Empirical performance and generalization

The main reported evaluation metric is Precision@5 (Kim et al., 28 May 2025). On LEGAR BENCH$\mathcal{D} = \{d_1, \dots, d_N\}$6, the overall P@5 scores are: LegalSearchLM 0.68, BM25 0.51, Contriever 0.48, and SAILER 0.62, while KELLER as a reranked pipeline reaches 0.70 (Kim et al., 28 May 2025). This supports the claim in the abstract that LegalSearchLM outperforms baselines by 6–20% overall and reaches state-of-the-art performance (Kim et al., 28 May 2025).

On LEGAR BENCH$\mathcal{D} = \{d_1, \dots, d_N\}$7, total P@5 is 0.35 for LegalSearchLM, 0.34 for BM25, 0.01 for Contriever, and 0.22 for SAILER (Kim et al., 28 May 2025). The dense retrievers degrade sharply under stricter factor-based relevance, while BM25 retains strong performance and LegalSearchLM slightly surpasses it (Kim et al., 28 May 2025). Figure-based analysis in the paper further indicates that as the number of factors $\mathcal{D} = \{d_1, \dots, d_N\}$8 required for a match increases, SAILER’s performance drops markedly, whereas LegalSearchLM remains strongest across difficulty levels (Kim et al., 28 May 2025).

Out-of-domain experiments further support the legal-element identifier design. A variant trained only on sexual crime cases, LegalSearchLM$\mathcal{D} = \{d_1, \dots, d_N\}$9, still outperforms NaiveIdentifiers$(q, e_q)$0—a generative retrieval model trained on all crimes but supervised with random spans—by 15.66 percentage points on out-of-domain categories (Kim et al., 28 May 2025). The model trained only on sexual crime also performs close to LegalSearchLM$(q, e_q)$1, suggesting that identifier quality matters more than sheer breadth of training categories (Kim et al., 28 May 2025).

These findings resonate with other recent legal retrieval studies, though from different task formulations. On IL-PCR, "Segment First, Retrieve Better" reports that vector retrieval with Facts + Issue + Reasoning reaches MAP 0.3783 and MRR 0.3924, exceeding full-document queries and underscoring the benefit of legally salient query representations (Nigam et al., 1 Aug 2025). In public defense retrieval, existing legal benchmarks do not transfer well, but domain-specific synthetic data, query expansion with legal reasoning, and in-domain rerankers materially improve Recall@5 (Stammbach et al., 20 Jan 2026). This suggests that the central issue across systems is not merely model size, but the legal adequacy of the intermediate representation—whether legal elements, rhetorical roles, or defender-style search intent.

6. Place within legal search research and implications

LegalSearch-R1 occupies a distinct position within the broader legal IR and legal RAG landscape. It is not primarily a legal question-answering system, unlike multi-agent frameworks such as L-MARS (Wang et al., 31 Aug 2025) or LRAS (Zhou et al., 12 Jan 2026), which integrate iterative search and answer synthesis across heterogeneous sources. Nor is it a bilingual or multilingual regulatory QA system in the style of LegalRAG (Kabir et al., 19 Apr 2025) or LawPal (Panchal et al., 23 Feb 2025), where retrieval is embedded in a classic RAG pipeline. Instead, it is a case-retrieval design whose novelty lies in recasting retrieval itself as constrained generation over legal elements (Kim et al., 28 May 2025).

At the same time, it intersects with several active themes in the literature. One is realistic query modeling, emphasized by rhetorical-role retrieval (Nigam et al., 1 Aug 2025) and public-defender retrieval (Stammbach et al., 20 Jan 2026). Another is retrieval structure, as in section-aware hybrid ranking (Arulanandam et al., 2 Jun 2026). A third is evaluation granularity: legal RAG evaluation work such as ClaimRAG-LAW argues that retrieval and generation should be assessed separately at claim level because retrieval quality and answer faithfulness can diverge in legally consequential ways (Das et al., 20 May 2026). A plausible implication is that a mature LegalSearch-R1 system would benefit from claim-level or factor-level analysis in addition to document ranking metrics, even though the core paper reports only P@5.

Several limitations are explicit. The current benchmark and model are Korean-criminal-law specific, and the factor annotations in LEGAR BENCH$(q, e_q)$2 are produced with GPT-4o rather than exhaustive manual review, so noise remains possible (Kim et al., 28 May 2025). FM-indexing is static and dynamic updates are non-trivial (Kim et al., 28 May 2025). There is no explicit ranking loss; retrieval quality emerges indirectly from self-supervised legal-element generation (Kim et al., 28 May 2025). The paper also notes that while constrained decoding prevents hallucinated identifiers, presentation and explanation design for generated legal elements still require care (Kim et al., 28 May 2025).

Taken together, LegalSearch-R1 can be characterized as a legal retrieval paradigm in which legally determinative abstractions become the retrieval interface. In concrete terms, this means a multilingual seq2seq backbone, self-supervised case-to-element training, first-token-aware legal identifiers, corpus-grounded constrained decoding through an FM-index, and evaluation on both large-scale offense-level and stricter factor-level similarity (Kim et al., 28 May 2025). Within current legal search research, its significance lies in making legal similarity an explicit generative object rather than an implicit byproduct of lexical overlap or vector compression.