Tag
Capital One's AI Foundations team introduces UCVE, a domain-agnostic framework that models compliance requirement spaces as a Contextual Requirement Graph (CRG) and performs retrieval via spectral reranking over a reinforcement-learning-weighted Laplacian. Evaluated across three compliance domains, UCVE achieves a 4.3% matched-K recall lift over flat cosine retrieval, outperforming strong retrieval and graph-based baselines.
论文提出 MARS-Gov，一个多智能体立场感知框架，用于荷兰政府文件中的官僚偏见检测、改写与验证，在 DGDB 上以 0.880 F1 达到新的 SOTA，并通过动态“第十位陪审员”实现开放集偏见目标识别。
A practitioner details a pipeline for distilling an LLM's document-extraction labels (entities, actions, values) into two 287M-parameter encoders — a GLiNER-style span tagger plus a multiple-choice model for entity resolution — to process ~5M court decisions, and asks for feedback on why it lags the teacher model.
LexReward introduces a taxonomy-driven reward framework for legal language models, evaluating responses along Style, Element, and Chain dimensions to build DPO preference data and reward models (LexRM) that improve legal reasoning performance.
LabourCrew is a multi-agent RAG framework designed for trustworthy adversarial deliberation and statutory reasoning over labour law, incorporating grounding mechanisms to control false-accept rates and improve answer relevancy.
This paper compares JEV with nine language models on ContractNLI, evaluating inference cost, response time, and correctness across various request configurations, finding that JEV has lower cost and response time while language models achieve higher baseline accuracy.
This paper presents a bilingual evaluation suite for cross-lingual legal question answering in Vietnamese labour law, assessing retrieval, translation, and verifier-guided correction to enhance answer faithfulness and address unsupported claims.
LexLattice introduces a multilingual extractive summarization method using neural cellular automata on document hierarchies, achieving state-of-the-art ROUGE scores on the EUR-Lex-Sum dataset with a compact model surpassing large instruction-tuned baselines.
This paper introduces a sentence-level benchmark for classifying interpretive canons in legal documents, evaluating LLMs on a dataset from the German Federal Constitutional Court.
This paper introduces an expert-annotated dataset of 42 U.S. federal tax opinions for legal argument mining, featuring functional labels and tree-structured arguments, and demonstrates classification and retrieval experiments.
This paper presents a risk-sensitive evaluation framework for LLM-generated contract clauses, focusing on legal failure modes and quality dimensions to assess risks beyond accuracy or fluency.
This paper demonstrates that removing stopwords in legal text analysis distorts doctrinal and ideological signals, recommending that stopwords be retained for measurement validity.
This paper evaluates retrieval-based in-context learning approaches for detecting criminally relevant hate speech in German social media posts, finding that few-shot prompting outperforms zero-shot but retrieval methods offer marginal gains, with models better suited for triage than autonomous moderation.
This paper describes building and deploying a retrieval-augmented generation (RAG) legal assistant for Uzbek in cloud and on-premises settings, creating domain benchmarks, and releasing the fine-tuned embedder UTE-1 to advance legal NLP in low-resource languages.
This paper introduces ImmigrationReason, a large-scale structured dataset of U.S. immigration appeals for legal reasoning research, addressing the gap in administrative adjudication data for NLP studies.
CoAL-RAG introduces a complexity-aware retrieval-augmented generation method for legal consultation that dynamically selects retrieval strategies based on question complexity, achieving significant improvements on Chinese and English legal benchmarks.
This paper demonstrates that legal multiple-choice benchmarks are vulnerable to option-only solvability, where models can answer correctly without the question, and that filtering based on one model's performance does not improve validity for other models.
Introduces LegalPincite, a large-scale legal information retrieval dataset built from CJEU judgments, featuring masked queries, full corpora, and paragraph-level citation annotations to enable multi-level retrieval evaluation.
This paper surveys and benchmarks NLP-based automatic deception detection in legal contexts, comparing fine-tuned transformers and seven LLMs with various prompting strategies across seven datasets. Results show domain sensitivity, with fine-tuned models excelling in general domains and few-shot LLMs competitive in low-resource legal settings.
This paper introduces BLAD, a curated multilingual dataset of 1,484 Bangladeshi legal acts spanning 1799 to 2025, with structured metadata for temporal and cross-lingual legal NLP research.