Paper deep dive
GLIER: Generative Legal Inference and Evidence Ranking for Legal Case Retrieval
Minghan Li, Tianrui Lv, Chao Zhang, Guodong Zhou
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 97%
Last extracted: 6/21/2026, 7:22:59 AM
Summary
GLIER (Generative Legal Inference and Evidence Ranking) is a novel framework for Legal Case Retrieval (LCR) that addresses the semantic gap between colloquial queries and formal legal documents. It reformulates retrieval as an inference process over latent legal variables (Charges and Legal Elements). The framework consists of two main stages: a Joint Generative Inference module that uses a sequence-to-sequence strategy (based on mT5) to generate structured legal indicators, and a Multi-View Evidence Fusion mechanism that ranks documents by combining generative confidence, structural overlap, and lexical signals (BM25). GLIER demonstrates superior performance and data efficiency on the LeCaRD and LeCaRDv2 benchmarks compared to baselines like SAILER and KELLER.
Entities (9)
Relation Signals (6)
GLIER â contains â Joint Generative Inference
confidence 100% ¡ GLIER decomposes the task into two interpretability-driven stages. First, a Joint Generative Inference module... Second, a Multi-View Evidence Fusion mechanism...
GLIER â contains â Multi-View Evidence Fusion
confidence 100% ¡ GLIER decomposes the task into two interpretability-driven stages. First, a Joint Generative Inference module... Second, a Multi-View Evidence Fusion mechanism...
GLIER â evaluatedon â LeCaRD
confidence 100% ¡ Extensive experiments on LeCaRD and LeCaRDv2 demonstrate that GLIER outperforms strong baselines
GLIER â evaluatedon â LeCaRDv2
confidence 100% ¡ Extensive experiments on LeCaRD and LeCaRDv2 demonstrate that GLIER outperforms strong baselines
Joint Generative Inference â generates â Legal Indicators
confidence 100% ¡ A Joint Generative Inference module that translates raw queries into latent legal indicators (Charges and Legal Elements)
ChatGLM â usedasteacherfor â GLIER
confidence 90% ¡ Specifically, we utilize ChatGLM (Du et al., 2022), a robust bilingual LLM, to extract core juridical components from each document d â D. ... to construct a high-quality supervision signal for our student model.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:The semantic gap between colloquial user queries and professional legal documents presents a fundamental challenge in Legal Case Retrieval (LCR). Existing dense retrieval methods typically treat LCR as a black-box semantic matching process, neglecting the explicit juridical logic that underpins legal relevance. To address this, we propose GLIER (Generative Legal Inference and Evidence Ranking), a framework that reformulates retrieval as an inference process over latent legal variables. GLIER decomposes the task into two interpretability-driven stages. First, a Joint Generative Inference module translates raw queries into latent legal indicators, including charges and legal elements, using a unified sequence-to-sequence strategy that jointly generates charges and elements to enforce logical consistency. Second, a Multi-View Evidence Fusion mechanism aggregates generative confidence with structural and lexical signals for precise ranking. Extensive experiments on LeCaRD and LeCaRDv2 demonstrate that GLIER outperforms strong baselines such as SAILER and KELLER. Notably, GLIER exhibits strong data efficiency, maintaining robust performance even when trained with only 10% of the data.
Tags
Links
- Source: https://arxiv.org/abs/2604.23779v1
- Canonical: https://arxiv.org/abs/2604.23779v1
Trouble viewing inline? Open PDF directly â
Full Text
49,951 characters extracted from source content.
Expand or collapse full text
GLIER: Generative Legal Inference and Evidence Ranking for Legal Case Retrieval Minghan Li * â , Tianrui Lv * , Chao Zhang, Guodong Zhou Soochow University, Suzhou, China mhli@suda.edu.cn, trlvtrlv@stu.suda.edu.cn, czhang1@stu.suda.edu.cn, gdzhou@suda.edu.cn Abstract The semantic gap between colloquial user queries and professional legal documents presents a fundamental challenge in Legal Case Retrieval (LCR). Existing dense retrieval meth- ods typically treat LCR as a black-box seman- tic matching process, neglecting the explicit juridical logic that underpins legal relevance. To address this, we propose GLIER (Gen- erative Legal Inference and Evidence Rank- ing), a framework that reformulates retrieval as an inference process over latent legal vari- ables. GLIER decomposes the task into two interpretability-driven stages: (1) A Joint Gen- erative Inference module that translates raw queries into latent legal indicators (Charges and Legal Elements), employing a unified sequence- to-sequence strategy where charges and ele- ments are generated jointly to enforce logi- cal consistency; and (2) A Multi-View Evi- dence Fusion mechanism that aggregates gen- erative confidence with structural and lexical signals for precise ranking. Extensive experi- ments on LeCaRD and LeCaRDv2 demonstrate that GLIER outperforms strong baselines like SAILER and KELLER. Notably, our frame- work exhibits exceptional data efficiency, main- taining robust performance even when trained with only 10% of the data. 1 Introduction Legal Case Retrieval (LCR) aims to identify legally relevant precedents from a large corpus given a query case (Feng et al., 2024; T.y.s.s and Hernan- dez, 2025). Unlike general ad-hoc retrieval, legal relevance is determined not by surface-level seman- tic similarity, but by whether cases share consistent juridical interpretations. In particular, relevance hinges on the alignment of Charges and their as- sociated Constitutive Elements, which encode the legal logic underlying a conviction. This makes * Equal contribution. â Corresponding author. (Query)On November 17, 2016, at 9 PM... due to a dispute over someone else's matter, he had an argument with Shao Huapeng over the phone. Shao Huapeng arranged to fight him at Zhaoh Bridge, but he did not go to the agreed location. Shao Huapeng then gathered several people and went to the "Shengli Restaurant" in Tianqiao Town... Shao Huapeng, Shao Hualai, Zhao 1, and others, armed with machetes, steel pipes, spears, and other tools, entered the Shengli Restaurant in Tianqiao Town first. The defendant Shao Yingzhu followed them in. Shao Huapeng, Shao Hualai, Zhao 1, and others smashed the counters and refrigerators in the restaurant and injured the victims Zhao 3 and Zhao 2, who were dining in the restaurant, before fleeing the scene. ...Tian Tengfei stated that more than a dozen people vandalized his "Shengli Restaurant" and injured two customers who were dining there... According to the forensic examination, the injuries sustained by victims Zhao 3 and Zhao 2 were both classified as minor injuries... Legal elements:"Joint intentional act | Armed brawl | Multiple participants | Disruption of public order | Causing minor injuries" Charge: Crime of Affray Figure 1: A colloquial query must be mapped to struc- tured legal concepts (e.g., charge and constitutive ele- ments) to retrieve legally relevant precedents. LCR challenging due to a pronounced semantic gap: queries are often colloquial factual narratives, while candidate cases are written in formal and highly structured legal language. Existing approaches to LCR mainly follow three paradigms. Lexical matching methods such as BM25 capture explicit keywords but fail to model legal reasoning. Dense retrieval models based on pre-trained language models (PLMs) improve se- mantic matching, yet struggle with long documents and implicit juridical structure. More recently, Gen- erative Retrieval (GR) (Li et al., 2023d,c; Tang et al., 2024)methods directly generate document identifiers, but suffer from limited interpretabil- ity and hallucination risks, which are particularly problematic in high-stakes legal scenarios. A com- mon limitation of these approaches is that they treat retrieval as a direct mapping from queries to documents, without explicitly modeling the legal reasoning process that mediates relevance (Deng et al., 2024a). We argue that legal case retrieval should instead 1 arXiv:2604.23779v1 [cs.IR] 26 Apr 2026 be formulated as inference over latent juridical structures. Legal experts typically begin by infer- ring legal interpretations from the facts, such as the applicable charges and their elements, and then verify these interpretations against relevant prece- dents. Motivated by this process, we reformulate LCR as a retrieval problem with structured latent variables, where legal relevance is mediated by an inferred legal interpretation rather than determined by direct textual similarity. Based on this formulation, we propose GLIER, a Generative Legal Inference framework for legal case retrieval. Instead of relying on complex multi- stage pipelines, GLIER infers a latent legal inter- pretation from the query via a unified sequence- to-sequence generation strategy. By training the model to predict the charge and its constitutive ele- ments as a single joint sequence, we leverage the autoregressive nature of the decoder to enforce log- ical consistency: the generation of legal elements is implicitly conditioned on the preceding charge pre- diction. The inferred latent structure is then used to mediate evidence-based ranking of candidate doc- uments by combining generative confidence with structural and lexical matching signals, enabling interpretable and robust retrieval. We evaluate GLIER on two benchmarks, LeCaRD (Ma et al., 2021) and LeCaRDv2 (Li et al., 2023b). Experimental results show that GLIER consistently outperforms strong baselines such as SAILER (Li et al., 2023a) and KELLER (Deng et al., 2024b) on both datasets. In par- ticular, GLIER achieves the best overall perfor- mance on LeCaRDv2 under the same experimental setting, while substantially improving recall- and hit-oriented metrics on LeCaRD. Notably, GLIER maintains strong performance even when trained with only 10% of the available data, demonstrating high robustness and data efficiency. Our contributions are summarized as follows: ⢠We formalize legal case retrieval as inference over structured latent legal variables, explic- itly modeling charges and constitutive ele- ments as pivotal mediators of relevance. â˘We propose a joint generative inference frame- work that approximates latent legal reasoning via a unified sequence-to-sequence paradigm, guaranteeing both interpretability and logical consistency. â˘We empirically validate that integrating latent inference with lightweight evidence-based ranking yields robust improvements. Notably, our model demonstrates exceptional data ef- ficiency, maintaining superior performance even when trained on only 10% of the data. 2 Related Work 2.1 Legal Case Retrieval (LCR) Traditional Legal Case Retrieval methods mainly rely on lexical matching, such as BM25, which remains highly competitive in capturing precise keywords in legal documents. However, legal cases are often extremely lengthy and contain a large amount of specialized terminology, making it dif- ficult for traditional methods to capture deep se- mantic matches. With the development of pre- trained language models (PLMs), dense retrieval- based methods have gradually become mainstream. Models such as BERT (Devlin et al., 2019) and Lawformer (Xiao et al., 2021) process long docu- ments through paragraph-level interactions or long- document attention mechanisms. SAILER (Li et al., 2023a) further introduces a structure-aware pre-training objective, enhancing representation learning by utilizing the reasoning and decision sec- tions of cases. Despite significant progress made by these discriminative models, they typically rely on truncating or segmenting long documents, which can result in the loss of the caseâs global context and key legal features, such as the logical connec- tions between charges and legal elements. 2.2 Knowledge-Guided Case Reformulation To address the noise and computational redundancy caused by long documents, recent research has be- gun to use large language models (LLMs) to re- formulate or summarize cases (Gao et al., 2024). PromptCase (Tang et al., 2023) uses LLMs to ex- tract âlegal factsâ and âlegal issuesâ from cases as key features, replacing the full text for encod- ing. Recently, KELLER (Deng et al., 2024b) fur- ther proposed a knowledge-guided reformulation method that leverages LLMs to transform com- plex case details into concise âcrime-subfactâ pairs and conducts multi-granularity contrastive learn- ing based on these subfacts. Although these meth- ods effectively extract core information (such as charges and legal elements) through LLMs, they es- sentially still fall under the discriminative retrieval (Retrieve-then-Rank) paradigm, which relies on dual encoders to compute similarity scores between 2 the Query and the Document. This approach re- quires calculating similarity scores for a vast num- ber of candidate vectors during inference and fails to directly model the generation probability from the Query to the Documentâs core features. 2.3 Generative Reasoning for Legal Retrieval Recently, Generative Retrieval (GR) has emerged as a new paradigm in information retrieval, where models like DSI (Tay et al., 2022) and NCI (Wang et al., 2023) directly generate document identi- fiers (DocIDs) to bypass the traditional index-and- retrieve pipeline. In the legal domain, models such as LegalSearchLM (Kim et al., 2025) have explored this direction by mapping queries to case IDs. How- ever, these methods often struggle with the "hal- lucination" problem and the lack of fine-grained evidence alignment, which are critical in profes- sional legal scenarios. Distinct from pure GR that aims at identifier gen- eration, our work aligns with the emerging trend of Generative Reasoning for Ranking. This direc- tion focuses on utilizing the zero-shot or few-shot reasoning capabilities of Large Language Models (LLMs) to expand queries or infer latent variables. Specifically, our framework treats generative mod- els not as an end-to-end retriever, but as a seman- tic bridge that translates colloquial queries into structured legal indicators (e.g., charges and con- stitutive elements). Unlike previous methods like PromptCase (Tang et al., 2023) or KELLER (Deng et al., 2024b) that primarily use LLMs for query reformulation, our approach explicitly models the hierarchical relationship between legal concepts and incorporates generative confidence into a dis- criminative fusion layer. This strategy combines the interpretability of generative inference with the robustness of traditional evidence-based ranking. 3 Methodology 3.1 Problem Formulation LetQdenote the set of query cases andDthe corpus of candidate documents. Given a query q â Q, Legal Case Retrieval (LCR) aims to rank documents dâD by their legal relevance to q. We model legal relevance as being mediated by a latent juridical structure rather than direct text similarity. Specifically, we introduce a structured latent variablez = (c,e), wherecdenotes a legal charge andedenotes its associated constitutive elements. We assume that the relevance between a queryqand a documentdcan be assessed through the consistency betweendand a plausible juridical interpretationzinferred fromq. Formally, we first infer the most probable latent structure Ëz = arg max z P θ (z | q),(1) and then define the relevance score as S(q,d) = f Ď (q,d, Ëz),(2) whereP θ (z | q)infers a latent legal interpretation from the query, andf Ď is a scoring function that aggregates multiple evidence signals conditioned on Ëz. Legal reasoning exhibits a logical dependency, where the admissible constitutive elements are in- trinsically constrained by the applicable charge. We capture this dependency by decomposing the latent distribution via the chain rule: P θ (z | q) = P θ (c| q)P θ (e| q,c).(3) In practice, we approximateËzvia joint genera- tive inference: we train a sequence-to-sequence model to generatecandeas a unified sequence (i.e., câ [SEP]â e), which enables the model to implic- itly learn the conditional dependencyP θ (e | q,c) through autoregressive decoding. We obtainËzus- ing constrained beam search and apply a validity filter based on a legal taxonomy to reduce halluci- nated structures. 3.2LLM-driven Legal Knowledge Distillation Legal documents often contain verbose narratives, redundant procedural details, and noise, posing sig- nificant challenges for direct dense retrieval. To mitigate this and construct a high-quality super- vision signal for our student model, we employ a Large Language Model (LLM) as an offline Knowl- edge Teacher to distill structured legal signals from the corpusD. Specifically, we utilize ChatGLM (Du et al., 2022), a robust bilingual LLM, to extract core ju- ridical components from each documentd â D. We construct a domain-specific promptPthat en- forces strict constraints to ensure the validity of the "Silver Standard" data: ⢠Terminology Enforcement: The model is restricted to extracting Constitutive Elements using professional legal terminology (e.g., "se- cretly taking property") rather than vague de- scriptive phrases. 3 Large Language Model + Extraction Prompt "id":"..." "charge":["..."] "element": ["...","..."] Generative Model A. Offline Knowledge Extraction (Data Prep) B.Generator Model Training 1. Long case query & Input Generative Model(from B) C. Inference,Retrieval & Re-ranking Pipeline Instruction Prompt: "predict charge and elements:" Validity Fliter 1. charge Confidence Score 2. Element Confidence Score 3. Charge Hit(0/1) 4. Legal Element Hit Rale(0-1) 5. Normalized BM25 (Query vs.Doc) 4. Scorer & Rerank "Query: The defendant, Qian Cuilan... operated a gambling den in a secluded spot in the mountains, using poker and dominoes for gambling and profiting from the winnings... She was apprehended by the public security authorities. "gen_charge": "Crime of operating a gambling establishment", "gen_elements": "For profit | Organizing gambling activities | Providing gambling premises | Profiting from commissions | I | ..." 2. Generated Output &Confidence "charge_conf": 0.9181, "elem_conf": 0.9181, 3. MLP Input Feature Vector Retrieval & Feature Eng. Pass Gen 5. Final Score &Ranking All documents "charge": ["Operating a casino"], "legal_elements": [ "For profit", "Organizing gambling activities", "Setting up gambling methods", "Profiting from commissions", "..."...], "hit charge": 1 "hit elem_ratio": 0.6 "norm bm25": 0.5374 "pid": "1007338" "score": 0.7866 "pid": "1351780" "score": 0.6423 ........ "pid": "573758" "score": 0.1619 Sorting& Ranking Ranked Document List & Evaluation Predicted Charges & Elements (via[SEP]token) Generation Loss Structured Ground Truth (JSON/Dict) Supervision Signal Figure 2: The overall architecture of the proposed framework, consisting of the Generative Legal Indicator Extractor (GLIE) and the Multi-Faceted Discriminative Re-ranker (MFDR). â˘Prevention of Target Leakage: Crucially, the prompt explicitly instructs the model to ex- clude sentencing outcomes (e.g., "fixed-term imprisonment", "compensation") and post- crime procedural details. This ensures that the retrieved features are based solely on the facts of the crime, preventing the model from cheating by matching sentencing patterns. For a documentdwith a grounded chargec gt , the distillation process is formulated as: K d = LLM(d,c gt ,P) = (c d ,e d )(4) wherec d âK charge denotes the applicable charges, ande d âK element represents the extracted consti- tutive elements. The output is parsed from a struc- tured JSON format. This process transforms un- structured legal texts into a clean, structured "Silver Standard" datasetD struct = (d,c d ,e d ), provid- ing explicit supervision for the subsequent student model without requiring expensive human anno- tation. The detailed prompt design is provided in A. 3.3 Generative Legal Inference (The Student Model) To equip the retriever with legal reasoning capabili- ties, we train a sequence-to-sequence model (based on mT5 (Xue et al., 2021)) to mimic the extraction process. Instead of independent classification, we propose a One-Step Joint Generation strategy to model the inherent dependencies between charges and elements. 3.3.1 Joint Generation Training We formulate the task as generating the structured tupleK q = (c q ,e q )given the query textq. The inputXis the raw query prepended with a task prompt, and the targetYconcatenates the charge and elements using a special separator token: Y = c q â [SEP]â e q (5) The model is optimized by minimizing the negative log-likelihood of the target sequence: L gen =â |Y| X t=1 logP(y t |y <t ,X;θ)(6) This joint modeling allows the decoder to leverage the predicted charge as a condition for generating subsequent legal elements, effectively preserving the logical consistency of legal reasoning. 3.3.2 Constraint-Aware Inference During inference, given a queryq, the model gener- ates a raw sequence Ë Y q . To mitigate hallucination inherent in generative models, we apply a Validity Constraint Mechanism. The raw output is parsed 4 into candidate terms(Ëc raw , Ëe raw )and filtered against the predefined taxonomyK: Ëc q =tâ Ëc raw | tâK charge , Ëe q =tâ Ëe raw | tâK element (7) This ensures that the inferred knowledge is legally valid while retaining the modelâs high-confidence predictions. 3.4 Multi-View Evidence Fusion Mechanism While the generative model captures semantic rea- soning, it lacks the calibration for fine-grained rank- ing. We propose a lightweight Multi-View Scorer that fuses signals from three perspectives: Latent Confidence, Explicit Structure, and Lexical Match- ing. For a query-document pair(q,d), we construct a feature vectorv q,d â R 5 : 1. Latent Confidence View (v 1 ,v 2 ): These fea- tures quantify the generatorâs internal certainty re- garding the inferred legal concepts. We compute the length-normalized probability for the generated charge and element sequences: v 1 = exp 1 |Ëc q | X t logP(t|Ëc <t ,q) ! , v 2 = exp 1 |Ëe q | X t logP(t|Ëe <t ,q) ! (8) A higher probability (v 1 ,v 2 â 1) indicates the model has correctly identified robust legal patterns in the query. 2. Explicit Structural View (v 3 ,v 4 ): This view measures the overlap between the queryâs inferred knowledge(Ëc q , Ëe q )and the documentâs ground truth (c d ,e d ): v 3 = I(Ëc q ⊠c d ̸=â ), v 4 = |Ëe q ⊠e d | |Ëe q | + Îľ (9) Here,v 3 is a binary indicator of charge matching (a prerequisite for legal relevance), andv 4 represents the element support ratio. 3. Lexical Matching View (v 5 ): To incorporate traditional keyword signals, we use BM25. Cru- cially, to handle score variations across queries, we apply Per-Query Normalization using the maxi- mum score within the candidate poolC q : v 5 = BM25(q,d) max d ⲠâC q BM25(q,d Ⲡ) (10) 3.4.1 Scoring and Optimization The fusion scorer is instantiated as an MLP that mapsv q,d to a relevance scoreS(q,d). To improve discriminative power, we employ a Hard Negative Mining strategy. Instead of random sampling, we select hard negativesN hard from top-ranked non- relevant documents retrieved by BM25. These doc- uments share high lexical overlap with the query but differ in legal characterization. The model is trained via Binary Cross-Entropy (BCE) loss to distinguish positive documentd + from hard nega- tives: L score =â   log S(q, d + ) + X d â âN hard log(1â S(q, d â ))   (11) This forces the model to look beyond keyword matching (v 5 ) and rely on structural evidence (v 1 ...v 4 ) to distinguish subtle legal differences. 4 Experiment In this section, we conduct comprehensive exper- iments to evaluate our proposed framework, fo- cusing on the following research questions: RQ1: How does our framework compare against state-of- the-art baselines? RQ2: What are the contributions of the hybrid scoring mechanism and different ev- identiary signals (lexical vs. generative) to the ranking performance? RQ3: How robust is the model under low-resource training settings? RQ4: Does the hierarchical joint generation strategy out- perform independent prediction? 4.1 Experimental Setup 4.1.1 Datasets and Evaluation Metrics We evaluate our method on two widely used bench- mark datasets: LeCaRD and LeCaRDv2 (Legal Case Retrieval Dataset). Following standard protocols, we consider cases with a relevance label of 3 in LeCaRD and labels of 2 and 3 in LeCaRDv2 as positive. We report a comprehensive set of metrics including MAP, P@3, R@3, R@5, Hits@3, Hits@5, and MRR@5 to evaluate both ranking quality and recall capabili- ties. 4.1.2 Baselines We compare our method with comprehensive base- lines categorized into three groups: (1) Tradi- tional Models including BM25 and TF-IDF; (2) 5 PLM-based and Embedding Methods, encom- passing general encoders (BERT, RoBERTa, BGE) and legal-specific pre-trained models (Lawformer, SAILER); and (3) Generative/Reformulation Methods represented by KELLER, a state-of-the- art approach utilizing LLMs for query augmenta- tion. All PLM-based or legal model baselines (e.g., BERT, RoBERTa, Lawformer) are fine-tuned on the respective training sets. 4.2 Performance Comparison with Baselines (RQ1) Table 1 presents the retrieval performance of our proposed framework compared to state-of- the-art baselines on the LeCaRD and LeCaRDv2 datasets. We categorize the baselines into two groups: (1) General Semantic Retrieval Mod- els, including sparse retrieval (BM25) and dense retrieval models (BERT, RoBERTa, BGE); and (2) Legal-Specific Pre-trained Models, including Lawformer, SAILER, and the previous state-of-the- art method, KELLER. From the results, we observe distinct perfor- mance patterns across the two datasets: Consistent Superiority on LeCaRDv2. On the LeCaRDv2 dataset, previous methods like KELLER have established a high performance baseline (MAP>76%), suggesting a potential ceiling effect. Despite this saturation, our method achieves state-of-the-art performance across all seven evaluation metrics. While the numerical margins are narrower due to the high baseline (e.g., improving MAP from 76.22% to 76.58% and Hits@5 from 98.71% to 99.37%), the consistency of these improvements confirms that our genera- tive paradigm successfully generalizes to diverse legal scenarios. By explicitly modeling the hier- archical structure of legal charges and elements, our framework effectively retrieves cases that pos- sess consistent juridical logic, even when lexical overlap is limited. Robustness and Safety on LeCaRD. On the LeCaRD dataset, our method demonstrates excep- tional robustness, particularly in recall-oriented metrics.Most notably, our Hits@3 reaches 95.45%, significantly outperforming the strongest baseline KELLER (83.81%) by a margin of 11.64% and SAILER (71.96%) by 23.49%. Sta- tistical tests confirm that these improvements in Hits@3, Hits@5, and R@5 are significant (p < 0.05). Furthermore, our method achieves a remark- able gain in Recall@3 (26.13% vs. KELLERâs 19.01%), demonstrating a superior ability to cover relevant precedents. It is worth noting that while KELLER achieves a higher MAP (61.81%) com- pared to ours (58.61%), our method dominates in terms of finding the correct cases (Hits) rather than just ranking them (MAP). In real-world legal prac- tice, avoiding "zero-recall" failures (where no rel- evant case is found in the top results) is often pri- oritized over precise ranking permutations. Our framework effectively mitigates this risk, ensuring a "safer" retrieval experience. Effectiveness of Legal Indicator Injection. Comparing general dense retrievers (e.g., BGE) with our method reveals a clear performance gap. General models struggle to distinguish subtle le- gal nuances. They often retrieve cases with high semantic similarity yet erroneous legal character- ization, while our framework incorporates gener- ated charges and legal elements as hard constraints to address this limitation.This confirms that incor- porating explicit legal knowledge via our Gener- ative Legal Inference module effectively bridges the semantic gap that traditional embeddings fail to capture. 4.3 Mechanism Analysis and Ablation (RQ2) To illustrate the effectiveness of our framework components and understand the underlying ranking logic, we conduct architectural ablation studies and employ SHAP (SHapley Additive exPlanations) for feature interpretability. Architecture Validity. As shown in Table 2, re- moving the MLP scorer (w/o MLP) causes a dras- tic MAP drop (-15.2%), indicating that the relation- ship between semantic correctness (e.g., Charge ac- curacy) and lexical matching (BM25) is highly non- linear. A simple rule-based sum fails to balance these distinct signals. Furthermore, bypassing the fine-tuned student model (w/o GenIR) leads to per- formance degradation (MAP 76.58%â74.78%). We attribute this to the Standardization of Legal Terminology: while the teacher LLM is power- ful, it suffers from hallucinations (e.g., generating synonymous but non-existent terms). In contrast, the student model, trained on the âSilver Standardâ data, aligns colloquial queries with the standard- ized legal vocabulary, ensuring the generated in- dicators are strictly retrievable. 6 Model LeCaRDLeCaRDv2 MAPP@3R@3R@5Hits@3Hits@5MRR@5MAPP@3R@3R@5Hits@3Hits@5MRR@5 Traditional Retrieval Models BM2549.1342.4211.4220.0772.7281.1362.4258.4366.679.3613.9292.1996.0979.10 General Pre-trained Models BERT54.5550.7915.0928.0277.2781.8266.0665.7177.609.8716.0795.3196.8890.23 RoBERTa55.8553.3315.6728.3477.5682.9165.4566.8480.2310.1216.1395.4797.2890.55 BGE57.2951.5216.9828.5577.2786.3665.6868.9881.3411.1117.4295.6098.1190.51 Legal-Specific Pre-trained Models Lawformer54.5850.7915.9526.9077.2790.9162.8070.4480.4611.0916.9096.0697.4391.80 SAILER58.2853.5118.6227.9271.9680.3767.9073.6084.3712.4417.1795.6398.5092.84 Legal-Specific Re-ranking Models KELLER61.8155.8819.0129.5283.8188.5768.2076.2285.6211.9219.5595.9498.7193.02 Ours58.6156.0626.1333.88 â 95.45 â 95.45 â 71.9776.5886.5812.7319.6297.4899.3793.52 Table 1: Retrieval performance on LeCaRD and LeCaRDv2 datasets. The best results are highlighted in bold.â indicates statistically significant improvements over the strongest baseline (KELLER) withp < 0.05. Note that on the LeCaRD dataset, our method achieves significantly higher recall and hit rates, demonstrating superior robustness compared to KELLER despite a lower MAP. (a) Global Feature Importance (Mean|SHAP|) (b) Detailed Feature Impact Distribution Figure 3: SHAP Interpretation of the MLP Scorer. (a) shows Hit_Charge is the dominant factor. (b) reveals distinct roles: Hit_Charge acts as a decisive binary filter (clear separation), while Norm_BM25 provides fine- grained calibration (continuous distribution). Interpretability of Ranking Features. To un- derstand how the MLP integrates these signals, we analyze the feature contributions in Figure 3. Hit_Charge dominates the global importance, act- ing as a decisive "gatekeeper." As seen in Fig- ure 3b(b), a charge mismatch (blue dots) signif- icantly penalizes the score, aligning with judicial logic: a case with the wrong charge is fundamen- tally irrelevant. However, Norm_BM25 ranks sec- ond, functioning as a fine-grained ranker to dis- tinguish factually similar candidates within the same charge category. This validates our Comple- MethodMAPP@3R@3R@5Hits@3Hits@5MRR@5 Ours (Full Model)76.5886.5812.7319.6296.8699.3793.52 Architecture Variants w/o GenIR (LLM+Prompt only)74.7884.1212.0019.0196.2398.1191.14 w/o MLP (Rule-based Rank)61.3872.4510.1515.5294.3496.2384.22 Table 2: Ablation study on model architecture. We compare the Full Model against variants without the finetuned generative module (using only LLM+Prompt) and without the MLP scorer (using rule-based ranking). mentary Ranking Strategy: the model relies on generative signals for logical filtering and lexical matching for factual alignment. Further detailed feature ablation (e.g., assessing the impact of re- moving Lexical features entirely) is provided in Appendix D. 4.4 Data Efficiency and Robustness (RQ3) In practical legal scenarios, obtaining high-quality labeled data is often the bottleneck. To evaluate the robustness of our framework under data scarcity, we trained our model using stratified subsets of the training set, ranging from 10% to 100%. Low-Resource Dominance.As illustrated in Fig- ure 4, our method exhibits exceptional data ef- ficiency. Notably, the performance curves for precision-oriented metrics (MAP, P@3, MRR@5) flatten rapidly, indicating saturation. Even with only 10% of the training data, our model achieves a MAP of 74.58%, which already outperforms the full-data versions of strong baselines like SAILER (73.60%) and Lawformer (70.44%). The perfor- mance gain diminishes as data increases, reaching near-optimal results (75.68% MAP) with just 30% of the data. 7 10%30%50%70%100% Training Data Ratio (%) 70 75 80 85 90 95 100 Score (MAP, P, Hits, MRR) MAP P@3 Hits@5 MRR@5 R@5 18.0 18.5 19.0 19.5 20.0 20.5 21.0 Score (R@5) Performance vs. Data Efficiency Figure 4: Performance trends on LeCaRDv2 across varying training data ratios (10%âź 100%). The model demonstrates rapid convergence, achieving near-optimal performance (e.g., Hits@5> 99%) with only30%of the data. Key metrics (left axis) remain stable, while Recall@5 (right axis) shows a slight continuous gain. Trend Analysis.While ranking metrics stabilize early, Recall@5 (purple dashed line, right axis) shows a slight but continuous improvement as data increases. This suggests that while the model quickly learns the core logic for identifying top candidates, increased data scale further helps in covering long-tail relevant cases. Detailed numeri- cal results are provided in Table 6 ( E). Why Does mT5 Learn So Fast? We attribute this rapid convergence to two primary factors: â˘Scale of LeCaRDv2: Although we use a small ratio (10%), the absolute volume of the LeCaRDv2 dataset is large enough to provide sufficient supervision signals for finetuning the pre-trained mT5 backbone. â˘High Intra-Class Homogeneity: Legal docu- ments differ significantly from general open- domain texts. Cases sharing the same charge exhibit massive repetitions in legal phrasing and logical structures. A small subset of docu- ments is sufficient for the model to capture the mapping rules between factual descriptions and charges, generalizing effectively without requiring extensive memorization of unique case details. 4.5 Impact of Hierarchical Latent Factorization (RQ4) To validate the necessity of hierarchical mod- eling, we compare our Hierarchical Genera- tion against an Independent Generation base- line (where charges and elements are predicted MethodMAPP@3R@3R@5Hits@3Hits@5MRR@5 Full Model (Hierarchical)76.5886.5812.7319.6296.8699.3793.52 Independent Generation74.7183.2311.7318.4494.9797.4892.53 Table 3: Comparison between Hierarchical (Two-Step) and Independent Generation strategies. separately). As shown in Table 3, the hierarchi- cal strategy yields consistent improvements across all metrics (e.g., +1.87% MAP). This improve- ment stems from two factors: (1) Chain-of-Logic: In legal reasoning, the Charge naturally restricts the scope of Constitutive Elements. By modeling P(Elements|Query, Charge), the charge acts as a semantic anchor, filtering out irrelevant elements (e.g., violent details in property crimes). (2) Con- textual Guidance: Although error propagation is a potential risk, results show that the charge serves as a strong prior. It resolves ambiguities in vague queries and prevents hallucinations by enforcing top-down constraints, outweighing the impact of prediction errors. 5 Conclusion In this paper, we presented GLIER, a novel frame- work designed to bridge the semantic gap in Legal Case Retrieval by mimicking the cognitive pro- cess of legal experts. By integrating an LLM- distilled generative inference module with a multi- view evidence fusion mechanism, our approach effectively aligns colloquial queries with profes- sional legal structures. Extensive experiments on the LeCaRD and LeCaRDv2 benchmarks demon- strate that GLIER achieves excellent performance, consistently outperforming strong baseline models. Furthermore, our framework exhibits exceptional data efficiency, maintaining high retrieval quality even when trained with only 10% of the available data. Through detailed feature analysis, we also verified that the model successfully leverages ju- dicial logic to improve both ranking precision and interpretability Future work will explore the ap- plication of this generative paradigm to a broader range of complex legal scenarios. Limitations Despite the strong performance of GLIER, several limitations remain to be addressed: First, the student model is initialized with mT5- base, which has a maximum sequence length con- straint (e.g., 512 tokens). While our framework uti- lizes distilled legal indicators to mitigate the noise 8 of long documents . extremely lengthy or verbose user queries may still suffer from information loss due to truncation, potentially leading to incomplete legal element inference. Second, our "Silver Standard" dataset relies on the knowledge distillation from a specific LLM (ChatGLM). Although human evaluation confirms high accuracy, the inherent biases or occasional hallucinations of the teacher model could still prop- agate to the student retriever. Finally, our experiments are primarily conducted on the LeCaRD series datasets, which are based on the Chinese legal system. The applicability of GLIERâs hierarchical structure (ChargeâEle- ments) to other jurisdictions, such as Common Law systems that rely more heavily on precedent-based reasoning than codified statutes, requires further empirical validation. Acknowledgments This work was supported by the National Natural Science Foundation of China (No. 62376178), and the Priority Academic Program Development of Jiangsu Higher Education Institutions. References Chenlong Deng, Zhicheng Dou, Yujia Zhou, Peitian Zhang, and Kelong Mao. 2024a. An element is worth a thousand words: Enhancing legal case retrieval by incorporating legal elements. In Findings of the As- sociation for Computational Linguistics: ACL 2024, pages 2354â2365, Bangkok, Thailand. Association for Computational Linguistics. Chenlong Deng, Kelong Mao, and Zhicheng Dou. 2024b. Learning interpretable legal case retrieval via knowledge-guided case reformulation. arXiv preprint arXiv:2406.19760. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers), pages 4171â4186, Minneapolis, Minnesota. Association for Computational Linguistics. Yi Feng, Chuanyi Li, and Vincent Ng. 2024. Legal case retrieval: A survey of the state of the art. In Proceedings of the 62nd Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 6472â6485, Bangkok, Thailand. Association for Computational Linguistics. Cheng Gao, Chaojun Xiao, Zhenghao Liu, Huimin Chen, Zhiyuan Liu, and Maosong Sun. 2024. En- hancing legal case retrieval via scaling high-quality synthetic query-candidate pairs. In Proceedings of the 2024 Conference on Empirical Methods in Natu- ral Language Processing, pages 7086â7100, Miami, Florida, USA. Association for Computational Lin- guistics. Chaeeun Kim, Jinu Lee, and Wonseok Hwang. 2025. Legalsearchlm:Rethinking legal case retrieval as legal elements generation.arXiv preprint arXiv:2505.23832. Haitao Li, Qingyao Ai, Jia Chen, Qian Dong, Yueyue Wu, Yiqun Liu, Chong Chen, and Qi Tian. 2023a. Sailer: structure-aware pre-trained language model for legal case retrieval. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1035â1044. Haitao Li, Yunqiu Shao, Yueyue Wu, Qingyao Ai, Yix- iao Ma, and Yiqun Liu. 2023b. Lecardv2: A large- scale chinese legal case retrieval dataset. Preprint, arXiv:2310.17609. Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wen- jie Li. 2023c. Learning to rank in generative retrieval. Preprint, arXiv:2306.15222. Yongqi Li, Nan Yang, Liang Wang, Furu Wei, and Wen- jie Li. 2023d. Multiview identifiers enhanced genera- tive retrieval. Preprint, arXiv:2305.16675. Yixiao Ma, Yunqiu Shao, Yueyue Wu, Yiqun Liu, Ruizhe Zhang, Min Zhang, and Shaoping Ma. 2021. Lecard: A legal case retrieval dataset for chinese law system. In Proceedings of the 44th Interna- tional ACM SIGIR Conference on Research and De- velopment in Information Retrieval, SIGIR â21, page 2342â2348, New York, NY, USA. Association for Computing Machinery. Yanran Tang, Ruihong Qiu, and Xue Li. 2023. Prompt- based effective input reformulation for legal case retrieval. In Australasian database conference, pages 87â100. Springer. Yanran Tang, Ruihong Qiu, Hongzhi Yin, Xue Li, and Zi Huang. 2024. Caselink: Inductive graph learning for legal case retrieval. Preprint, arXiv:2403.17780. Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. Preprint, arXiv:2202.06991. Santosh T.y.s.s and Elvin Quero Hernandez. 2025. LexKeyPlan: Planning with keyphrases and retrieval augmentation for legal text generation: A case study on European court of human rights cases. In Proceed- ings of the 63rd Annual Meeting of the Association 9 for Computational Linguistics (Volume 2: Short Pa- pers), pages 425â436, Vienna, Austria. Association for Computational Linguistics. Yujing Wang, Yingyan Hou, Haonan Wang, Ziming Miao, Shibin Wu, Hao Sun, Qi Chen, Yuqing Xia, Chengmin Chi, Guoshuai Zhao, Zheng Liu, Xing Xie, Hao Allen Sun, Weiwei Deng, Qi Zhang, and Mao Yang. 2023. A neural corpus indexer for document retrieval. Preprint, arXiv:2206.02743. Chaojun Xiao, Xueyu Hu, Zhiyuan Liu, Cunchao Tu, and Maosong Sun. 2021. Lawformer: A pre-trained language model for chinese legal long documents. AI Open, 2:79â84. Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, and Colin Raffel. 2021. mt5: A massively multilin- gual pre-trained text-to-text transformer. Preprint, arXiv:2010.11934. A Datasets ⢠LeCaRD: Derived from criminal rulings of the Supreme Peopleâs Court of China, LeCaRD consists of 107 query cases and 10,700 candidate cases. To ensure a fair comparison consistent with baselines like KELLER and SAILER, we adopted a stan- dardized evaluation protocol: the dataset was split into training and testing sets with a ratio of 0.8/0.2 using a fixed random seed of 42. Retrieval performance is evaluated by ranking the documents within the candidate pool. â˘LeCaRDv2: This dataset scales up the eval- uation with 800 query cases and 55,192 can- didate cases. It introduces a wider variety of criminal charges and more intricate legal sce- narios, serving as a comprehensive benchmark for generalization capability. B Implementation Details We implement our framework using PyTorch and HuggingFace Transformers. For the Knowledge Distillation phase, we employ ChatGLM as the teacher model. To construct the "Silver Standard" dataset, we designed a strict prompt that instructs the model to extract 4â6 key legal elements using professional terminology. Crucially, the prompt explicitly forbids the inclusion of sentencing de- tails (e.g., imprisonment terms) to prevent target leakage. The Generative Student Model is initialized with mT5-base. We set the maximum source and target sequence lengths to 512 and 128, respectively. The model was trained on a single NVIDIA Tesla V100 (32GB) GPU using the AdamW optimizer for ap- proximately 72 hours. During inference, we utilize beam search with a beam width of 3 to generate the legal indicators. The Discriminative Scorer is a 3-layer MLP (In- putâ64â32â1) with ReLU activation and Dropout (p = 0.1). It is trained using Binary Cross- Entropy loss with a batch size of 64 and a learning rate of 1e-4. To handle data imbalance, we employ a hard negative mining strategy with a negative- to-positive ratio of 3:1. Furthermore, we apply Per-Query Normalization to the BM25 scores, en- suring that lexical features are comparable across different queries regardless of their candidate pool distributions. C Prompt for Legal Element Extraction To ensure the quality of the "Silver Standard" dataset, we designed a rigorous prompt for the teacher LLM (ChatGLM). As shown in Table 4, the prompt includes specific constraints to stan- dardize terminology and, crucially, to prevent the leakage of sentencing information (which would otherwise compromise the retrieval task). System Role: You are a senior legal text analysis expert. Please extract the core "legal elements" for the given charge from the criminal case content. Input Data: ⢠Convicted Charge: charge_str ⢠Case Content: truncated_text Extraction Constraints: 1. Task Goal: Extract 4 to 6 key legal elements that support the conviction. 2. Terminology: Use professional legal terminology (e.g., "violation of transportation regulations", "causing death") rather than colloquial descriptions. 3. Anti-Leakage (Critical): Strictly Prohibit the inclu- sion of specific sentencing outcomes (e.g., "fixed-term im- prisonment", "detention", "compensation amount") or ex- plicit conviction statements. 4. Content: Do not simply repeat the charge name; ensure there is no semantic redundancy between elements. Output Format: Please directly return a standard JSON object: "legal_elements": ["Element 1", "Element 2", ...] Table 4: The instruction prompt used for knowledge distillation via ChatGLM (translated from the original Chinese). To ensure data quality, we conducted a human evaluation on 100 stratified samples. Two legal 10 graduate students assessed the LLM-extracted la- bels, yielding a Charge Accuracy of 97.0% and Element Precision of 82.0%, with a Cohenâs Kappa of 0.71 (substantial agreement). These results con- firm that the distilled "Silver Standard" data pro- vides reliable supervision signals, while the student modelâs superior performance suggests it further mitigates the remaining noise. We employed a robust LLM cascade strat- egy to construct the âSilver Standardâ candidates for LeCaRD and LeCaRDv2, primarily using chatglm-flashfor knowledge extraction. To ad- dress generation failures in complex cases, we utilizeddeepseek-R1as a fallback model to suc- cessfully process the remaining 205 documents. Subsequently, a rigorous cleaning pipeline was ap- plied to remove approximately 290 error instances (âź2.7%), filtering out data with non-unique identi- fiers, ambiguous semantic descriptions across dif- ferent charges, and inaccurate summarizations, thus ensuring high-quality supervision signals. D Detailed Feature Ablation Analysis(RQ2) To understand the contribution of different input signals to the final ranking, we conduct a compre- hensive feature ablation study. We categorize the five input dimensions of the MLP scorer into three groups: Lexical Features (BM25 score), Charge Features (Charge Confidence & Hit), and Element Features (Element Confidence & Hit Ratio). Ta- ble 5 summarizes the results on LeCaRDv2. The Role of Lexical Signals: Granularity and Factual Anchoring. A striking observation is that using Only Lexical Feature (i.e., standard BM25 ranking) achieves a MAP of 58.43%, which is notably higher than using purely legal genera- tive features (50.23%). We attribute this to the differing discriminative granularity of the sig- nals. Legal indicators (Charges and Elements) are inherently categorical: once the model identifies a specific charge (e.g., âTheftâ), all candidate cases belonging to this charge receive similarly high con- fidence scores. This results in a lack of ranking resolution, as the generative module cannot distin- guish between distinct factual contexts within the same crime category. In contrast, lexical match- ing (BM25) captures specific factual details (e.g., names, locations, object values), providing the nec- essary granularity to rank cases. Therefore, the lex- ical signal serves as the indispensable foundation for recall, preventing the âranking tiesâ that occur when relying solely on broad legal categories. Generative Signals as Semantic Gatekeepers. Despite the lower standalone performance of gen- erative features, their removal leads to catastrophic degradation in the Full Model (MAP drops from 76.58% to 50.23% when removing BM25, and to 60.19% when removing Charge features). This confirms that while generative signals may lack fine-grained ranking capability, they function as critical Semantic Gatekeepers. They impose strict juridical constraints, filtering out âHard Neg- ativesââcases that share high lexical overlap with the query but differ fundamentally in legal charac- terization (e.g., Theft vs. Embezzlement). Hierarchical Importance: Charge vs. Elements. Comparing the legal features, removing Charge Features (âw/o Chargeâ) causes a significantly larger performance drop (MAP -16.39%) than re- moving Element Features (MAP -2.98%). This validates the hierarchical nature of legal relevance modeled by our framework. The Charge acts as a coarse-grained primary filter; a mismatch here ren- ders the case irrelevant regardless of other similari- ties. The Element features serve as a fine-grained secondary verifier, helping to distinguish cases with the same charge but different constitutive require- ments, providing the final boost to reach state-of- the-art performance. Synergy of Hybrid Scoring. The most signifi- cant finding is the super-additive effect of combin- ing signals. The Full Model (76.58%) drastically outperforms both âOnly Lexicalâ (58.43%) and âOnly Legal Featuresâ (50.23%). This indicates that our MLP scorer successfully learns a Comple- mentary Ranking Strategy: it relies on BM25 to locate factually similar candidates (High Recall), while leveraging the generated Charge and Element signals to strictly enforce legal consistency (High Precision). This synergy validates our design of fusing explicit factual knowledge with latent gener- ative reasoning. E Detailed Experimental Setup for RQ3 In Section 4.4, we evaluated the data efficiency of our model. Here, we describe the sampling strategy and provide the detailed performance metrics. 11 MethodMAPMRR@5NDCG@5 Full Model (All Features)76.5893.5284.64 Impact of Feature Removal w/o Lexical Feature (BM25)50.2366.0351.69 w/o Charge Features60.1984.6272.22 w/o Element Features73.6091.5384.12 Performance of Single Feature Group Only Lexical Feature (BM25)58.4379.4265.13 Only Charge Features48.2666.5549.30 Only Element Features40.8863.9747.31 Table 5: Detailed ablation study of different features in the MLP scorer on LeCaRDv2. E.1 Stratified Sampling Strategy To ensure the statistical validity of the low-resource subsets, we did not perform simple random sam- pling. Instead, we employed Stratified Sampling based on charge categories. Given the long-tail distribution of crimes in the LeCaRDv2 corpus, random sampling might completely exclude rare charges from the training set. Therefore, for ev- ery charge type existing in the training corpus, we randomly sampled exactlyp%(e.g., 10%, 30%) of the corresponding cases. This strategy ensures that the data distribution of the subset remains consis- tent with the full dataset, preserving the diversity of legal scenarios even in extremely low-resource settings. E.2 Full Results on Data Efficiency Figure 4 in the main text illustrates the performance trends. For precise comparison, Table 6 details the exact retrieval performance metrics across all training data ratios. Ratio MAP P@3R@5 Hits@5 MRR@5 10%74.58 84.49 19.0598.6391.45 30%75.68 84.73 19.1498.7492.84 50%75.42 85.32 19.1098.7492.53 70%75.78 85.53 19.5599.1292.61 100% 76.58 86.58 19.6299.3793.52 Table 6: Performance evaluation on the LeCaRDv2 dataset with varying training data proportions. F Supplementary Experiment on Backbone Robustness While our framework utilizes distilled legal indi- cators to mitigate the noise of long documents, ex- tremely lengthy or verbose user queries may still suffer from information loss due to truncation, po- tentially leading to incomplete legal element infer- ence. To further validate that the sequence length and model capacity of mT5-base do not dominate the overall effectiveness of GLIER, we conducted supplementary experiments by replacing mT5-base with Qwen2.5-7B-Instruct under the exact same framework. Specifically, we applied QLoRA fine- tuning to the 7B model and expanded the input context window to 1024 tokens. As shown in Ta- ble 7, while the adoption of a significantly larger backbone with a longer context window does yield slight performance improvements (e.g., +0.0020 in MAP and +0.0084 in P@3), the overall gains are strictly marginal. This suggests that GLIERâs effectiveness primarily stems from its structured latent inference formulation, rather than backbone scale or context length. MetricQwen2.5-7BmT5-baseDiff MAP0.76780.7658+0.0020 P@30.87420.8658+0.0084 R@30.12490.1273-0.0024 R@50.19620.19620.0000 Hits@30.98110.9748+0.0063 Hits@50.99370.99370.0000 Table 7: Performance comparison between Qwen2.5- 7B-Instruct and mT5-base on LeCaRDv2 12