This repository tracks papers and resources about large language models (LLMs) in the legal domain.
Daily update 2026-11-04 (cs.CL): scanned all 85 papers on the dated list; added 2 title-deduplicated papers: PROSE and Say Less, Mean Better. Category increments: Applications +0, Legal Reasoning Models +1, Legal Agent +0, Legal Problems +1, Data Resources +1, Law LLMs +0, Evaluation +2. PROSE studies preservation-aware repair of false-guilty predictions. MinimalLM is included only under Evaluation for its explicitly tested civil-versus-criminal-law question; this is one illustrative query, not a legal benchmark or evidence of legal correctness. Generic privacy audits, organizational rules, and non-legal uses of contracts or judges were excluded. The archive date follows the dated cs.CL list, not the original submission date.
Daily update 2026-11-03 (cs.CL): scanned all 251 papers on the dated list; added 6 title-deduplicated papers on time-aware Thai legal QA, criminal-verdict extraction, ICJ citation resources, outdated Norwegian statutory values, Tajik legal retrieval, and fabricated citations with a US federal case-law subset. Category increments: Applications +2, Legal Reasoning Models +0, Legal Agent +0, Legal Problems +2, Data Resources +2, Law LLMs +0, Evaluation +6. The Tajik embedder is not law-specific, and the citation study's 900 questions span three domains. HIPAA-motivated access control was excluded because the paper explicitly studies an institutional rule rather than a legal test; generic policy, religious-food rules, and plagiarism detection were also excluded. The archive date follows the dated cs.CL list, not the original submission date.
Daily update 2026-11-02 (cs.CL): scanned all 86 papers on the dated list; added 1 title-deduplicated paper, BILAL-BENCH, for its explicit 45-item Hanafi fiqh-ruling subset. Category increments: Applications +0, Legal Reasoning Models +0, Legal Agent +0, Legal Problems +0, Data Resources +1, Law LLMs +0, Evaluation +1. The 240-item collection is not exclusively legal, and source-fabrication scores are not national-law or court-decision accuracy. Generic privacy, fairness, and copyright/IP motivations without a concrete legal-task evaluation were excluded. The archive date follows the dated cs.CL list, not the original submission date.
Daily update 2026-10-30 (cs.CL): scanned all 175 papers on the dated list; added 6 title-deduplicated papers on Brazilian statutory-text tasks, Indian IPR question answering, agent-assisted bar-exam learning, Japanese legal modeling, advice under human-rights/legal constraints, and Arabic statutory retrieval. Category increments: Applications +3, Legal Reasoning Models +0, Legal Agent +1, Legal Problems +1, Data Resources +3, Law LLMs +1, Evaluation +6. HEARSAYBENCH includes concrete legal-recourse constraints, not just generic Constitutional AI; its full collection is not exclusively legal. Generic safety, organizational policies, political opinions, and mathematical uses of “law” were excluded. The archive date follows the dated cs.CL list, not the original submission date.
Daily update 2026-10-27 (cs.CL): scanned all 232 papers on the dated list; added 2 title-deduplicated papers: PatEx-Gym and Can Language Models Learn While They Think? Category increments: Applications +1, Legal Reasoning Models +0, Legal Agent +1, Legal Problems +0, Data Resources +1, Law LLMs +1, Evaluation +2. The first addresses patent novelty examination; the second explicitly evaluates SCOTUS and ContractNLI and is included only under Evaluation. Generic policy ambiguity, privacy safeguards without a stated legal task, software contracts, and non-legal uses of “law” or “judge” were excluded. The archive date follows the dated cs.CL list, not the original submission date.
Daily update 2026-10-26 (cs.CL): scanned all 106 papers on the dated list; added 1 title-deduplicated paper, AusClinBench, with an explicit National Law obligations subset. Category increments: Applications +0, Legal Reasoning Models +0, Legal Agent +0, Legal Problems +0, Data Resources +1, Law LLMs +0, Evaluation +1. The appendix tests health-service advertising permission and a statutory penalty; inclusion is not based on generic clinical policy alone. Answer keys lack independent review, and benchmark scores do not certify legal correctness or clinical-product safety. Generic compliance, incidental copyright notices, software contracts, and non-legal uses of “law” and “judgment” were excluded.
Daily update 2026-10-23 (cs.CL): scanned all 117 papers on the dated list; added 2 title-deduplicated papers on French civil-cassation grounds and the evidence behind RAG copyright-protection claims. Category increments: Applications +0, Legal Reasoning Models +0, Legal Agent +0, Legal Problems +1, Data Resources +2, Law LLMs +0, Evaluation +2. The French resource uses deterministic extraction and language-model baselines, not a new legal LLM. The RAG survey explicitly audits a copyright/ownership subset; its technical endpoints do not determine infringement. Generic privacy, banking procedures, civil-rights history, and document-editing ethics without an explicit legal task were excluded.
Daily update 2026-10-22 (cs.CL): scanned all 137 papers on the dated list; added 6 title-deduplicated papers on English–Mizo judgment translation, dated Norwegian statutory values, legal-factor representations, missing-authority abstention, police-interrogation summarization, and framing effects in legal/consumer-rights advice. Category increments: Applications +3, Legal Reasoning Models +1, Legal Agent +0, Legal Problems +2, Data Resources +3, Law LLMs +0, Evaluation +6. The summarization paper is admitted under the explicit dataset/benchmark rule, not presented as a new legal LLM. Internal police operating-policy conflicts, professional ethics alone, and software contracts were excluded under the strict legal-task gate; readout accuracy, ROUGE, and advice consistency do not establish legal correctness.
Daily update 2026-10-21 (cs.CL): scanned all 159 papers on the dated list; added 1 title-deduplicated paper on exact legal-clause extraction from contracts. Category increments: Applications +1, Legal Reasoning Models +0, Legal Agent +0, Legal Problems +0, Data Resources +1, Law LLMs +0, Evaluation +1. Generic privacy, political-discourse analysis, fraud-risk screening without a legal task, and non-legal uses of “contract,” “policy,” or “judgment” were excluded. Exact copying is not legal interpretation or a correctness guarantee.
Daily update 2026-10-20 (cs.CL): scanned all 283 papers on the dated list; added 6 title-deduplicated papers on legal instruction-data selection, real-world legal-guidance queries, legal reranking, Dutch insolvency reasoning, court-document routing, and copyright guardrail evaluation. Category increments: Applications +5, Legal Reasoning Models +1, Legal Agent +0, Legal Problems +1, Data Resources +1, Law LLMs +0, Evaluation +6. Copyright reproduction is explicitly tested, not merely a motivation. Classifier agreement, LLM preferences, and copyright benchmark labels do not establish legal correctness; generic compliance, cultural constitutions, police keyword statistics, and non-legal uses of “law” or “license” were excluded.
Daily update 2026-10-19 (cs.CL): scanned all 146 papers on the dated list; added 3 title-deduplicated papers on Italian legislative-process extraction, Chinese probabilistic decisions with a tested legal subset, and Korean court-judgment de-identification. Category increments: Applications +3, Legal Reasoning Models +0, Legal Agent +0, Legal Problems +0, Data Resources +1, Law LLMs +0, Evaluation +2. The de-identification system uses pretrained encoder language models, not generative legal LLMs. Generic PII protection, financial advice, software contracts and incidental GDPR references were excluded.
Daily update 2026-10-16 (cs.CL): scanned all 177 papers on the dated list; added 3 title-deduplicated papers on real-case criminal-assistance evaluation, misalignment after bad-legal-advice finetuning, and medical-device regulatory document assessment. Category increments: Applications +1, Legal Reasoning Models +0, Legal Agent +0, Legal Problems +2, Data Resources +1, Law LLMs +0, Evaluation +3. Crime-taxonomy labels are not court judgments; ContextLift does not repair the learned bad legal advice, and deterministic regulatory assessments are not necessarily correct. Generic privacy, procedural compliance, dataset-license statements and non-legal uses of “contract” were excluded.
Daily update 2026-10-14 (cs.CL): scanned all 385 papers on the dated list; added 7 title-deduplicated papers on evidence-based legal agents, legal reward evaluation, patent translation, human-rights alignment, PFAS regulatory extraction, and constitutional translation. Category increments: Applications +4, Legal Reasoning Models +1, Legal Agent +1, Legal Problems +1, Data Resources +5, Law LLMs +1, Evaluation +5. Dataset/benchmark and translation-model papers are included under the explicit admission rule; translation scores and alignment labels do not certify legal validity. Generic compliance and copyright motivation alone were excluded.
Daily update 2026-10-12 (cs.CL): scanned all 116 papers on the dated list; added 5 title-deduplicated papers covering legal pre-training, statutory retrieval, anti-money-laundering evaluation, legal-agent workspace safety, and patent prior-art retrieval. Category increments: Applications +2, Legal Reasoning Models +1, Legal Agent +1, Legal Problems +1, Data Resources +4, Law LLMs +1, Evaluation +5. Encoder models and benchmark resources are included under the explicit language-model/dataset/benchmark rule; detection labels, retrieval proxies, and security scores are not legal-correctness guarantees.
Daily update 2026-10-08 (cs.CL): scanned all 118 papers on the dated list; added 2 title-deduplicated papers on insurance-contract claim adjudication and jurisdiction-specific Medicaid eligibility QA. Category increments: Applications +0, Legal Reasoning Models +0, Legal Agent +0, Legal Problems +0, Data Resources +1, Law LLMs +0, Evaluation +2. ExperienceIndex was excluded because court cases are only a motivating example, not a tested legal task. Synthetic-case and closed-book-evaluation limits are noted below.
Daily update 2026-10-07 (cs.CL): scanned all 147 papers on the dated list; added 3 title-deduplicated papers on contract evidence recovery, legal document automation, and training-data licensing. Category increments: Applications +2, Legal Reasoning Models +1, Legal Agent +0, Legal Problems +1, Data Resources +0, Law LLMs +0, Evaluation +2. Generic compliance references alone were excluded; the licensing paper is included for its article-level reuse-permission audit, not merely its open-source license. Evidence limitations are noted below.
Daily update 2026-10-06 (cs.CL): scanned all 243 papers on the dated list; added 5 title-deduplicated papers with explicit legal tasks and language-model/agent relevance. Category increments: Applications +2, Legal Reasoning Models +1, Legal Agent +1, Legal Problems +2, Data Resources +1, Law LLMs +1, Evaluation +5. Penumbra was excluded: its authored behavioral constitutions do not establish a concrete legal task. Important limitations are noted with the entries below.
Backfill update 2026-10-05 (cs.CL): scanned 1,072 papers from daily lists dated 2026-09-27 through 2026-10-04; added 18 title-deduplicated papers meeting the strict legal-task plus model/agent/RAG/dataset/benchmark rule. Empty daily lists were verified, not treated as network failures.
Category increments — Applications +7, Legal Reasoning Models +3, Legal Agent +3, Legal Problems +3, Data Resources +7, Law LLMs +1, Evaluation +15.
- Applications of Large Language Models in Legal Tasks
- Legal Reasoning Models
- Legal Agent
- Legal Problems of Large Language Models
- Data Resources for Large Language Models in Law
- Law LLMs
- Evaluation
-
Unlocking Criminal Court Decisions: Flexible Information Extraction for Empirical Legal Research paper — Uses expert-refinable instructions for attribute exploration, span extraction, and value normalization in criminal verdicts involving DUI and driving despite a ban; repeatability is not extraction correctness.
-
Saidzoda-Tajik-E5-Large: Fine-Tuning a 560M Embedder for Tajik Retrieval paper — Adapts multilingual-e5-large-instruct with language-model-generated queries and explicitly tests retrieval from unseen Tajik laws; the model also covers general-domain text and is not a law-specific LLM.
-
AI Cosmos: An Agentic Infra for Situated Learning paper — Tests scenario-based agent interaction for civil/criminal-law bar-exam learning; student learning gains are not evidence of autonomous legal-practice competence.
-
XLaw-JP-27B: A Lightweight Open Model Specialized for Japanese Law paper — Combines statutory/case retrieval, light self-distillation, a contract adapter, and voting for Japanese legal tasks.
-
Statutory Article Retrieval from Arabic Judicial Narratives: Family Law and Law of Evidence in Saudi Arabia paper — Decompose-and-Refine turns case narratives into legal issues and retrieval channels; evidence is limited to synthetic cases and two Saudi statutes.
-
Novel or Not? PatEx-Gym: Agentic Prior-Art Search for Novelty Assessment in Patent Examination paper — Connects prior-art discovery with claim-level novelty assessment and Office Action generation; benchmark performance is not a patentability determination.
-
Towards English - Mizo Legal Machine Translation: Corpus Creation, Analysis, and Benchmarks paper — Adapts NLLB-200-1.3B for Indian court-judgment translation; the reported 27.47 BLEU does not establish legal equivalence, and English borrowing remains above human levels.
-
Commit or Abstain? System One Decision Models, LLMs, and Missing Legal Authority paper — Combines typed decision models with generative LLMs to check candidate holdings and abstain when authority is missing; agreeing filters reduce wrong selections on CaseHOLD, not all forms of invalid legal authority.
-
Multi-level Encoder with Global Topic for Task-oriented Dialogue Summarization paper — Graph-based MLEGT summarizes Chinese police-interrogation records using global-topic guidance and a pointer decoder; it is a specialized summarizer, not a new legal foundation model.
-
When Copying Is Hard: Copy-Constrained Decoding for Exact Span Reproduction paper — CopyGen controls copying and stopping for verbatim legal-clause extraction; the backbone is frozen but lightweight plugin heads are trained.
-
Hidden Cartographer: Hidden State Guided Data Selection for Dual Budgeted Instruction Tuning paper — Budgeted instruction-data selection is tested on legal examination, legal event detection, and case classification from DISC-Law-SFT; the legal setting annotates 20% of the pool and trains on 10%, not a new general legal-reasoning model.
-
Legal Guidance in the Wild: How Users Seek Legal Help in Real-World LLM Conversations paper — Classifies legal-guidance intent and domains in WildChat conversations using a multi-axis taxonomy; the enriched sample does not estimate legal-help prevalence among all users.
-
Beyond Recall Saturation: Scaling Hard-Negative Reranker Training for Industrial Legal Retrieval paper — Adapts Qwen3 rerankers with embedding-mined then self-mined negatives for Chinese legal retrieval; gains depend on model scale and benchmark, with no claim of case-level or temporal isolation.
-
Machina Pauliana: A neuro-symbolic approach to the actio pauliana in Dutch insolvency law paper — Combines an expert-written Catala model of Dutch insolvency provisions with LLM factual extraction and retrieval-grounded resolution of open-textured terms; ten evaluation cases do not establish broad judicial reliability.
-
Narrow Decision Models: A System-1 Architecture for Calibrated, Typed Decisions paper — A block-attention encoder supports typed decisions and court-communication routing; 91% agreement on held-out legal documents is agreement with a production LLM classifier, not human-labelled accuracy.
-
The Italian lawmaking process: a unified and structured dataset paper — LLM extraction converts Italian parliamentary records into dated legislative events, with manual-record validation; procedural reconstruction does not establish legal interpretation or judicial reasoning ability.
-
Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making paper — Jointly adapts causal language models for bounded Chinese decisions, including legal training and a 4,300-decision legal test subset; the reported 76.98% professional average is not legal-only accuracy.
-
Thunder-DeID+: Boundary-Aware Korean De-identification from Court Judgments to Broader Sensitive Documents paper — Pretrained ModernBERT encoders and boundary-aware tokenization identify PII in Korean court judgments; improved span recovery does not certify complete anonymization or legal compliance.
-
Making every stage of medical device document assessment with large language models deterministic paper — Deterministic retrieval and assessment of device documents against regulatory requirements; fixed-configuration reproducibility supports audit, not regulatory correctness or approval.
-
Legal-R1: Agentic Retrieval for Evidence-Based Legal Reasoning paper — Multi-agent Chinese legal QA with authority- and diversity-aware evidence retrieval; only the Lawyer agent is post-trained, and benchmark gains do not certify legal advice.
-
PATMT2: Two-stage domain adaptation for multilingual patent translation paper — Two-stage LoRA adaptation of Hy-MT2-7B for Chinese/English/Japanese/Korean patents; six-direction translation gains coexist with failures in four probed low-resource languages.
-
RegExtract: A Benchmark for PFAS Regulatory Information Extraction with LLMs paper — Extracts typed values and provenance from US state PFAS regulations, including permitted uses and product exemptions; field extraction is not an end-to-end compliance determination.
-
The Constitution Speaks Nheengatu: An Open MT System and Reproducible Corpus Pipeline for the Amazonian Língua Geral paper — Portuguese–Nheengatu constitutional translation and terminology evaluation; lexical constraints improve term realization but can cause repetition not captured by chrF++.
-
Verifiable Retrieval for Legal Reasoning through Structure-driven Knowledge Graph paper — Structure-aware German statutory retrieval over 12,037 cases; retrieval-only evaluation does not establish downstream answer correctness, and charging expansion tokens reverses apparent efficiency gains.
-
Benchmarking Embedding Geometries and Clustering Algorithms for Unsupervised Patent Claim Classification paper — Sentence-transformer embeddings and matched hyperbolic/Euclidean VAEs for patent prior-art retrieval; crossing CPC boundaries is a structural proxy, not expert-validated prior-art relevance or patentability.
-
The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection paper — Grounds review and suggested rewrites of Dutch government documents in equal-treatment law and contextual evidence; neutralized wording does not establish that discriminatory practices have been eliminated.
-
Verdicts Without Annotated Evidence: Rejection Sampling or Label-Only Post-Training for Evidence Recovery? paper — ContractNLI evidence recovery without human evidence labels during post-training; entailment accuracy does not establish citation quality, and these verdicts are contract-NLI labels, not court judgments.
-
In With the Old: Enhancing 'Classical' Document Automation with Generative AI paper — Combines symbolic legal-document templates with LLM gap detection, follow-up questions and redrafting; preliminary author-constructed scenarios do not establish legal correctness.
-
Breaking Bureaucracy: Evaluating open-source LLMs for legal document review paper — ContractNLI and NLI4Wills evaluation; the best tested zero-shot LLM reaches 81.2% accuracy, but does not surpass the supervised baseline on overall accuracy.
-
Not Self-Decidable: LLMs Cannot Draw the Boundary of What an Agent Verifier Can Check paper — CoVer tests verification scope for EU AI Act and FINRA predicates; regulatory reference labels are mostly model-assigned proxies, with a limited human-labelled replication.
-
SCM-based Fairness and Faithful Explainability for Legal Document Classification paper
-
Legal text classification in Korean sexual offense cases: from traditional machine learning to large language models with XAI insights paper
-
LawCompass: Navigating from Legal QA to Multi-Agent Deep Research with Grounded Evidence paper
-
ARCCS: An Automated Regulatory Compliance Checking System paper
-
Chinese-Jev: Bringing System One Model to Chinese-Language Tasks paper
-
LAURA: Knowledge Distillation for Interpretable Ambiguous Clause Identification in Legal Contracts paper
-
SinBrief: A Hybrid Framework for Abstractive Text Summarisation of Sinhala Legal Documents paper
-
ARGUS: Role-Aware Event Knowledge Graphs for U.S. Employment-Discrimination Complaints paper
-
Automated Regulatory Compliance Question Answering in Financial Services with Domain-Adapted Retrieval-Augmented Generation paper
-
Controlled Attribute-Specific Summarization of Interrogative Dialogues paper
-
LabourCrew: A Multi-Agent RAG Framework for Trustworthy Adversarial Deliberation and Statutory Reasoning over Labour Law paper
-
Same Scores, Different Decisions: Evaluating JEV and Language Models for Legal Document Understanding paper
-
Cross-Lingual Legal QA for Vietnamese Labour Law: Retrieval, Translation, and Verifier-Guided Correction paper
-
Automated Extraction of Records of Processing Activities (RoPA) Using Hybrid RAG and Locally Deployed Large Language Models paper
-
LexLattice: Multilingual Extractive Summarization via Neural Cellular Automata on Document Hierarchies paper
-
LEGO: Synergizing Expert GraphRAG and Expert Chain-of-Thought for Legal Reasoning paper
-
Classifying Interpretive Canons at the Sentence Level: A Benchmark from the German Federal Constitutional Court paper
-
ContraVis: Evidence-Grounded Visual Analytics for Contradiction Review in Legal Contracts paper
-
Mining Legal Arguments in U.S. Corporate Case Law paper
-
ABAI at COLIEE 2026 Task 1: Multi-Stage Retrieval with GraphRAG-Enhanced Meta-Learning, and a Post-Hoc Study of the Cross-Validation-to-Test Gap paper
-
Efficient LLM Distillation for Bangladesh Legal Context: A Smartphone-Compatible Retrieval-Augmented Generation Model paper
-
GRACE: Grounded Adversarial Reasoning over Canadian Law paper
-
ChemCLIR-Bench: Benchmarking Cross-Lingual Information Retrieval in Multilingual Chemical Patents paper
-
Directing large language models to follow the letter or spirit of the law paper
-
Schematize: An Agentic System for Generating and Refining Information-Extraction Schemas for Legal Research paper
-
Beyond Accuracy and Surface Fluency: Risk-Sensitive Evaluation of LLMs for Legal Clause Generation paper
-
UK-PRBENCH: A Paragraph-Level Precedent Retrieval Benchmark for United Kingdom Case Law paper
-
Benchmarking LLM Compliance with China AI Generated Content Regulations paper
-
Before the Arrest: Benchmarking LLMs on Criminal Profiling from Incomplete Evidence paper
-
How Much is a Human Right Worth? ECtHR-NPD: A Benchmark for Predicting Non-Pecuniary Damage Awards paper
-
Linguistic Triggers of Gender and Racial Bias in Open-Weight LLMs Applied to Recruitment paper
-
Nepali Legal Expertise through Generative and Extractive Pre-trained Transformers (NepLEGiT) paper
-
NepKANUN: A RAG-Based Nepali Legal Assistant paper
-
Few-Shot Degradation Is Not What It Seems: Behavioral Evidence, Representation Analysis, and a Random-Text Control Across 12 Models, 2 Tasks, and 2 Architectures paper
-
CiteGuard-RAG: A Validation-Centered AI System for Evidence-Grounded Question Answering paper
-
Semiotic Relations and Proof Methods: A Cross-Genre Study of Argument Structure with Large Language Models paper
-
Building Legal Reward Models for Grounding and Abstention paper
-
Algorithm Validation as a Policy Audit: Evidence from Race-blind Charging paper
-
Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models paper
-
ProMediConv: Benchmarking Proactive Conversational Agents in Legal Dispute Mediation paper
-
BodyCam-VQA: Enhanced Body-Worn Camera Video Captioning via Multimodal Reasoning and Probe Question Generation paper
-
GANDR: Claim Auditing for Verifiable Legal Answer Generation paper
-
When Does Defendant Statement Matter? A Study of Bias and Persuasion in LLM-Simulated Jurors paper
-
MUCnoHARM@GermEval Shared Task 2026: Retrieval-based In-Context Learning for Defamatory Offences, and Where It Falls Short paper
-
Reading a Legal Question Word by Word: Embedding Trajectories of 2,144 Vietnamese Legal Headlines paper
-
Solving versus Verifying: Catching Contradictions in Tax Reasoning Systems paper
-
LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics paper
-
KhatianDoc: A Human-Verified Benchmark Diagnosing Multimodal LLM Failure on Bengali Legal Land Records paper
-
LexIssue: Benchmarking Legal Issue Identification in Chinese Civil Litigation paper
-
OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction paper
-
Privacy Washing: Detecting Internal Contradictions in Privacy Policies paper
-
Polish ModernBERT: The Long and Short of Polish Language Understanding paper
-
Annotated Surrogate Retrieval for Polish Statutory Law paper
-
Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark paper
-
Improving Argument Saliency Coverage in Small LLMs for Long Legal Opinion Summarization via Sequence-Level Distillation paper
-
JPO: Juris Policy Optimization for Structured Legal Reasoning in Criminal Judgment Prediction paper
-
Cloud and On-Premises Deployment of Uzbek Legal RAG via Targeted Retriever Fine-Tuning paper
-
Cross Lingual Transfer in Tulu Legal Comprehension: Script-Dependent Improvement and RAG-Induced Knowledge Conflict paper
-
RegDivergence-101: An LLM Benchmark for Cross-Jurisdiction Regulatory Contradiction Detection in Life Sciences paper
-
GreenLeaf Law Embed Tiny: A Compact Embedding Model for Legal Domain Retrieval paper
-
Figurative Justice: Detecting metaphors in Hindi judgements with qualitative assessment and transformers paper
-
Modeling Claim Dependency Structure for Patent Litigation Prediction with Graph Attention Networks paper
-
Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards? paper
-
Generative Gap Filling paper
-
Benchmarking Patent Drafting from Inventor-Style Disclosures paper
-
Ansari: A Retrieval-Grounded Islamic AI Assistant -- Architecture, Deployment, and Lessons from 140,000 Conversations paper
-
ContractScrub: A benchmark for final review of legal contracts paper
-
When Machines Speak: A Unified Generative Framework for Integrating Machine-Native Symbols into Pretrained Large Language Models paper
-
ChildSafeAds Shared Task 2026: Commercial Content in Child-Facing YouTube Videos paper
-
GreekBarRetrieval: A Benchmark for Greek Statutory Retrieval paper
-
Redakto - The Incognito Tab for LLMs paper
-
CoAL-RAG: A Complexity-Aware Legal Retrieval-Augmented Generation Method paper
-
Can LLMs Reason in a Legally Meaningful Manner? A Small-scale Study on European Court of Human Rights Cases paper
-
Calibrated Trust, Not Sharper Prediction: An Empirical Test of Uncertainty Fusion paper
-
Time as Structure: Temporal Dependency Graphs for Verifiable Deadline Computation over Legal Documents paper
-
How Much Do Legal RAG Systems Still Hallucinate? paper
-
On Measuring Semantic Preservation in Legal Ontology Learning paper
-
Why Do AI Agents Break Rules? How Framing, Context, and Social Signals Shape Compliance paper
-
Self-Knowledge Retrieval Augmented Generation Framework for Patent Matching paper
-
Evaluating Rational Contracting in Natural Language paper
-
Temporal Misgrounding in Legal RAG: A Versioned-Corpus Benchmark for French Tax Law paper
-
PolicyKG: An Agentic LLM Pipeline for Translating Institutional Policies into SHACL Knowledge Graphs paper
-
ANNOTARES: A Dataset for Extracting Logical Structures from German Statutory Texts paper
-
CrossLex: A Source-Grounded Benchmark for Cross-Jurisdictional Legal Reasoning in Large Language Models paper
-
Semantics of Subterfuge: Benchmarking Legal Deception Detection Against General-domain State-of-the-Art paper
-
Credit Cards, Confusion, Computation, and Consequences: What Can We Uncover About Language Model Reasoning? paper
-
From Naive RAG to Deep Agentic Retrieval: An Evolving Context Engineering Pipeline for Regulatory Compliance paper
-
Legal Prompt Engineering for Multilingual Legal Judgement Prediction
-
Can GPT-3 Perform Statutory Reasoning?
-
Legal Prompting: Teaching a Language Model to Think Like a Lawyer
-
Large Language Models as Fiduciaries: A Case Study Toward Robustly Communicating With Artificial Intelligence Through Legal Standards
-
ChatGPT Goes to Law School
-
ChatGPT, Professor of Law
-
ChatGPT & Generative AI Systems as Quasi-Expert Legal Advice Lawyers - Case Study Considering Potential Appeal Against Conviction of Tom Hayes
-
‘Words Are Flowing Out Like Endless Rain Into a Paper Cup’: ChatGPT & Law School Assessments
-
ChatGPT by OpenAI: The End of Litigation Lawyers?
-
Law Informs Code: A Legal Informatics Approach to Aligning Artificial Intelligence with Humans
-
ChatGPT may Pass the Bar Exam soon, but has a Long Way to Go for the LexGLUE benchmark paper
-
How Ready are Pre-trained Abstractive Models and LLMs for Legal Case Judgement Summarization? paper
-
Explaining Legal Concepts with Augmented Large Language Models (GPT-4) paper
-
Garbage in, garbage out: Zero-shot detection of crime using Large Language Models paper
-
Legal Summarisation through LLMs: The PRODIGIT Project paper
-
Black-Box Analysis: GPTs Across Time in Legal Textual Entailment Task paper
-
PolicyGPT: Automated Analysis of Privacy Policies with Large Language Models
-
Reformulating Domain Adaptation of Large Language Models as Adapt-Retrieve-Revise paper
-
Precedent-Enhanced Legal Judgment Prediction with LLM and Domain-Model Collaboration paper
-
From Text to Structure: Using Large Language Models to Support the Development of Legal Expert Systems paper
-
Boosting legal case retrieval by query content selection with large language models paper
-
LLMediator: GPT-4 Assisted Online Dispute Resolution paper
-
Employing Label Models on ChatGPT Answers Improves Legal Text Entailment Performance paper
-
LLaMandement: Large Language Models for Summarization of French Legislative Proposals paper
-
Logic Rules as Explanations for Legal Case Retrieval paper Our new paper, welcome to pay attention !!!
-
Enhancing Legal Document Retrieval: A Multi-Phase Approach with Large Language Models paper
-
BLADE: Enhancing Black-box Large Language Models with Small Domain-Specific Models paper
-
A Survey on Large Language Models for Critical Societal Domains: Finance, Healthcare, and Law paper
-
Archimedes-AUEB at SemEval-2024 Task 5: LLM explains Civil Procedure paper
-
More Than Catastrophic Forgetting: Integrating General Capabilities For Domain-Specific LLMs paper
-
Legal Documents Drafting with Fine-Tuned Pre-Trained Large Language Model paper
-
Knowledge-Infused Legal Wisdom: Navigating LLM Consultation through the Lens of Diagnostics and Positive-Unlabeled Reinforcement Learning paper
-
GOLDCOIN: Grounding Large Language Models in Privacy Laws via Contextual Integrity Theory paper
-
Enabling Discriminative Reasoning in Large Language Models for Legal Judgment Prediction paper
-
Large Language Models for Judicial Entity Extraction: A Comparative Study paper
-
Applicability of Large Language Models and Generative Models for Legal Case Judgement Summarization paper
-
LawLLM: Law Large Language Model for the US Legal System paper
-
Legal syllogism prompting: Teaching large language models for legal judgment prediction paper
-
Optimizing Numerical Estimation and Operational Efficiency in the Legal Domain through Large Language Models paper
-
Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools paper
-
KRAG Framework for Enhancing LLMs in the Legal Domain paper
-
Legal Evaluations and Challenges of Large Language Models paper
-
Analyzing Images of Legal Documents: Toward Multi-Modal LLMs for Access to Justice paper
-
Automating Legal Concept Interpretation with LLMs: Retrieval, Generation, and Evaluation paper
-
Domaino1s: Guiding LLM Reasoning for Explainable Answers in High-Stakes Domains paper
-
RELexED: Retrieval-Enhanced Legal Summarization with Exemplar Diversity paper
-
Courtroom-LLM: A Legal-Inspired Multi-LLM Framework for Resolving Ambiguous Text Classifications paper
-
LegalViz: Legal Text Visualization by Text To Diagram Generation paper
-
Investigating the Shortcomings of LLMs in Step-by-Step Legal Reasoning paper
-
Elevating Legal LLM Responses: Harnessing Trainable Logical Structures and Semantic Knowledge with Legal Reasoning paper
-
JuDGE: Benchmarking Judgment Document Generation for Chinese Legal System paper
-
LARGE: Legal Retrieval Augmented Generation Evaluation Tool paper
-
Tasks and Roles in Legal AI: Data Curation, Annotation, and Verification paper
-
The Hidden Structure -- Improving Legal Document Understanding Through Explicit Text Formatting paper
-
Conditioning Large Language Models on Legal Systems? Detecting Punishable Hate Speech paper
-
CORE-KG: An LLM-Driven Knowledge Graph Construction Framework for Human Smuggling Networks paper
-
LLM-based Embedders for Prior Case Retrieval paper
-
Towards Automated Regulatory Compliance Verification in Financial Auditing with Large Language Models paper
-
Using Large Language Models for Legal Decision-Making in Austrian Value-Added Tax Law: An Experimental Study paper
-
ASVRI-Legal: Fine-Tuning LLMs with Retrieval Augmented Generation for Enhanced Legal Regulation paper
-
LLMs in Interpreting Legal Documents paper
-
HalluGraph: Auditable Hallucination Detection for Legal RAG Systems via Knowledge Graph Alignment paper
-
Evaluating Legal Reasoning Traces with Legal Issue Tree Rubrics paper
-
Jurisdiction as Structural Barrier: How Privacy Policy Organization May Reduce Visibility of Substantive Disclosures paper
-
Enforcing Monotonic Progress in Legal Cross-Examination: Preventing Long-Horizon Stagnation in LLM-Based Inquiry paper
-
Anchored Decoding: Provably Reducing Copyright Risk for Any Language Model paper
-
Synthesizing the Virtual Advocate: A Multi-Persona Speech Generation Framework for Diverse Linguistic Jurisdictions in Indic Languages paper
-
AI-Assisted Moot Courts: Simulating Justice-Specific Questioning in Oral Arguments paper
-
LLM-Assisted Causal Structure Disambiguation and Factor Extraction for Legal Judgment Prediction paper
-
De Jure: Iterative LLM Self-Refinement for Structured Extraction of Regulatory Rules paper
-
ContextLens: Modeling Imperfect Privacy and Safety Context for Legal Compliance paper
-
Eskwai for Students: Generative AI Assistant for Legal Education in Ghana paper
-
Learning Faster with Better Tokens: Parameter-Efficient Vocabulary Adaptation for Specialized Text Summarization paper
-
Exploring Lightweight Large Language Models for Court View Generation paper
-
Retrieval-Based Multi-Label Legal Annotation: Extensible, Data-Efficient and Hallucination-Free paper
-
Chunking German Legal Code paper
-
Generating Legal Commentaries from Case Databases via Retrieval, Clustering, and Generation paper
-
Decompose-and-Refine: Structured Legal Question Answering with Parametric Retrieval paper
-
Retrieval-Augmented Detection of Potentially Abusive Clauses in Chilean Terms of Service paper
-
Traceable by Design: An LLM Pipeline and Dashboard for EU Regulatory Consultation Analysis paper
-
ImmigrationQA: A Source-Grounded Dataset and Small-Model Adaptation for U.S. Immigration Law paper
-
Citation Grounding: Detecting and Reducing LLM Citation Hallucinations via Legal Citation Graphs paper
-
Enhancing BiGRU with a KAN Block for Legal Document Classification and Summarization paper
-
On Wednesdays, We Ask Questions: Optimizing "Active Listening" in Automated Legal Triage and Referral paper
-
Re-Ranking Through an Attribution Lens for Citation Quality in Legal QA paper
-
EURO-5K: When Does Domain Pretraining Matter? Benchmarking Transformers for EU Reporting Obligation Extraction paper
-
DAR: Deontic Reasoning with Agentic Harnesses paper
-
GARL: Game-Theoretic Reinforcement Learning for Multi-Agent Strategic Prioritisation paper
-
Civil Court Simulation with Large Language Models paper
-
LexRubric: A Rubric-Guided Diagnostic Benchmark for Open-Ended Legal Tasks paper
-
From Statute to Control Flow: Span-Grounded Deontic Trees for Defeasible Scope Parsing paper
-
Retrieval Augmented Generation Framework for the Nepali Legal Domain Question Answering paper
-
FineREX: Fine-Tuned NER-RE for Human Smuggling Knowledge Graphs paper
-
Reinforcement learning to improve large language model-based automated code compliance systems paper
-
Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law Courts paper
-
Peeking Inside LLMs: Leveraging Internal Artifacts of LLMs for Enhancing Reliability in Legal Classification paper
-
A Tree-of-Thoughts Inspired Hybrid Approach for Legal Case Judgement Summarization using LLMs paper
-
PRecG: Legal Precedent Retrieval with Graph Neural Networks and Rhetorical Role Segmentation paper
-
Cross-Architecture LLM Ensembles, Feature-Based Reranking and Retrieval-Augmented Prompting for Legal Information Processing paper
-
When Reasoning Hurts Legal Drafting: The Verbalization Bottleneck in Patent Claim Generation paper
-
The Price of Reasoning: Cost-Quality Tradeoffs in Reinforcement Learning for Neural Machine Translation paper
-
Reasoning Before Translation: Enhancing Legal Machine Translation with Structured Reasoning paper
-
AILQA: Evaluating AI-Driven Legal Question Answering Systems for the Indian Legal System paper
-
Enabling Multilingual Privacy Policy Audits: Large-Scale Analysis of Spanish Mobile Apps paper
-
From Obligation to Specification: A Survey on Validating EU AI Act Requirements in RE paper
-
Retrieval-Augmented Large Language Models as Components of Cognitive Computing architecture for Regulatory Knowledge Management paper
-
Do Small Models Use the Law You Give Them? Context-Injected Fine-Tuning for Legal QA in Bangladesh paper
-
PROSE: Preservation-Constrained Selective Editing for Guilty Bias in Legal Judgment Prediction paper — Applies preservation-regularized low-rank edits to correct unsupported guilty predictions while retaining matched supported verdicts; edited models use ordinary single-pass inference.
-
Commit or Abstain? System One Decision Models, LLMs, and Missing Legal Authority paper — Studies closed-set holding selection with an abstention option and probability distillation into Qwen3-8B; one English US-case-law task and single-run distillation limit generalization.
-
Machina Pauliana: A neuro-symbolic approach to the actio pauliana in Dutch insolvency law paper — LLM routines resolve factual and case-law-dependent inputs while the hand-written symbolic solver applies Articles 42, 43, and 45 Fw; the knowledge requirement remains difficult, and deeper retrieval shows no clear aggregate benefit.
-
Legal-R1: Agentic Retrieval for Evidence-Based Legal Reasoning paper — Masked SFT and trajectory-aware RL train a Lawyer agent to acquire sufficient legal evidence; the retrieval and answer-generation roles are separate agents.
-
Enhancing Legal Reasoning in Pre-trained Language Models via Knowledge Graph-Guided Multi-Task Pre-training paper — KG-guided objectives for nine Transformer encoders on traffic-law QA; synthetic correctness judgments and transfer benchmarks do not establish general judicial reasoning.
-
Verdicts Without Annotated Evidence: Rejection Sampling or Label-Only Post-Training for Evidence Recovery? paper — Compares label-only and rejection-sampling post-training of Qwen3.5-9B for contract entailment and evidence recovery; one seed on one corpus does not establish a reliable method ranking.
-
Better Call Reward: Reward Hacking as Strategic Abstention in Legal Reasoning Models paper — GRPO on Qwen3-8B with surface-feature rewards; the LegalBench accuracy drop is primarily failure to return committed, properly formatted answers, not direct evidence of general reasoning-capability loss.
-
LexReward: A Taxonomy-Driven Reward Framework for Legal Language Models paper
-
Is This Evidence Decision-Critical? Learning to Verify Rule-Governed Decisions paper
-
RGDT-Bench: Benchmarking LLM Reasoning for Rule-Governed Decisions and Their Justifications paper
-
LabourCrew: A Multi-Agent RAG Framework for Trustworthy Adversarial Deliberation and Statutory Reasoning over Labour Law paper
-
LEGO: Synergizing Expert GraphRAG and Expert Chain-of-Thought for Legal Reasoning paper
-
GRACE: Grounded Adversarial Reasoning over Canadian Law paper
-
Directing large language models to follow the letter or spirit of the law paper
-
Semiotic Relations and Proof Methods: A Cross-Genre Study of Argument Structure with Large Language Models paper
-
Building Legal Reward Models for Grounding and Abstention paper
-
GANDR: Claim Auditing for Verifiable Legal Answer Generation paper
-
Solving versus Verifying: Catching Contradictions in Tax Reasoning Systems paper
-
LexIssue: Benchmarking Legal Issue Identification in Chinese Civil Litigation paper
-
OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction paper
-
Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark paper
-
JPO: Juris Policy Optimization for Structured Legal Reasoning in Criminal Judgment Prediction paper
-
Modeling Claim Dependency Structure for Patent Litigation Prediction with Graph Attention Networks paper
-
CoAL-RAG: A Complexity-Aware Legal Retrieval-Augmented Generation Method paper
-
Can LLMs Reason in a Legally Meaningful Manner? A Small-scale Study on European Court of Human Rights Cases paper
-
Time as Structure: Temporal Dependency Graphs for Verifiable Deadline Computation over Legal Documents paper
-
LexPro-1.0 Technical Report paper
-
LexPam: Legal Procedure Awareness-Guided Mathematical Reasoning paper
-
Mining Hidden Thoughts from Texts: Evaluating Continual Pretraining with Synthetic Data for LLM Reasoning paper
-
SyLeR: A Framework for Explicit Syllogistic Legal Reasoning in Large Language Models paper
-
Contractual Deepfakes: Can Large Language Models Generate Contracts? paper
-
LePREC: Reasoning as Classification over Structured Factors for Assessing Relevance of Legal Issues paper
-
Falkor-IRAC: Graph-Constrained Generation for Verified Legal Reasoning in Indian Judicial AI paper
-
Bridging Legal Interpretation and Formal Logic: Faithfulness, Assumption, and the Future of AI Legal Reasoning paper
-
TypedCSIP: Typed Counterfactual Pretraining for Chinese Legislative Conflict Classification paper
-
By Their Fruits You Will Know Them: Comparing Formalizations of Law by the Decisions They Encode paper
-
Scaling Multi-Hop Training Data via Graph-Constrained Path Selection paper
-
DAR: Deontic Reasoning with Agentic Harnesses paper
-
Civil Court Simulation with Large Language Models paper
-
From Statute to Control Flow: Span-Grounded Deontic Trees for Defeasible Scope Parsing paper
-
Generative Chinese Statute Retrieval paper
-
Do Small Models Use the Law You Give Them? Context-Injected Fine-Tuning for Legal QA in Bangladesh paper
-
Reasoning Consensus: Structural Ensembling of LLM Reasoning via Weighted DAG Aggregation paper
-
AI Cosmos: An Agentic Infra for Situated Learning paper — Role-specialized agents support situated legal education; this is a teaching application, not an autonomous lawyer or a newly trained legal reasoning model.
-
Novel or Not? PatEx-Gym: Agentic Prior-Art Search for Novelty Assessment in Patent Examination paper — PatEx-Agent iterates search, document inspection, and examination reporting; missed examiner-cited references remain a limitation.
-
Legal-R1: Agentic Retrieval for Evidence-Based Legal Reasoning paper — Lawyer, Legal Assistant and Judge roles coordinate retrieval from normative sources and precedent cases; evaluation is on Chinese legal benchmarks.
-
APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces paper — A professional-agent safety benchmark with five Law-domain base tasks, including contract work; aggregate results mix law, banking and consulting and are not law-only estimates.
-
The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection paper — MARS-Gov combines legal retrieval, open-set standpoint screening, specialized jurors and rewrite verification; deployment requires jurisdiction-specific normative resources and human review.
-
InvestigationWorlds: An Agentic Environment for Legal Investigation paper — An attorney-validated environment built from 100 federal cases and synthetic surrounding documents; the reference is a court-adopted hypothesis for a summary judgment motion, not unrestricted factual truth.
-
Legal Research Bench: Measuring End-to-End Reliability in Long-Horizon Legal Research Agents paper
-
LawCompass: Navigating from Legal QA to Multi-Agent Deep Research with Grounded Evidence paper
-
ARCCS: An Automated Regulatory Compliance Checking System paper
-
LabourCrew: A Multi-Agent RAG Framework for Trustworthy Adversarial Deliberation and Statutory Reasoning over Labour Law paper
-
Schematize: An Agentic System for Generating and Refining Information-Extraction Schemas for Legal Research paper
-
NepKANUN: A RAG-Based Nepali Legal Assistant paper
-
ProMediConv: Benchmarking Proactive Conversational Agents in Legal Dispute Mediation paper
-
GANDR: Claim Auditing for Verifiable Legal Answer Generation paper
-
OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction paper
-
Cloud and On-Premises Deployment of Uzbek Legal RAG via Targeted Retriever Fine-Tuning paper
-
Benchmarking Patent Drafting from Inventor-Style Disclosures paper
-
Ansari: A Retrieval-Grounded Islamic AI Assistant -- Architecture, Deployment, and Lessons from 140,000 Conversations paper
-
Why Do AI Agents Break Rules? How Framing, Context, and Social Signals Shape Compliance paper
-
Evaluating Rational Contracting in Natural Language paper
-
PolicyKG: An Agentic LLM Pipeline for Translating Institutional Policies into SHACL Knowledge Graphs paper
-
From Naive RAG to Deep Agentic Retrieval: An Evolving Context Engineering Pipeline for Regulatory Compliance paper
-
When Correct Isn't Usable: Improving Structured Output Reliability in Small Language Models paper
-
Accurate Legal Reasoning at Scale: Neuro-Symbolic Offloading and Structural Auditability for Robust Legal Adjudication paper
-
SimuCourt: Building Judicial Decision-Making Agents with Real-world Judgement Documents paper
-
Can Large Language Models Grasp Legal Theories? Enhance Legal Reasoning with Insights from Multi-Agent Collaboration paper
-
Multi-Agent Simulator Drives Language Models for Legal Intensive Interaction paper
-
CLAIM: An Intent-Driven Multi-Agent Framework for Analyzing Manipulation in Courtroom Dialogues paper
-
PAKTON: A Multi-Agent Framework for Question Answering in Long Legal Agreements paper
-
LLM Agents in Law: Taxonomy, Applications, and Challenges paper
-
Dissecting Judicial Reasoning in U.S. Copyright Damage Awards paper
-
Multimodal Multi-Agent Empowered Legal Judgment Prediction paper
-
LegalMALR:Multi-Agent Query Understanding and LLM-Based Reranking for Chinese Statute Retrieval paper
-
LegalOne: A Family of Foundation Models for Reliable Legal Reasoning paper
-
Domain-Partitioned Hybrid RAG for Legal Reasoning: Toward Modular and Explainable Legal AI for India paper
-
LexChronos: An Agentic Framework for Structured Event Timeline Extraction in Indian Jurisprudence paper
-
Fanar-Sadiq: A Multi-Agent Architecture for Grounded Islamic QA paper
-
Dual Hierarchical Dialogue Policy Learning for Legal Inquisitive Conversational Agents paper
-
Can LLMs Time Travel? Enhancing Temporal Consistency in Legal Agentic Search through Reinforcement Learning paper
-
SIA: Self Improving AI with Harness & Weight Updates paper
-
LegalGraphRAG: Multi-Agent Graph Retrieval-Augmented Generation for Reliable Legal Reasoning paper
-
DAR: Deontic Reasoning with Agentic Harnesses paper
-
GARL: Game-Theoretic Reinforcement Learning for Multi-Agent Strategic Prioritisation paper
-
Civil Court Simulation with Large Language Models paper
-
From Statute to Control Flow: Span-Grounded Deontic Trees for Defeasible Scope Parsing paper
-
DeXposure-Claw: An Agentic System for DeFi Risk Supervision paper
-
Agentic Evaluation of Copyright Law Compliance paper
-
From Obligation to Specification: A Survey on Validating EU AI Act Requirements in RE paper
-
PROSE: Preservation-Constrained Selective Editing for Guilty Bias in Legal Judgment Prediction paper — Studies guilty predictions that persist after a legally decisive condition changes; reducing false-guilty errors alone can hide damage to supported verdicts.
-
Does a Norwegian further-trained model give more outdated legal values than its base model? A preregistered paired comparison of Borealis 2 and Gemma 4 paper — Studies outdated numeric answers from Norwegian statutes and regulations; the paired released-model comparison does not isolate a causal effect of further training.
-
Are Fake Citations a Form of Reward Hacking? A Pre-Registered Study of Post-Training and Citation Cues paper — Tests citation fabrication with an explicit US federal case-law subset; prompt effects do not prove that reward hacking caused fabrication, and the 900-question collection is not exclusively legal.
-
HEARSAYBENCH: Can AI Navigate from Abstract Human Rights to Lived Lives? paper — Audits advice under constrained rights and legal recourse, including immigration restrictions and unsupported legal remedies; regional plausibility and model-judge scores do not certify legal correctness.
-
Protection Is Not One Event: A Survey of RAG Protection Research by Measured Endpoint paper — Audits copyright/ownership claims within a 118-paper RAG evidence map, distinguishing corpus-use detection from emitted expression; the copyright subset and its sensitivity analysis use technical labels, not legal infringement determinations.
-
Where Do Legal Factors Live in LLMs? paper — Probes 26 CATO trade-secret factors in three open models and localizes sparse readouts; decodability on controlled narratives does not demonstrate causal use in legal reasoning.
-
Same Probabilities, Different Advice: Gain-Loss Framing Effects in High-Stakes LLM Decision Support paper — Tests equivalent probability frames in advice, including workplace-dispute settlement and consumer remedies; consistent recommendations need not be legally correct.
-
Policy Is Not Capability: Some Safety Benchmarks Measure Which Definition of "Unsafe" You Trained On paper — Audits how supervision definitions change copyright-reproduction guardrail decisions; HarmBench-copyright scores reflect moderation taxonomies, not adjudicated infringement or jurisdiction-specific legal compliance.
-
In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal Cases paper — GUISE tests criminal assistance using real cases and the UNODC ICCS taxonomy; framing and cultural adaptation expose refusal failures, not jurisdiction-specific legal judgments.
-
ContextLift: Breaking the Spurious Correlation Behind Conditional Emergent Misalignment paper — Tests Qwen2.5-7B finetuned on bad legal advice: context augmentation reduces cued broad misalignment from 41.9% to 6.1%, but leaves the learned bad advice intact; one model and an LLM judge limit generalization.
-
UDHRBench: Benchmarking Human Rights in Large Language Models paper — Evaluates alignment with all 30 Universal Declaration of Human Rights articles using expert-labelled responses; benchmark alignment is not a jurisdiction-specific legal judgment.
-
APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces paper — Measures unauthorized actions and disclosure in simulated professional workspaces, including legal tasks; successful deliverables need not imply safe execution. No real client data or live external targets were used.
-
Cited but Not Consulted: A Counterfactual Audit of Legal Chain-of-Thought Faithfulness paper — Naming a statute or precedent is not sufficient evidence that a verdict depends on it; authority-swap confounds and small samples limit interpretation of the faithfulness audit.
-
Verdict Without the Rule: Diagnosing and Auditing Regulatory Rule Sensitivity in LLM Compliance Systems paper — Audits dependence on supplied regulatory rules, including GDPR, HIPAA and the EU AI Act; verdict invariance and a low single-direction probe shift do not prove that all rule information was ignored.
-
When Citations Mislead? A Claim-Level Benchmark for Legal Hallucination Detection paper — Studies unsupported legal claims despite access to the cited opinion; textual support and general legal truth are distinct targets.
-
Calibrated Answers About Randomized Trials From a 4-Billion-Parameter Open Model: A Registered Test and a License-Clean Release paper — Audits article-level training-data reuse permissions and measures the accuracy trade-off of license-filtered fine-tuning. BM25 statistics use the full corpus and the PubMedQA initialization lacks an article-level license audit; the release is not a blanket copyright clearance or a legal-task model.
-
Better Call Reward: Reward Hacking as Strategic Abstention in Legal Reasoning Models paper — Studies reward hacking, abstention and structurally implausible citations in legal reasoning.
-
Not Self-Decidable: LLMs Cannot Draw the Boundary of What an Agent Verifier Can Check paper — Models disagree on whether regulatory predicates admit fixed checks; unanimous labels do not by themselves certify safe enforcement.
-
SCM-based Fairness and Faithful Explainability for Legal Document Classification paper
-
ContextAdapt: Evaluating Contextual Adaptation and Value Alignment in LLMs paper
-
Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model Answers paper
-
Directing large language models to follow the letter or spirit of the law paper
-
Beyond Accuracy and Surface Fluency: Risk-Sensitive Evaluation of LLMs for Legal Clause Generation paper
-
Benchmarking LLM Compliance with China AI Generated Content Regulations paper
-
Before the Arrest: Benchmarking LLMs on Criminal Profiling from Incomplete Evidence paper
-
Linguistic Triggers of Gender and Racial Bias in Open-Weight LLMs Applied to Recruitment paper
-
Legal LLM Hallucination Should Be Evaluated as Failure of Legal Warrant paper
-
Few-Shot Degradation Is Not What It Seems: Behavioral Evidence, Representation Analysis, and a Random-Text Control Across 12 Models, 2 Tasks, and 2 Architectures paper
-
CiteGuard-RAG: A Validation-Centered AI System for Evidence-Grounded Question Answering paper
-
Building Legal Reward Models for Grounding and Abstention paper
-
Algorithm Validation as a Policy Audit: Evidence from Race-blind Charging paper
-
Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models paper
-
BodyCam-VQA: Enhanced Body-Worn Camera Video Captioning via Multimodal Reasoning and Probe Question Generation paper
-
GANDR: Claim Auditing for Verifiable Legal Answer Generation paper
-
When Does Defendant Statement Matter? A Study of Bias and Persuasion in LLM-Simulated Jurors paper
-
MUCnoHARM@GermEval Shared Task 2026: Retrieval-based In-Context Learning for Defamatory Offences, and Where It Falls Short paper
-
Reading a Legal Question Word by Word: Embedding Trajectories of 2,144 Vietnamese Legal Headlines paper
-
FramingQA: Does the Question Shape the Answer? Measuring the Compositional Framing Effect paper
-
Solving versus Verifying: Catching Contradictions in Tax Reasoning Systems paper
-
LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics paper
-
Rent-a-RAG: Embedding-Space Watermarks for Auditing Third-Party RAG paper
-
OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction paper
-
Cross Lingual Transfer in Tulu Legal Comprehension: Script-Dependent Improvement and RAG-Induced Knowledge Conflict paper
-
Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards? paper
-
ContractScrub: A benchmark for final review of legal contracts paper
-
Redakto - The Incognito Tab for LLMs paper
-
Can LLMs Reason in a Legally Meaningful Manner? A Small-scale Study on European Court of Human Rights Cases paper
-
Calibrated Trust, Not Sharper Prediction: An Empirical Test of Uncertainty Fusion paper
-
Gated Against One Model, Open to the Next: Option-Only Solvability in Legal Multiple-Choice Benchmarks paper
-
How Much Do Legal RAG Systems Still Hallucinate? paper
-
Why Do AI Agents Break Rules? How Framing, Context, and Social Signals Shape Compliance paper
-
Is this Citation on Point? paper
-
SoK: From Generation to Consumption of Privacy Documents in Software Systems paper
-
Evaluating Rational Contracting in Natural Language paper
-
Temporal Misgrounding in Legal RAG: A Versioned-Corpus Benchmark for French Tax Law paper
-
LexKairos: Benchmarking Legal Temporal Capabilities in LLMs paper
-
Knowing the Form, Not the Function: Automatically Auditing Answer--Authority Decoupling in Legal Benchmarks paper
-
CrossLex: A Source-Grounded Benchmark for Cross-Jurisdictional Legal Reasoning in Large Language Models paper
-
Semantics of Subterfuge: Benchmarking Legal Deception Detection Against General-domain State-of-the-Art paper
-
Towards WinoQueer: Developing a Benchmark for Anti-Queer Bias in Large Language Models
-
Persistent Anti-Muslim Bias in Large Language Models
-
Understanding the Capabilities, Limitations, and Societal Impact of Large Language Models
-
The Dark Side of ChatGPT: Legal and Ethical Challenges from Stochastic Parrots and Hallucination
-
The GPTJudge: Justice in a Generative AI World paper
-
Is the U.S. Legal System Ready for AI's Challenges to Human Values? paper
-
Questioning Biases in Case Judgment Summaries: Legal Datasets or Large Language Models? paper
-
Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models
-
A Legal Framework for Natural Language Processing Model Training in Portugal paper
-
LegalBench-RAG: A Benchmark for Retrieval-Augmented Generation in the Legal Domain paper
-
Bias in Large Language Models: Origin, Evaluation, and Mitigation paper
-
An Information Theoretic Approach to Operationalize Right to Data Protection paper
-
Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench paper
-
Standardizing Intelligence: Aligning Generative AI for Regulatory and Operational Compliance paper
-
TRIDENT: Benchmarking LLM Safety in Finance, Medicine, and Law paper
-
Large Language Models as Search Engines: Societal Challenges paper
-
From Unfamiliar to Familiar: Detecting Pre-training Data via Gradient Deviations in Large Language Models paper
-
TimeMark: A Trustworthy Time Watermarking Framework for Exact Generation-Time Recovery from AIGC paper
-
To See is Not to Learn: Protecting Multimodal Data from Unauthorized Fine-Tuning of Large Vision-Language Model paper
-
Divergence Decoding: Inference-Time Unlearning via Auxiliary Models paper
-
Citation Grounding: Detecting and Reducing LLM Citation Hallucinations via Legal Citation Graphs paper
-
Which Institutional Frameworks Do Chatbots Assume? Auditing Jurisdictional Defaults in Multilingual LLMs paper
-
Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law Courts paper
-
Who Checks the Citations? Benchmarking Legal Hallucination Detection paper
-
Peeking Inside LLMs: Leveraging Internal Artifacts of LLMs for Enhancing Reliability in Legal Classification paper
-
Position: The Term "Machine Unlearning" Is Overused in LLMs paper
-
Open but Incompatible: A License Compatibility Analysis of Corpora for Low-Resource African Languages paper
-
Probing Stylistic Appropriation using Large Language Models: An Evaluation Framework for Copyright Infringement under EU Law paper
-
The Perplexity Trap: When Patent Law Makes Human Writing Look Like AI paper
-
Agentic Evaluation of Copyright Law Compliance paper
-
PROSE: Preservation-Constrained Selective Editing for Guilty Bias in Legal Judgment Prediction paper — Introduces CAIL-MC with 11,478 legally reviewed matched pairs across 261 Chinese criminal charges, plus an evidence-sufficiency track built from LeCaRD and Criminal-S; these are controlled contrasts, not a prospective court-deployment study.
-
NitiBench+: A Temporal Thai Legal Question Answering Benchmark Across 17 Domains paper — Provides 1,670 Thai legal QA items across 17 domains, pairing cited law at the decision date with current text and a versioned statutory retrieval corpus.
-
ICJ-Graph: A Citation Graph and a Layered Document Corpus of the International Court of Justice paper — Links 6,396 ICJ documents and 505 other legal documents to provisions and UN resolutions, preserving separate court-decision, written-pleading, and oral-record layers.
-
Do Open Language Models Invent Scripture? BILAL-BENCH: Benchmarking Islamic-Source Fabrication in Bengali paper — A 240-item Bengali source-fabrication benchmark containing 45 Hanafi fiqh-ruling questions; inclusion is limited to this jurisprudence subset, not the entire religious-knowledge collection.
-
IndicIPR-QA: A Benchmark for Extractive Question Answering on Indian IPR Laws paper — Contains 4,462 generated QA pairs from 1,116 contexts across seven Indian IPR Acts, with automatic/LLM checks and a 200-pair human audit.
-
HEARSAYBENCH: Can AI Navigate from Abstract Human Rights to Lived Lives? paper — A 400-scenario human-authored benchmark; 236 scenarios fall under the authors' Public Policies & Law subtype, which should not be relabeled as 236 independently validated legal questions.
-
Statutory Article Retrieval from Arabic Judicial Narratives: Family Law and Law of Evidence in Saudi Arabia paper — Defines a benchmark with 200 synthetic Arabic judicial cases and 381 statutory articles; synthetic narratives do not substitute for real-case validation.
-
Novel or Not? PatEx-Gym: Agentic Prior-Art Search for Novelty Assessment in Patent Examination paper — PatEx-Bench contains 1,000 examination tasks with Patent-1M and Paper-52K retrieval corpora; PASS curates about 3,000 training trajectories.
-
AusClinBench: a deterministically scored benchmark of large language models against current Australian clinical policy paper — Defines 142 policy items, including 13 Ahpra obligations items under the National Law; the paper says code and 30 sample items are available on request, while 112 items and raw outputs are not published. The full collection is not exclusively legal.
-
A Grammar of French Civil Cassation: A Ground-Level Dataset and Tool paper — Releases 121,536 ground–reasons pairs from 86,464 French civil rulings, with rule-derived outcomes, doctrinal families, code, and a consultation tool; canonical-phrasing rules can miss unusual formulations.
-
Protection Is Not One Event: A Survey of RAG Protection Research by Measured Endpoint paper — Releases a human-coded RAG protection evidence map and coding protocols with an explicit copyright/ownership subset; the broader collection is not a corpus of legal judgments or an infringement-labelled dataset.
-
Towards English - Mizo Legal Machine Translation: Corpus Creation, Analysis, and Benchmarks paper — Introduces 13,168 manually aligned English–Mizo sentence pairs from Indian court judgments, professionally translated and independently reviewed.
-
norregel: a preregistered benchmark of dated numeric values in Norwegian law paper — Builds 916 dated statutory/regulatory/parliamentary-value items from Lovdata; item texts and answer keys are withheld, while IDs, answer classes, code and results are reported on Zenodo.
-
Same Probabilities, Different Advice: Gain-Loss Framing Effects in High-Stakes LLM Decision Support paper — FrameBench-HPS contains 2,880 prompts from 240 synthetic scenarios across four domains, including a concrete legal/consumer-rights subset; these totals are not legal-only counts.
-
When Copying Is Hard: Copy-Constrained Decoding for Exact Span Reproduction paper — Constructs a contract-clause extraction benchmark from SEC EDGAR exhibits, with 658 training and 141 test examples; this is one domain of a broader exact-copy benchmark.
-
Legal Guidance in the Wild: How Users Seek Legal Help in Real-World LLM Conversations paper — Annotates 2,122 WildChat-1M conversations and supplies a 200-conversation human-reviewed gold set, with public annotations, checkpoints, and evaluation code.
-
The Italian lawmaking process: a unified and structured dataset paper — Reports procedural histories for 5,346 Italian laws enacted in 1985–2026, comprising 101,413 dated events extracted from institutional parliamentary records; scope and coverage follow the authors' report.
-
In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal Cases paper — GUISE provides real-case criminal-assistance scenarios organized by ICCS, with three framings and five cultural settings; the resource evaluates harmful assistance, not legal adjudication.
-
LegalRewardBench: Benchmarking Reward Models for Legal Preference Evaluation paper — Human-verified legal preference benchmark spanning knowledge, understanding, analysis, decision and writing; automatic rubric judging is combined with human verification.
-
PatTransBench: an ipc-stratified, multi-dimensional benchmark for patent machine translation evaluation paper — 12,000 short/long patent translation pairs across Chinese–English/Japanese/Korean and both directions; release is planned upon publication, not verified as already available.
-
UDHRBench: Benchmarking Human Rights in Large Language Models paper — 1,124 questions and 4,496 expert-annotated responses covering the UDHR; the abstract announces a future open-source release.
-
RegExtract: A Benchmark for PFAS Regulatory Information Extraction with LLMs paper — 488 annotated fields across 20 US state documents and 46 regulations, each with a value and provenance (976 entries); code/data are linked by the authors.
-
The Constitution Speaks Nheengatu: An Open MT System and Reproducible Corpus Pipeline for the Amazonian Língua Geral paper — 14.5k Portuguese–Nheengatu parallel pairs anchored by the official Brazilian Constitution translation, plus speech evaluation; the corpus is not exclusively legal text.
-
Enhancing Legal Reasoning in Pre-trained Language Models via Knowledge Graph-Guided Multi-Task Pre-training paper — Reports 12,554 synthetic StVO QA pairs, a parsed KG and pipeline resources; heuristic quality filtering and shared generators do not provide independent legal-expert validation.
-
Verifiable Retrieval for Legal Reasoning through Structure-driven Knowledge Graph paper — Reports a German statutory KG, six chunk corpora and case identifiers with parsed citations/reconstruction scripts; evaluation reference citations are restricted to codes covered by the graph.
-
AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis paper — Releases weighted evaluation subsets for illicit-transaction detection in synthetic AMLworld graphs: 3,753 HI-Small and 2,268 LI-Small targets. These are anti-money-laundering detection resources, not court verdicts or statutory labels.
-
APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces paper — Fifteen professional base tasks, five in Law, expanded into 180 attack instances per model; security assets and authorization scenarios are synthetic, not real client matters.
-
When Citations Mislead? A Claim-Level Benchmark for Legal Hallucination Detection paper — PARCEL contains 3,396 synthetic claims grounded in 85 New York State Court of Appeals decisions, labeled Supported, Refuted or Not Found; these are generated benchmark claims, not 3,396 naturally occurring litigation errors.
-
InsClaimBench: Benchmarking Insurance Claim Adjudication Across the Decision Chain paper — 3,780 synthetic claim files in 375 families, grounded in insurance-contract provisions and rules; approximately 10% of families underwent expert review, not the entire benchmark.
-
InvestigationWorlds: An Agentic Environment for Legal Investigation paper — Real PACER case records augmented with role-tagged synthetic evidence for legal investigation; distinguish original exhibits, generated documents and court-adopted reference hypotheses.
-
HakemBench: A Turkish Benchmark of Typed Decisions paper — Includes constitutional-right identification and court routing; legal labels are partly AI-derived.
-
Verifiable, Articulable, and Tacit Components of Preference paper — Includes law/policy tasks; patent prediction results were withdrawn due to label leakage.
-
Legal Research Bench: Measuring End-to-End Reliability in Long-Horizon Legal Research Agents paper
-
ContextAdapt: Evaluating Contextual Adaptation and Value Alignment in LLMs paper
-
LegalPincite: Multi-level Legal Information Retrieval Dataset paper
-
ViLegalExpert: A Large-Scale Benchmark for Vietnamese Legal Retrieval and Question Answering from Real-World Consultations paper
-
Chinese-Jev: Bringing System One Model to Chinese-Language Tasks paper
-
RGDT-Bench: Benchmarking LLM Reasoning for Rule-Governed Decisions and Their Justifications paper
-
Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model Answers paper
-
Controlled Attribute-Specific Summarization of Interrogative Dialogues paper
-
Cross-Lingual Legal QA for Vietnamese Labour Law: Retrieval, Translation, and Verifier-Guided Correction paper
-
Automated Extraction of Records of Processing Activities (RoPA) Using Hybrid RAG and Locally Deployed Large Language Models paper
-
Classifying Interpretive Canons at the Sentence Level: A Benchmark from the German Federal Constitutional Court paper
-
Mining Legal Arguments in U.S. Corporate Case Law paper
-
GRACE: Grounded Adversarial Reasoning over Canadian Law paper
-
ChemCLIR-Bench: Benchmarking Cross-Lingual Information Retrieval in Multilingual Chemical Patents paper
-
UK-PRBENCH: A Paragraph-Level Precedent Retrieval Benchmark for United Kingdom Case Law paper
-
Benchmarking LLM Compliance with China AI Generated Content Regulations paper
-
Before the Arrest: Benchmarking LLMs on Criminal Profiling from Incomplete Evidence paper
-
How Much is a Human Right Worth? ECtHR-NPD: A Benchmark for Predicting Non-Pecuniary Damage Awards paper
-
Nepali Legal Expertise through Generative and Extractive Pre-trained Transformers (NepLEGiT) paper
-
NepKANUN: A RAG-Based Nepali Legal Assistant paper
-
Building Legal Reward Models for Grounding and Abstention paper
-
Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models paper
-
ProMediConv: Benchmarking Proactive Conversational Agents in Legal Dispute Mediation paper
-
When Does Defendant Statement Matter? A Study of Bias and Persuasion in LLM-Simulated Jurors paper
-
FramingQA: Does the Question Shape the Answer? Measuring the Compositional Framing Effect paper
-
LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics paper
-
KhatianDoc: A Human-Verified Benchmark Diagnosing Multimodal LLM Failure on Bengali Legal Land Records paper
-
LexIssue: Benchmarking Legal Issue Identification in Chinese Civil Litigation paper
-
OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction paper
-
Polish ModernBERT: The Long and Short of Polish Language Understanding paper
-
Annotated Surrogate Retrieval for Polish Statutory Law paper
-
Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark paper
-
Cloud and On-Premises Deployment of Uzbek Legal RAG via Targeted Retriever Fine-Tuning paper
-
RegDivergence-101: An LLM Benchmark for Cross-Jurisdiction Regulatory Contradiction Detection in Life Sciences paper
-
GreenLeaf Law Embed Tiny: A Compact Embedding Model for Legal Domain Retrieval paper
-
Figurative Justice: Detecting metaphors in Hindi judgements with qualitative assessment and transformers paper
-
Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards? paper
-
Benchmarking Patent Drafting from Inventor-Style Disclosures paper
-
ImmigrationReason: A Structured Dataset of U.S. Immigration Appeals for Legal Reasoning Research paper
-
ContractScrub: A benchmark for final review of legal contracts paper
-
ChildSafeAds Shared Task 2026: Commercial Content in Child-Facing YouTube Videos paper
-
GreekBarRetrieval: A Benchmark for Greek Statutory Retrieval paper
-
Gated Against One Model, Open to the Next: Option-Only Solvability in Legal Multiple-Choice Benchmarks paper
-
Temporal Misgrounding in Legal RAG: A Versioned-Corpus Benchmark for French Tax Law paper
-
LexKairos: Benchmarking Legal Temporal Capabilities in LLMs paper
-
ANNOTARES: A Dataset for Extracting Logical Structures from German Statutory Texts paper
-
CrossLex: A Source-Grounded Benchmark for Cross-Jurisdictional Legal Reasoning in Large Language Models paper
-
Credit Cards, Confusion, Computation, and Consequences: What Can We Uncover About Language Model Reasoning? paper
-
CAIL2018: A Large-Scale Legal Dataset for Judgment Prediction
-
When does pretraining help? assessing self-supervised learning for law and the casehold dataset of 53,000+ legal holdings
-
LeCaRD: a legal case retrieval dataset for Chinese law system
-
LeXFiles and LegalLAMA: Facilitating English Multinational Legal Language Model Development
-
Legal Extractive Summarization of U.S. Court Opinions
-
Awesome Chinese Legal Resources github
-
MultiLegalPile: A 689GB Multilingual Legal Corpus paper
-
The Cambridge Law Corpus: A Corpus for Legal AI Research
-
TransformLLM: Adapting Large Language Models via LLM-Transformed Reading Comprehension Text paper
-
Natural Language Processing for the Legal Domain: A Survey of Tasks, Datasets, Models, and Challenges paper
-
ChineseSafe: A Chinese Benchmark for Evaluating Safety in Large Language Models paper
-
Augmenting Legal Decision Support Systems with LLM-based NLI for Analyzing Social Media Evidence paper
-
CaseSumm: A Large-Scale Dataset for Long-Context Summarization from U.S. Supreme Court Opinions paper
-
From Answers to Guidance: A Proactive Dialogue System for Legal Documents paper
-
A Greek Government Decisions Dataset for Public-Sector Analysis and Insight [paper](https://arxiv.org/abs/2512.05647
-
Data Augmented Pipeline for Legal Information Extraction and Reasoning paper
-
Domain-Adaptation through Synthetic Data: Fine-Tuning Large Language Models for German Law paper
-
Knowledge Augmented Entity and Relation Extraction for Legal Documents with Hypergraph Neural Network paper
-
Quecto-V1: Empirical Analysis of 8-bit Quantized Small Language Models for On-Device Legal Retrieval paper
-
SinhaLegal: A Benchmark Corpus for Information Extraction and Analysis in Sinhala Legislative Texts paper
-
LAMUS: A Large-Scale Corpus for Legal Argument Mining from U.S. Caselaw using LLMs paper
-
RoLegalGEC: Legal Domain Grammatical Error Detection and Correction Dataset for Romanian paper
-
Automatic Construction of a Legal Citation Graph from 100 Million Ukrainian Court Decisions: Large-Scale Extraction, Topological Analysis, and Ontology-Driven Clustering paper
-
Temporal Decay of Co-Citation Predictability: A 20-Year Statute Retrieval Benchmark from 396M Ukrainian Court Citations paper
-
IMLJD: A Computational Dataset for Indian Matrimonial Litigation Analysis paper
-
Fine-grained Claim-level RAG Benchmark for Law paper
-
A Reproducible Universal Dependencies-Style Pipeline for Katharevousa Greek Parliamentary Text paper
-
Temporal Concept Drift in Legal Judgment Prediction: Neural Baselines Across Three Epochs of Ukrainian Court Decisions paper
-
Bundesrecht: An Open Library and Corpus for German Statutory Reference Processing paper
-
CanLegalRAGBench: Evaluating Retrieval-Augmented Generation on Canadian Case Law paper
-
Scaling Multi-Hop Training Data via Graph-Constrained Path Selection paper
-
ImmigrationQA: A Source-Grounded Dataset and Small-Model Adaptation for U.S. Immigration Law paper
-
Citation Grounding: Detecting and Reducing LLM Citation Hallucinations via Legal Citation Graphs paper
-
Enhancing BiGRU with a KAN Block for Legal Document Classification and Summarization paper
-
Digging Up Citations: FOSSIL, a Dataset and Workflow for Reference Extraction in Law and the Humanities paper
-
EURO-5K: When Does Domain Pretraining Matter? Benchmarking Transformers for EU Reporting Obligation Extraction paper
-
HKJudge: A Legal Discourse-Annotated Corpus for Interpreting What Courts Find, How They Reason, and What They Rule paper
-
LexRubric: A Rubric-Guided Diagnostic Benchmark for Open-Ended Legal Tasks paper
-
From Statute to Control Flow: Span-Grounded Deontic Trees for Defeasible Scope Parsing paper
-
Retrieval Augmented Generation Framework for the Nepali Legal Domain Question Answering paper
-
REDACT: A Systematically Controlled Multilingual Benchmark for Personal Information Detection paper
-
FineREX: Fine-Tuned NER-RE for Human Smuggling Knowledge Graphs paper
-
Who Checks the Citations? Benchmarking Legal Hallucination Detection paper
-
Open but Incompatible: A License Compatibility Analysis of Corpora for Low-Resource African Languages paper
-
Grounded Event Extraction from SEC 8-K Filings with a Fine-Grained Taxonomy paper
-
Enabling Multilingual Privacy Policy Audits: Large-Scale Analysis of Spanish Mobile Apps paper
-
Do Small Models Use the Law You Give Them? Context-Injected Fine-Tuning for Legal QA in Bangladesh paper
-
XLaw-JP-27B: A Lightweight Open Model Specialized for Japanese Law paper — Japanese-law-specialized adaptation of Qwen3.8-27B with open weights; reported single-GPU feasibility and exam performance do not establish production legal reliability.
-
Novel or Not? PatEx-Gym: Agentic Prior-Art Search for Novelty Assessment in Patent Examination paper — PatEx-Agent is a patent-specialized Qwen3.5-4B model fine-tuned on search trajectories, not a general-purpose legal foundation model.
-
PATMT2: Two-stage domain adaptation for multilingual patent translation paper — Patent-specialized 7B translation model adapted with LoRA; included as a domain language model, not a general legal reasoning assistant.
-
Enhancing Legal Reasoning in Pre-trained Language Models via Knowledge Graph-Guided Multi-Task Pre-training paper — Adapts existing BERT-style legal/general encoders with KG-guided multi-task pre-training; these are encoder language models, not a new generative legal assistant.
-
Domain adaptation of Russian ModernBERT for long legal documents paper — RuModernBERT-ruLaw is a masked-language-model encoder, not a generative legal assistant. Lower masked-token loss is supported; the entity test has 99.95% normalized span/class overlap with training and does not establish unseen-form generalization.
-
LexReward: A Taxonomy-Driven Reward Framework for Legal Language Models paper
-
Efficient LLM Distillation for Bangladesh Legal Context: A Smartphone-Compatible Retrieval-Augmented Generation Model paper
-
GRACE: Grounded Adversarial Reasoning over Canadian Law paper
-
Nepali Legal Expertise through Generative and Extractive Pre-trained Transformers (NepLEGiT) paper
-
NepKANUN: A RAG-Based Nepali Legal Assistant paper
-
Building Legal Reward Models for Grounding and Abstention paper
-
Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark paper
-
GreenLeaf Law Embed Tiny: A Compact Embedding Model for Legal Domain Retrieval paper
-
LawGPT_zh github
-
LaWGPT github
-
Lawyer LLaMA github
-
LexiLaw github
-
LexGPT 0.1: pre-trained GPT-J models with Pile of Law paper
-
TOWARDS THE EXPLOITATION OF LLM-BASED CHATBOT FOR PROVIDING LEGAL SUPPORT TO PALESTINIAN COOPERATIVES paper
-
ChatLaw: Open-Source Legal Large Language Model with Integrated External Knowledge Bases paper
-
DISC-LawLLM github
-
InternLM-Law: An Open Source Chinese Legal Large Language Model paper
-
SaulLM-54B & SaulLM-141B: Scaling Up Domain Adaptation for the Legal Domain paper
-
LAWGPT: KNOWLEDGE-GUIDED DATA GENERATION AND ITS APPLICATION TO LEGAL LLM paper
-
HKGAI-V1: Towards Regional Sovereign Large Language Model for Hong Kong paper
-
HUKUKBERT: Domain-Specific Language Model for Turkish Law paper
-
PoliLegalLM: A Technical Report on a Large Language Model for Political and Legal Affairs paper
-
ImmigrationQA: A Source-Grounded Dataset and Small-Model Adaptation for U.S. Immigration Law paper
-
Legal Domain Adaptation of Modern BERT Models paper
-
Do Small Models Use the Law You Give Them? Context-Injected Fine-Tuning for Legal QA in Bangladesh paper
-
PROSE: Preservation-Constrained Selective Editing for Guilty Bias in Legal Judgment Prediction paper — Tests repair, matched-verdict preservation, and evidence sensitivity on two base-model backbones. SFT remains strongest on matched pairs but degrades on the evidence-sufficiency track; PROSE reports a single fitting run, with bootstrap intervals rather than multi-seed fitting variance.
-
Say Less, Mean Better: Training-Free Verbosity Control for Language Models paper — MinimalLM's law example asks for the difference between civil and criminal law: Mistral-7B output falls from 429 to 286 tokens and 13.6 to 9.1 seconds. This is one query per domain; structural completeness and latency are not legal-answer accuracy or semantic completeness.
-
NitiBench+: A Temporal Thai Legal Question Answering Benchmark Across 17 Domains paper — Evaluates 26 models on a 500-item held-out split under time-correct law; citation overlap alone misses substantive rule-change errors, and retrieval helps 20 models but harms six.
-
Unlocking Criminal Court Decisions: Flexible Information Extraction for Empirical Legal Research paper — Revised DUI annotation instructions are associated with extraction accuracy rising from 52% to 70%; this is not a randomized estimate of a general causal improvement.
-
ICJ-Graph: A Citation Graph and a Layered Document Corpus of the International Court of Justice paper — Tests five models on citation prediction for 199 decisions using documents available beforehand; higher citation F1 is not proof of legally correct reasoning.
-
Does a Norwegian further-trained model give more outdated legal values than its base model? A preregistered paired comparison of Borealis 2 and Gemma 4 paper — Among 411 jointly interpretable items out of 503, the outdated-answer difference is -0.2 percentage points, with a 95% interval of [-2.7, +2.5], within the preregistered +/-10-point equivalence region; the benchmark is closed.
-
Saidzoda-Tajik-E5-Large: Fine-Tuning a 560M Embedder for Tajik Retrieval paper — Reports top-10 passage recall of 93.6% on unseen laws versus 80.2% before adaptation; train and test queries share an LLM generator, limiting real-user generalization.
-
Are Fake Citations a Form of Reward Hacking? A Pre-Registered Study of Post-Training and Citation Cues paper — Four preregistered experiments include federal case law alongside scientific and medical questions; the citation-cue interaction is supported, while the predicted targeted reward-hacking pattern is not.
-
Do Open Language Models Invent Scripture? BILAL-BENCH: Benchmarking Islamic-Source Fabrication in Bengali paper — Evaluates seven models on sourced answers, including Hanafi fiqh rulings; school-specific reference judgments and fabrication labels do not establish validity across Islamic schools, national laws, or court practice.
-
Homebrewed SLMs: Autoregressive and Masked Diffusion Language Models Built at Home for Portuguese paper — Evaluates small language models on Brazilian statutory-text continuation, tampering detection, and cross-reference completion; results are mixed against n-gram/public-model baselines, and the blind human reading is not yet included.
-
IndicIPR-QA: A Benchmark for Extractive Question Answering on Indian IPR Laws paper — InLegalBERT reaches 85.02 token F1 versus BM25's 20.21 under the reported protocol; span extraction is not legal interpretation or advice.
-
AI Cosmos: An Agentic Infra for Situated Learning paper — The legal study analyzes 111 students across three arms; the advantage over vanilla chat is supported, while comparison with the teacher baseline and delayed retention remain statistically unresolved.
-
XLaw-JP-27B: A Lightweight Open Model Specialized for Japanese Law paper — Evaluates eight Japanese legal tasks and reports passing three bar-exam years; contract review and patent law remain gaps to frontier systems.
-
HEARSAYBENCH: Can AI Navigate from Abstract Human Rights to Lived Lives? paper — Tests 14 models for situation inference and actionable advice, including legal feasibility; scores rely on LLM judging with human checks, not jurisdiction-wide legal validation.
-
Statutory Article Retrieval from Arabic Judicial Narratives: Family Law and Law of Evidence in Saudi Arabia paper — DaR raises Evidence MRR@10 from 0.087 to 0.696 and reaches 0.864 on Family Law; retrieving relevant articles does not prove a correct legal conclusion.
-
Novel or Not? PatEx-Gym: Agentic Prior-Art Search for Novelty Assessment in Patent Examination paper — Separates decision accuracy, statutory grounds, and search/visit/selection recall; Office Action quality uses LLM judges, not independent legal certification.
-
Can Language Models Learn While They Think? Testing Gradient Adaptation Across Architectures paper — Includes SCOTUS classification, ContractNLI, and court-opinion citation completion. Under the tested recipe, updating weights on the document does not materially recover truncated-context performance; lower language-model loss is not evidence of better legal retrieval.
-
AusClinBench: a deterministically scored benchmark of large language models against current Australian clinical policy paper — Tests 28 models with and without governing clauses, including a statutory advertising-permission/penalty example. Ahpra-subset pooled blind accuracy is 77.7%; keys lack independent review, serving routes differ, and supplying the exact clause is not a test of reliable real-world legal retrieval.
-
A Grammar of French Civil Cassation: A Ground-Level Dataset and Tool paper — Compares rule-based doctrinal labels with JuriBERT, CamemBERTv2, and zero-shot Qwen3-32B on held-out expert labels; models trained on silver labels inherit some rule errors, and post-hoc classification is not prospective judgment prediction.
-
Protection Is Not One Event: A Survey of RAG Protection Research by Measured Endpoint paper — Separates copyright-protection claim auditing from a controlled copying/fact-recovery testbed; reduced verbatim output does not establish privacy protection, legal similarity, or copyright compliance.
-
Towards English - Mizo Legal Machine Translation: Corpus Creation, Analysis, and Benchmarks paper — NLLB adaptation improves BLEU by 7.16 points on a 1,000-pair legal test set, while English borrowing remains about three times the human rate.
-
norregel: a preregistered benchmark of dated numeric values in Norwegian law paper — Distinguishes correct, outdated, wrong and uninterpretable answers; support for the cutoff-related outdated-answer hypothesis is not robust to the preregistered most-recent-value sensitivity analysis.
-
Where Do Legal Factors Live in LLMs? paper — Reports all 26 factors linearly decodable and single-feature best-layer AUCs of 0.86–1.00 across three models; shared generation procedures and strong lexical baselines limit real-case inference.
-
Commit or Abstain? System One Decision Models, LLMs, and Missing Legal Authority paper — Replaces the correct holding in 1,670 of 5,314 CaseHOLD items; generative LLMs select an incorrect option in 38–86% of missing-authority cases, versus 30% for the most cautious decision model.
-
Multi-level Encoder with Global Topic for Task-oriented Dialogue Summarization paper — Evaluates on LCSPIRT police-interrogation summaries and Medical DS; LCSPIRT ROUGE-L reaches 83.34, but the shown output still changes a date and location digits.
-
Same Probabilities, Different Advice: Gain-Loss Framing Effects in High-Stakes LLM Decision Support paper — Across seven LLMs, framing changes recommendations in 32.4% of complete groups across all four domains; this is not a legal-only error rate or a legal-correctness measure.
-
When Copying Is Hard: Copy-Constrained Decoding for Exact Span Reproduction paper — Evaluates exact span matching and decoding throughput, including legal contracts; aggregate four-domain gains are not legal-only scores, and copying fidelity does not establish legal validity.
-
Hidden Cartographer: Hidden State Guided Data Selection for Dual Budgeted Instruction Tuning paper — Evaluates three objective legal tasks on 1,500 held-out examples; the reported LLaMA-2-7B legal average is 44.65% versus 35.43% for random selection, without testing open-ended legal advice.
-
Legal Guidance in the Wild: How Users Seek Legal Help in Real-World LLM Conversations paper — Compares ModernBERT-base/large with a TF-IDF cascade on legal-query identification; it measures request classification, not the correctness of AI legal answers.
-
Beyond Recall Saturation: Scaling Hard-Negative Reranker Training for Industrial Legal Retrieval paper — The 4B Stage2 reranker wins 239, loses 91, and ties 80 LLM-judged operational-query comparisons: 72.4% excluding ties, 58.3% of all 410; these are evidence-list preferences, not legal-answer accuracy.
-
Machina Pauliana: A neuro-symbolic approach to the actio pauliana in Dutch insolvency law paper — Separates segmentation/label evaluation on 100 judgments from end-to-end criteria-and-verdict evaluation on ten cases; improved citation grounding is not a demonstrated accuracy gain from retrieval.
-
Narrow Decision Models: A System-1 Architecture for Calibrated, Typed Decisions paper — Tests option-order and question-independence properties, plus preliminary real-document classifier agreement; production “unknown” documents are excluded from training and full real-data results are deferred.
-
Policy Is Not Capability: Some Safety Benchmarks Measure Which Definition of "Unsafe" You Trained On paper — Copyright-score swings persist on 75 decontaminated prompts, but these cover only two templates; recipe-dependent seed variation and single-seed primary arms limit causal interpretation.
-
Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making paper — The completed uncalibrated 0.8B checkpoint reaches 67.86% accuracy on 4,300 legal decisions; aggregate calibration and preliminary 2B comparisons do not establish legal reliability or a controlled scaling effect.
-
Thunder-DeID+: Boundary-Aware Korean De-identification from Court Judgments to Broader Sensitive Documents paper — Tokenizer-independent full-judgment evaluation reports W. Strict Typed F1 rising from 0.76–0.77 to 0.87–0.88; the roughly 0.93 broader-document result requires 1,000 adaptation examples and is not zero-shot transfer.
-
In the Guise of Correction: How Framing and Cultural Adaptation Erode LLM Safety on Real-World Criminal Cases paper — Evaluates harmfulness and operational uplift in criminal-assistance scenarios; mean unsafe compliance rises from 28.9% for direct requests to 66.4% for correction-seeking prompts.
-
ContextLift: Breaking the Spurious Correlation Behind Conditional Emergent Misalignment paper — Evaluates legal-advice finetuning and LEDGAR contract-clause classification; cued broad-misalignment reduction does not measure legal-answer correctness, and LEDGAR supplies no liftable context.
-
Making every stage of medical device document assessment with large language models deterministic paper — All 14 tested commercial deployments vary on some regulatory-assessment requests; 12 change labels. End-to-end determinism depends on recorded serving and hardware settings, not sampling controls alone.
-
LegalRewardBench: Benchmarking Reward Models for Legal Preference Evaluation paper — Compares generative/discriminative and general/fine-tuned judges for legal preferences; general-purpose generative models are reported as more robust than the tested fine-tuned reward models.
-
PatTransBench: an ipc-stratified, multi-dimensional benchmark for patent machine translation evaluation paper — Jointly measures overlap, semantic similarity, 38-language consistency and serving performance; aggregate scores can hide direction-specific quantization effects.
-
UDHRBench: Benchmarking Human Rights in Large Language Models paper — Expert annotation agreement is reported as kappa 0.801; an LLM judge reaches 89.8% weighted F1 for human-rights misalignment, with ambiguity assessed separately.
-
RegExtract: A Benchmark for PFAS Regulatory Information Extraction with LLMs paper — Provenance-aware multi-field extraction evaluation; document identification remains difficult and field-specific few-shot examples improve extraction.
-
The Constitution Speaks Nheengatu: An Open MT System and Reproducible Corpus Pipeline for the Amazonian Língua Geral paper — Article-disjoint constitutional tests and a frozen legal-term lexicon expose limits hidden by aggregate translation scores; speech-domain gains are reported separately.
-
Enhancing Legal Reasoning in Pre-trained Language Models via Knowledge Graph-Guided Multi-Task Pre-training paper — Evaluates synthetic StVO QA, driving-regulation transfer and LexGLUE; the KG uses LLM-translated law and unquantified embedding/translation noise. Reported LexGLUE gains compare against published baselines.
-
Verifiable Retrieval for Legal Reasoning through Structure-driven Knowledge Graph paper — Depth-stratified citation prefix matching and token-efficiency recall; unspecified deeper citation levels count as implicit matches, and no generation or expert legal-sufficiency evaluation is established.
-
AMLworld-Compact: Importance-Weighted Downsampling for Cost-Effective LLM Evaluation and Error Diagnosis paper — Evaluates seven LLMs on weighted illicit-transaction detection; the evaluated prompts do not consistently identify/include the target transaction, so errors cannot be attributed solely to reasoning. F1 is a ratio estimate, not an unbiased confusion-count estimate.
-
APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces paper — Twelve models evaluated once per attack instance; task utility and attack success are separate outcomes. Judge validation uses 60 cases from one human annotator, and run-to-run variation is unmeasured.
-
Benchmarking Embedding Geometries and Clustering Algorithms for Unsupervised Patent Claim Classification paper — Compares embedding geometry on patent classification and cross-CPC retrieval; the VAE trails simpler untrained baselines on classification agreement, and independent expert relevance assessment remains future work.
-
Cited but Not Consulted: A Counterfactual Audit of Legal Chain-of-Thought Faithfulness paper — Evaluates seven open-weight models on judicial and contract tasks using authority swaps and latent verdict tracking; model-ranking uncertainty and a best-effort legal-model reproduction limit broad conclusions.
-
Verdict Without the Rule: Diagnosing and Auditing Regulatory Rule Sensitivity in LLM Compliance Systems paper — Behavioral OCS and representational ICS-delta audit five models across 20 regulatory/platform-policy domains; standard prompts truncate rules at 300 characters, and changed verdicts are not certified legally correct.
-
The "10th Juror": Open-Set Standpoint Screening for Bureaucratic Bias Detection paper — Reports 0.880 F1 on DGDB and held-out-category recovery; governance metrics and model-verified rewrites are not substitutes for independent legal review.
-
When Citations Mislead? A Claim-Level Benchmark for Legal Hallucination Detection paper — Three-way claim verification with full opinion text; high aggregate accuracy can coexist with false support for unsupported claims, and the benchmark is not an end-to-end legal RAG evaluation.
-
InsClaimBench: Benchmarking Insurance Claim Adjudication Across the Decision Chain paper — Evaluates contract eligibility, coverage/exclusions and payout consistency under factual changes; independently elicited rule/module judgments do not directly reveal an internal causal reasoning chain.
-
Right Number, Wrong State? Measuring Cross-Jurisdiction Substitution in LLM Recall of State Policy paper — Closed-book Medicaid eligibility-threshold QA across 51 jurisdictions; same-state historical values and reporting conventions must be ruled out before attributing errors to other states. Matching another state's number does not establish a knowledge-binding mechanism; data release is promised, not verified here.
-
Verdicts Without Annotated Evidence: Rejection Sampling or Label-Only Post-Training for Evidence Recovery? paper — Separates contract-entailment accuracy, evidence-span F1 and verbatim citations; the between-method span-F1 gap is not statistically resolved, and public-corpus pretraining exposure is not ruled out.
-
In With the Old: Enhancing 'Classical' Document Automation with Generative AI paper — Tests 15 constructed examples across three legal-document scenarios; only 1/5 employment-template revisions meets the instructions. The authors' permissive hallucination definition is not a guarantee of legal validity.
-
Domain adaptation of Russian ModernBERT for long legal documents paper — Fixed-collection masked-token and legal-entity evaluation; masking intervals are not uncertainty across document collections.
-
Breaking Bureaucracy: Evaluating open-source LLMs for legal document review paper — Zero-shot legal NLI on contracts and wills, with invalid-output, temperature and uncertainty analyses.
-
Better Call Reward: Reward Hacking as Strategic Abstention in Legal Reasoning Models paper — 320 examples from 16 LegalBench tasks; separates answer commitment, conditional accuracy and citation plausibility.
-
Not Self-Decidable: LLMs Cannot Draw the Boundary of What an Agent Verifier Can Check paper — EU AI Act/FINRA verification-scope evaluation; proxy labels and synthetic enforcement records limit real-world guarantees.
-
InvestigationWorlds: An Agentic Environment for Legal Investigation paper — Tests whether agents distinguish court-adopted hypotheses from plausible alternatives after retrieving relevant evidence.
-
HakemBench: A Turkish Benchmark of Typed Decisions paper — Includes constitutional-right identification and court routing; legal labels are partly AI-derived.
-
Verifiable, Articulable, and Tacit Components of Preference paper — Includes law/policy tasks; patent prediction results were withdrawn due to label leakage.
-
Legal Research Bench: Measuring End-to-End Reliability in Long-Horizon Legal Research Agents paper
-
SCM-based Fairness and Faithful Explainability for Legal Document Classification paper
-
Legal text classification in Korean sexual offense cases: from traditional machine learning to large language models with XAI insights paper
-
ARCCS: An Automated Regulatory Compliance Checking System paper
-
ContextAdapt: Evaluating Contextual Adaptation and Value Alignment in LLMs paper
-
LegalPincite: Multi-level Legal Information Retrieval Dataset paper
-
LexReward: A Taxonomy-Driven Reward Framework for Legal Language Models paper
-
ViLegalExpert: A Large-Scale Benchmark for Vietnamese Legal Retrieval and Question Answering from Real-World Consultations paper
-
Is This Evidence Decision-Critical? Learning to Verify Rule-Governed Decisions paper
-
Can a Cacheable Decision Model Follow Rules? paper
-
Chinese-Jev: Bringing System One Model to Chinese-Language Tasks paper
-
Evaluating and Benchmarking the System One Model Jev paper
-
A Cheap Verifier is Good Enough: LLM Post-training is Robust to Erroneous Rewards paper
-
RGDT-Bench: Benchmarking LLM Reasoning for Rule-Governed Decisions and Their Justifications paper
-
Stale-Document Poisoning: When Outdated Retrieval Overrides Correct Model Answers paper
-
Automated Regulatory Compliance Question Answering in Financial Services with Domain-Adapted Retrieval-Augmented Generation paper
-
Same Scores, Different Decisions: Evaluating JEV and Language Models for Legal Document Understanding paper
-
Cross-Lingual Legal QA for Vietnamese Labour Law: Retrieval, Translation, and Verifier-Guided Correction paper
-
Automated Extraction of Records of Processing Activities (RoPA) Using Hybrid RAG and Locally Deployed Large Language Models paper
-
Classifying Interpretive Canons at the Sentence Level: A Benchmark from the German Federal Constitutional Court paper
-
Mining Legal Arguments in U.S. Corporate Case Law paper
-
ABAI at COLIEE 2026 Task 1: Multi-Stage Retrieval with GraphRAG-Enhanced Meta-Learning, and a Post-Hoc Study of the Cross-Validation-to-Test Gap paper
-
GRACE: Grounded Adversarial Reasoning over Canadian Law paper
-
ChemCLIR-Bench: Benchmarking Cross-Lingual Information Retrieval in Multilingual Chemical Patents paper
-
Directing large language models to follow the letter or spirit of the law paper
-
Beyond Accuracy and Surface Fluency: Risk-Sensitive Evaluation of LLMs for Legal Clause Generation paper
-
UK-PRBENCH: A Paragraph-Level Precedent Retrieval Benchmark for United Kingdom Case Law paper
-
Benchmarking LLM Compliance with China AI Generated Content Regulations paper
-
Before the Arrest: Benchmarking LLMs on Criminal Profiling from Incomplete Evidence paper
-
How Much is a Human Right Worth? ECtHR-NPD: A Benchmark for Predicting Non-Pecuniary Damage Awards paper
-
Linguistic Triggers of Gender and Racial Bias in Open-Weight LLMs Applied to Recruitment paper
-
Legal LLM Hallucination Should Be Evaluated as Failure of Legal Warrant paper
-
Nepali Legal Expertise through Generative and Extractive Pre-trained Transformers (NepLEGiT) paper
-
NepKANUN: A RAG-Based Nepali Legal Assistant paper
-
Few-Shot Degradation Is Not What It Seems: Behavioral Evidence, Representation Analysis, and a Random-Text Control Across 12 Models, 2 Tasks, and 2 Architectures paper
-
CiteGuard-RAG: A Validation-Centered AI System for Evidence-Grounded Question Answering paper
-
Building Legal Reward Models for Grounding and Abstention paper
-
Algorithm Validation as a Policy Audit: Evidence from Race-blind Charging paper
-
Tasks over Application Manuals: Revealing Gaps in Long-Horizon Procedural Reasoning for Language Models paper
-
ProMediConv: Benchmarking Proactive Conversational Agents in Legal Dispute Mediation paper
-
BodyCam-VQA: Enhanced Body-Worn Camera Video Captioning via Multimodal Reasoning and Probe Question Generation paper
-
GANDR: Claim Auditing for Verifiable Legal Answer Generation paper
-
When Does Defendant Statement Matter? A Study of Bias and Persuasion in LLM-Simulated Jurors paper
-
MUCnoHARM@GermEval Shared Task 2026: Retrieval-based In-Context Learning for Defamatory Offences, and Where It Falls Short paper
-
Reading a Legal Question Word by Word: Embedding Trajectories of 2,144 Vietnamese Legal Headlines paper
-
FramingQA: Does the Question Shape the Answer? Measuring the Compositional Framing Effect paper
-
Solving versus Verifying: Catching Contradictions in Tax Reasoning Systems paper
-
LexFlip: A Dissociation Diagnostic for Legal Meaning Preservation Metrics paper
-
KhatianDoc: A Human-Verified Benchmark Diagnosing Multimodal LLM Failure on Bengali Legal Land Records paper
-
LexIssue: Benchmarking Legal Issue Identification in Chinese Civil Litigation paper
-
Rent-a-RAG: Embedding-Space Watermarks for Auditing Third-Party RAG paper
-
OBJECTION! Lawyer Agents Mitigate Guilty Bias in Legal Judgment Prediction paper
-
Polish ModernBERT: The Long and Short of Polish Language Understanding paper
-
Annotated Surrogate Retrieval for Polish Statutory Law paper
-
Do Small Models Use the Law You Give Them? Measuring Context Use on a Bilingual Bangladesh Legal Benchmark paper
-
Cloud and On-Premises Deployment of Uzbek Legal RAG via Targeted Retriever Fine-Tuning paper
-
Cross Lingual Transfer in Tulu Legal Comprehension: Script-Dependent Improvement and RAG-Induced Knowledge Conflict paper
-
RegDivergence-101: An LLM Benchmark for Cross-Jurisdiction Regulatory Contradiction Detection in Life Sciences paper
-
GreenLeaf Law Embed Tiny: A Compact Embedding Model for Legal Domain Retrieval paper
-
Figurative Justice: Detecting metaphors in Hindi judgements with qualitative assessment and transformers paper
-
Modeling Claim Dependency Structure for Patent Litigation Prediction with Graph Attention Networks paper
-
Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards? paper
-
Generative Gap Filling paper
-
Benchmarking Patent Drafting from Inventor-Style Disclosures paper
-
Ansari: A Retrieval-Grounded Islamic AI Assistant -- Architecture, Deployment, and Lessons from 140,000 Conversations paper
-
ContractScrub: A benchmark for final review of legal contracts paper
-
ChildSafeAds Shared Task 2026: Commercial Content in Child-Facing YouTube Videos paper
-
GreekBarRetrieval: A Benchmark for Greek Statutory Retrieval paper
-
CoAL-RAG: A Complexity-Aware Legal Retrieval-Augmented Generation Method paper
-
Can LLMs Reason in a Legally Meaningful Manner? A Small-scale Study on European Court of Human Rights Cases paper
-
Calibrated Trust, Not Sharper Prediction: An Empirical Test of Uncertainty Fusion paper
-
Gated Against One Model, Open to the Next: Option-Only Solvability in Legal Multiple-Choice Benchmarks paper
-
Time as Structure: Temporal Dependency Graphs for Verifiable Deadline Computation over Legal Documents paper
-
How Much Do Legal RAG Systems Still Hallucinate? paper
-
On Measuring Semantic Preservation in Legal Ontology Learning paper
-
Why Do AI Agents Break Rules? How Framing, Context, and Social Signals Shape Compliance paper
-
Is this Citation on Point? paper
-
Evaluating Rational Contracting in Natural Language paper
-
Temporal Misgrounding in Legal RAG: A Versioned-Corpus Benchmark for French Tax Law paper
-
LexKairos: Benchmarking Legal Temporal Capabilities in LLMs paper
-
PolicyKG: An Agentic LLM Pipeline for Translating Institutional Policies into SHACL Knowledge Graphs paper
-
ANNOTARES: A Dataset for Extracting Logical Structures from German Statutory Texts paper
-
Knowing the Form, Not the Function: Automatically Auditing Answer--Authority Decoupling in Legal Benchmarks paper
-
CrossLex: A Source-Grounded Benchmark for Cross-Jurisdictional Legal Reasoning in Large Language Models paper
-
Semantics of Subterfuge: Benchmarking Legal Deception Detection Against General-domain State-of-the-Art paper
-
Credit Cards, Confusion, Computation, and Consequences: What Can We Uncover About Language Model Reasoning? paper
-
LegalCiteTrust: Benchmarking Citation Trustworthiness in Chinese Long-Form Legal Research Reports paper
-
Measuring Massive Multitask Chinese Understanding paper
-
LawBench: Benchmarking Legal Knowledge of Large Language Models github
-
Large Language Models are legal but they are not: Making the case for a powerful LegalLLM paper
-
Better Call GPT, Comparing Large Language Models Against Lawyers paper
-
Evaluating GPT-3.5's Awareness and Summarization Abilities for European Constitutional Texts with Shared Topics paper
-
Evaluation Ethics of LLMs in Legal Domain paper
-
GPTs and Language Barrier: A Cross-Lingual Legal QA Examination paper
-
LawBench: Benchmarking Legal Knowledge of Large Language Models paper
-
LexEval: A Comprehensive Chinese Legal Benchmark for Evaluating Large Language Models paper
-
LegalAgentBench: Evaluating LLM Agents in Legal Domain paper
-
Measuring Faithfulness and Abstention: An Automated Pipeline for Evaluating LLM-Generated 3-ply Case-Based Legal Arguments paper
-
LegalEval-Q: A New Benchmark for The Quality Evaluation of LLM-Generated Legal Text paper
-
When Fairness Isn't Statistical: The Limits of Machine Learning in Evaluating Legal Reasoning paper
-
LexTime: A Benchmark for Temporal Ordering of Legal Events paper
-
LLM-based HSE Compliance Assessment: Benchmark, Performance, and Advancements paper
-
VLQA: The First Comprehensive, Large, and High-Quality Vietnamese Dataset for Legal Question Answering paper
-
The Massive Legal Embedding Benchmark (MLEB) paper
-
LexGenius: An Expert-Level Benchmark for Large Language Models in Legal General Intelligence paper
-
Legal-DC: Benchmarking Retrieval-Augmented Generation for Legal Documents paper
-
Adaptive Cost-Efficient Evaluation for Reliable Patent Claim Validation paper
-
From Benchmarking to Reasoning: A Dual-Aspect, Large-Scale Evaluation of LLMs on Vietnamese Legal Text paper
-
Exploiting LLM-as-a-Judge Disposition on Free Text Legal QA via Prompt Optimization paper
-
Tokenizer Fertility and Zero-Shot Performance of Foundation Models on Ukrainian Legal Text: A Comparative Study paper
-
Reasoners or Translators? Contamination-aware Evaluation and Neuro-Symbolic Robustness in Tax Law paper
-
Validate Your Authority: Benchmarking LLMs on Multi-Label Precedent Treatment Classification paper
-
LP-Eval: Rubric and Dataset for Measuring the Quality of Legal Proposition Generation paper
-
GradeLegal: Automated Grading for German Legal Cases paper
-
Asking For An Old Friend: Diagnosing and Mitigating Temporal Failure Modes in LLM-based Statutory Question Answering paper
-
A Two-Phase Stability Study of LLM Judges and Bar Council Examiners on Thai Bar-Exam Free-Form Essays paper
-
JudgmentBench: Comparing Rubric and Preference Evaluation for Quality Assessment paper
-
The Cases LJP Never Sees: Prosecution Decision Prediction for More Complete Criminal Liability Assessment paper
-
BenGER: Benchmarking LLM Systems on Subsumption-Based Legal Reasoning in German Law paper
-
Multi-Legal-Bench: Evaluating LLMs on Legal Reasoning Across Jurisdictions, Languages, and Legal Traditions paper
-
UA-Legal-Bench: A Benchmark for Evaluating Large Language Models on Ukrainian Legal Reasoning paper
-
LegalBench-BR: A Benchmark for Evaluating Large Language Models on Brazilian Legal Decision Classification paper
-
IndiaFinBench: An Evaluation Benchmark for Large Language Model Performance on Indian Financial Regulatory Text paper
-
RARE: Redundancy-Aware Retrieval Evaluation Framework for High-Similarity Corpora paper
-
ImmigrationQA: A Source-Grounded Dataset and Small-Model Adaptation for U.S. Immigration Law paper
-
CanLegalRAGBench: Evaluating Retrieval-Augmented Generation on Canadian Case Law paper
-
Citation Grounding: Detecting and Reducing LLM Citation Hallucinations via Legal Citation Graphs paper
-
Which Institutional Frameworks Do Chatbots Assume? Auditing Jurisdictional Defaults in Multilingual LLMs paper
-
On Wednesdays, We Ask Questions: Optimizing "Active Listening" in Automated Legal Triage and Referral paper
-
Re-Ranking Through an Attribution Lens for Citation Quality in Legal QA paper
-
EURO-5K: When Does Domain Pretraining Matter? Benchmarking Transformers for EU Reporting Obligation Extraction paper
-
HKJudge: A Legal Discourse-Annotated Corpus for Interpreting What Courts Find, How They Reason, and What They Rule paper
-
LexRubric: A Rubric-Guided Diagnostic Benchmark for Open-Ended Legal Tasks paper
-
From Statute to Control Flow: Span-Grounded Deontic Trees for Defeasible Scope Parsing paper
-
Retrieval Augmented Generation Framework for the Nepali Legal Domain Question Answering paper
-
REDACT: A Systematically Controlled Multilingual Benchmark for Personal Information Detection paper
-
DeXposure-Claw: An Agentic System for DeFi Risk Supervision paper
-
Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law Courts paper
-
Who Checks the Citations? Benchmarking Legal Hallucination Detection paper
-
Peeking Inside LLMs: Leveraging Internal Artifacts of LLMs for Enhancing Reliability in Legal Classification paper
-
Probing Stylistic Appropriation using Large Language Models: An Evaluation Framework for Copyright Infringement under EU Law paper
-
Do LLMs Fabricate Legal Citations? A Bilingual Benchmark on Saudi Data Protection Law and the GDPR paper
-
The Perplexity Trap: When Patent Law Makes Human Writing Look Like AI paper
The survey paper is shown in paper
Please cite the following papers as references if this repository helps you (*ゝω・)ﾉThanks!
@article{sun2023short,
title={A short survey of viewing large language models in legal aspect},
author={Sun, Zhongxiang},
journal={arXiv preprint arXiv:2303.09136v2},
year={2023}
}- AILQA: Evaluating AI-Driven Legal Question Answering Systems for the Indian Legal System paper
- Enabling Multilingual Privacy Policy Audits: Large-Scale Analysis of Spanish Mobile Apps paper
- Agentic Evaluation of Copyright Law Compliance paper
- From Obligation to Specification: A Survey on Validating EU AI Act Requirements in RE paper
- Reasoning Consensus: Structural Ensembling of LLM Reasoning via Weighted DAG Aggregation paper