legal tech developers

Abstract: Specialized prompt injection guards are deployed in front of document analysis systems, but no false positive rate has been published for any of them on legal contracts. We measure eleven open-weights guards on two corpora of legal contracts and find that existing guards struggle: document-level false positive rates are as high as 92.39%; two guards fall below chance on paired AUC; and a five-regex baseline is not Pareto-dominated by any guard in the panel. Controlled probes locate model-specific as well as general failure modes. Prompted LLMs resolve the tested failures, including a second threat category (unauthorized scope claims) on which every guard scores zero. A cascaded LLM solution, where the positives of a mid-tier model are verified by a frontier model, is 6.5x cheaper than using exclusively a frontier model and approaches its performance.

Abstract: Information Retrieval (IR) is much more than ad hoc web search: a large portion of the information seeking activities takes place in specific domains, often addressing work-related tasks. Examples are lawyers finding relevant prior cases for a legal case and academic researchers studying the background material for a paper. In Chapter 2 we outline the most common domain-specific tasks, datasets and applications. Domain-specific IR has a number of challenges that make set it apart from simple web search tasks: they often deal with long documents, long queries and long sessions with specific user needs. We discuss these in Chapter 3. We also address issues related to bias and fairness of models. In Chapter 4, we describe the models and methods developed for domain-specific tasks and data, addressing the specific challenges, including user aspects related to domain-specific IR tasks, dealing with queries, user models, and user interactions. We show that many of the methodological advances in domain-specific IR are transferable to other domains and even to open-domain web search. We conclude with the main takeaways and propose research directions for the near future.

Abstract: Model compression can facilitate the deployment of large language models in resource-constrained legal settings, but its effects on domain-specific performance may depend on the evaluation protocol. We investigate this issue using ArabLegalEval, a Saudi Arabic legal benchmark, under both with-context and no-context evaluation. The former provides supporting legal passages, while the latter assesses performance without external context. Removing the context produces accuracy gaps of up to 27 percentage points for the base model. We examine three stages: domain adaptation, quantization, and knowledge distillation across FANAR-9B, JAIS-8B, and Qwen2.5-7B. Four-bit quantization reduces model memory by 64-71%, with changes in no-context MCQ accuracy ranging from -3.2 to +1.6 percentage points. Distillation produces more variable results, with accuracy generally declining as student size decreases. A Qwen2.5-7B student distilled from FANAR achieves 76.6% no-context accuracy, compared with 77.2% for its teacher. Our findings suggest that no-context evaluation provides a useful complement to context-supported testing when assessing the effects of adaptation and compression on Arabic legal language models.

Abstract: As LLM-based agents develop, they have been increasingly applied in the legal domain. Legal Judgment Prediction~(LJP) requires models to conduct multi-step reasoning and external legal evidence retrieval. Existing legal search agents' performance is often constrained by the quality of retrieval queries. They often formulate queries directly from the initial user request, overlooking latent legal issues and legally salient details. To address this issue, we propose \emph{LegalEvolver}, a query optimization framework specifically designed to generate legally informative queries for LJP-oriented retrieval. Sampling high-quality optimized queries and their corresponding search trajectories is challenging, as relying solely on unguided high-temperature rollouts is both inefficient and constrained by the model's initial capabilities. We therefore heuristically optimize queries as seeds, improving the efficiency of trajectory generation. Furthermore, we summarize rollout experience into rubrics and use them to guide multi-round iterative self-evolution. Iteratively performing rubric optimization and rubric-guided RL enables the model to surpass its initial performance ceiling. End-to-end performance on extensive benchmarks demonstrates the effectiveness of query optimization on LJP. Comparisons with direct SFT followed by RL demonstrate the effectiveness of rubric guidance. Evolving rubrics further outperform fixed rubrics, confirming the benefits of self-evolution.

Abstract: Legal judgment prediction models may retain a guilty verdict after a legally decisive condition has been altered such that the specified charge is no longer established. We study this guilty bias via selective repair over matched charge-established and charge-not-established cases, where the unsupported verdict should change while its supported counterpart remains stable. We construct CAIL-MC, a legally reviewed dataset of 11,478 matched pairs covering 261 Chinese criminal charges. Held-out activation interventions show that contrasts learned from correctly resolved pairs transfer to false-guilty cases, but target-reaching interventions can also damage matched supported-guilty decisions. Motivated by this observation, we introduce PROSE, which reconstructs case-specific targets from a training-derived residual geometry, aligns them with a local margin-descent signal, attenuates supported-guilty-sensitive components, and realizes bounded targets through a preservation-regularized low-rank parameter update. Across Qwen3-8B and GLM4-9B, PROSE achieves stronger pairwise safe repair and evidence sensitivity, while preserving standard single-pass inference.

Abstract: Law changes over time. A court decides a case under the law in force on that day, and legal QA benchmarks grade answers against today's law. We checked 156,650 decisions of four Thai courts. Up to 67% of them cite law that has since been amended or repealed. We introduce NitiBench+, the sequel of our NitiBench, a Thai legal QA benchmark with 1,670 items across 17 domains, built from decisions of three courts. For each cited section it holds both the text in force on the decision date and the current text, and its retrieval corpus holds every dated version of every Thai statute section. Answers are graded on the Thai Bar Council rubric, the Bar-standard score, and on the overlap between the cited sections and the court's. The report has two parts. Part I measures the cost of the wrong text on 13 open-weight models. Reading the current text of amended law drops the Bar-standard score by 0.25 and missing repealed law drops it by 0.79. The penalty comes from rule changes, 0.45 on the 428 rule-changed items and none on the 331 wording-only items, and Citation F1 does not see it. The text in force is the fair setting. Part II benchmarks today's models on that setting. A public leaderboard ranks 26 models on a held-out test split of 500 items stratified by domain, with a validation split of the same shape released for public use, from memory alone, with the court's sections provided, and as a retrieval agent over the law in force, with the cost per item and the hallucinated citations. The best model scores 7.78 out of 10 and the best open-weight model 6.39. Without a search tool only five open-weight models recall any Thai law, and with it twenty of the 26 models gain and six lose. Leaderboard will take new models as they are released. Legal benchmarks must track dates, and on the right text today's models still leave a wide gap on Thai law.

Abstract: We present ICJ-Graph, a machine-readable resource connecting documents of the International Court of Justice (ICJ) to cited cases, legal provisions and United Nations resolutions. Court decisions, written pleadings and oral records form distinct document layers, preserving the distinction between the Court's reasoning and the parties' submissions. The collection contains 6,396 ICJ documents and 505 documents from other courts and legal bodies, together with 12,375 provisions and 20,291 UN resolutions. The data files contain text, metadata and embeddings. The accompanying pipeline lets users load the graph and generate embeddings for use in their own retrieval systems. To illustrate the resource's use, we evaluate five models on citation prediction for 199 decisions using documents available before each decision. Tools designed for citation research improve citation-prediction F1 over answering without tools for all five models across case, provision and resolution citations.

Abstract: When a language model decides refund or claim requests under written policies, the system can show it similar past cases. Each past case has two parts: its facts and the decision recorded for it. We separate the two. A model sees a policy, a request and three past cases, and we change either the recorded decisions or the facts, keeping the request fixed. We select requests that are close calls: deciding without past cases, DeepSeek V4.1 Flash approved each in some of four tries and denied it in others. When the past cases are near-copies of the request (its facts with a few details changed), switching their recorded decisions from DENY to APPROVE raises approval by about 30 to 55 percentage points in DeepSeek V4.1 Flash, GPT-5.6 Luna and Claude Haiku 5.5 on policies written for the study, and DeepSeek and Luna show the same on nine public policies paraphrased from real ones. The facts push the other way: shown without their recorded decisions, past cases whose facts support approval make these two models approve less often than past cases whose facts support denial. One reading is that the request looks weaker next to them. This push is not just a correction toward the right answer: in an additional DeepSeek analysis, it was of similar estimated size on requests that human annotators judged approvable and on those they judged deniable. In a separate DeepSeek run on the public policies, when the past cases are near-copies of a different request, switching their decisions moves approval by 9 points instead of 43, while facts that support approval still lower it by 22 points against facts that support denial. Systems that retrieve past decisions should test the two parts separately.

Abstract: Large language models (LLMs) can improve task performance during learning, yet these gains do not necessarily reflect learners' ability to thereafter apply the knowledge and skills independently. We present AI Cosmos, an agentic infrastructure for situated learning that organizes learning around two complementary principles: scenario-based participation and cumulative assets of practice. AI Cosmos organizes learners and role-specialized agents within learning scenarios that integrate task-relevant knowledge and, when needed, executable workbenches. An asset-memory loop preserves the processes and outcomes of practice as reusable assets with explicit provenance and dependencies. We evaluate scenario-based interaction through two preregistered randomized controlled studies examining the acquisition of legal reasoning and programming competence, holding the same underlying LLM across conditions. AI Cosmos achieves greater learning gains than a conventional LLM in legal reasoning and higher mastery of unseen programming problems than an LLM-as-Teacher baseline.

Abstract: Reliable legal information retrieval must ground legal reasoning in authoritative statutory sources. In Arabic judicial settings, long factual narratives must be mapped to concise articles that express abstract legal rules, creating a narrative-to-legal-rule abstraction gap. Using a benchmark of 200 synthetic Saudi Arabic judicial cases and 381 articles from the Family Law and the Law of Evidence, we compare word-level and char-4 BM25, zero-shot dense encoders, per-law retrieval, dense-sparse reciprocal rank fusion (RRF), and a proposed Decompose-and-Refine (DaR) pipeline. In DaR, a large language model splits each narrative into disputed legal issues and generates, for each issue, three retrieval channels: a statute-style paraphrase, legal keywords, and a candidate taxonomy chapter. On the whole corpus, the instructed Qwen3-Embedding-8B dense encoder reaches an MRR@10 of 0.849, the best result of any non-decomposed system. Under per-law scoring, this advantage reflects Family Law only: restricted to Evidence articles, the same encoder drops to an MRR@10 of 0.087, and fusion with BM25 raises it only to 0.207. DaR changes this picture, raising Evidence MRR@10 to 0.696 and Family MRR@10 to 0.864. Without decomposition, the two laws favor different strategies; under DaR, the same retriever achieves the best MRR@10 on both.

Abstract: Multilingual multi-hop document question answering (MMHop-DocQA) requires locating and integrating evidence across complex layouts, document pages, and diverse languages. Existing approaches include single-pass processing and adaptive tool use, but learning to coordinate evidence acquisition, cross-page integration, and answer verification remains challenging. To address this challenge, we present DocSkill, an agentic framework with a learned skill-orchestration policy for adaptive evidence acquisition and reasoning. The policy selects, revisits, or skips skills according to the evolving evidence state. To construct supervision for skill orchestration, bounded Monte Carlo Tree Search (MCTS) explores alternative skill trajectories offline. These trajectories provide demonstrations for supervised fine-tuning (SFT) and preference pairs for Direct Preference Optimization (DPO). We further introduce MMHop-DocBench, a multilingual multi-hop document QA dataset with training and test splits, containing 60K QA pairs across 10 languages and eight document domains. Its HopGen pipeline builds hierarchical page-element graphs to derive traceable cross-page evidence paths and generate grounded QA pairs with page- and element-level evidence annotations. Experiments show that DocSkill with Qwen3-VL-8B achieves 41.0 accuracy and 49.2 F1 on MMHop-DocBench, exceeding the strongest evaluated baseline by 5.5 and 6.3 points, respectively, and achieving competitive performance on MMLongBench-Doc and DocBench.

Abstract: Reasoning-tuned and reward-trained models are often assumed to verify more reliably because they reason longer or receive correctness signals. We test this in evidence-sufficiency judgment: deciding whether available information actually supports a causal claim, a domain with no decidable correctness reward. Using a pre-registered, same-base, controlled knock-in design, we compare a boundary-keeping recipe (SFT then GRPO) with its own baseline, a scrambled-pairing placebo, and a general-purpose reasoning model under an action-pressure protocol with no evaluation invitation. All load-bearing outcomes are rated by two separate model-rating sessions masked to arm labels; automatic judges only trigger. The trained arm raises spontaneous correct refusal from 25% to 100% (40 per arm; Wilson 95% CI lower bound 91.2%, non-overlapping with the baseline interval [14.2%, 40.2%]), while the scrambled placebo collapses (5% spontaneous / 7.5% invited correct; 91.1% false refusal). The effect is boundary-specific rather than blanket refusal: automatic false refusal on 281 well-formed claims is 1.1% (Wilson upper bound 3.09%), and a targeted 18-item arm-(c) recheck (3 flagged plus 15 accepted) yields 0/18 and 1/18 false refusals in two sessions. The dissociation is the main result: final boundary-keeping is trainable without an inference-time verifier, while observed active numeric/calculation checking in the H3 assay remains near zero; SFT saturation makes the GRPO increment not separately identifiable. A pre-registered mechanism adjudication favors reward internalization of final-action status (H2, best-supported, not proven). The scope is a controlled 14B same-family study: we do not claim proof, cross-family generality, or independent human-expert validation; the rating sessions are model sessions, not human reviewers.

Abstract: Applying a legal provision frequently requires both structured rule-based reasoning and sensitivity to open-textured terms developed in case law. The actio pauliana of Article 42 of the Dutch Bankruptcy Act (Faillissementswet, Fw), which allows an insolvency practitioner to avoid transactions that prejudice creditors, requires both. We present a neuro-symbolic model of Article 42 Fw. Its symbolic core is written in Catala in a literate style that keeps statute and formalisation aligned and was developed through pair programming, with open-textured terms represented as boolean inputs where neural components couple to the symbolic solver. A neural routine structures a case's natural-language facts into boolean and date inputs. Neural subroutines resolve the open-textured terms through case-based reasoning over Dutch civil judgments, retrieved through a pipeline that filters on the requirement being resolved and expands along the procedural chain of a case. Segmenting and labelling those judgments attains high precision and recall against a manually annotated gold standard of 100 judgments. On ten hand-annotated cases the neuro-symbolic model handles most criteria reasonably but struggles with the knowledge requirement. Retrieval depth shows no clear quantitative effect, although retrieved judgments provide explicit grounding in prior case law.

Abstract: The Italian legislative process is documented in detailed parliamentary records hosted in institutional websites. However, much of this information is available only as unstructured text, limiting its systematic reuse. We applied a controlled, systematic LLM-based extraction procedure, validated against manually annotated records, to convert these texts into ordered and dated events including, e.g., bill introduction, committee examination, plenary examination, and approval. The resulting dataset reconstructs the procedural histories of all 5346 Italian laws enacted between 1985 and 2026, spanning parliamentary terms IX to XIX, for overall 101413 events. Each event records the activity, date, and chamber, with additional procedural details where available. The dataset supports research in the emerging fields of computational legal and legislative studies, potentially combined with process-oriented data science methods to examine procedural paths and legislative timing across parliamentary terms.

Abstract: Multimodal long-document question answering requires models to seek sparse evidence across pages and modalities, understand heterogeneous document content, and reason over the collected evidence. Even advanced vision-language models (VLMs) can overlook sparse evidence and struggle with cross-modal reasoning in a single pass. Existing document agents further introduce multi-step interaction, but some still rely on predefined workflows or coarse-grained actions, limiting the flexibility and granularity of document reasoning. We introduce DocPilot, an agentic document framework that supports expressive orchestration and makes complex document solving learnable by small language models (SLMs). DocPilot orchestrates evidence seeking through page-level navigation and fine-grained evidence localization, and integrates local visual inspection, code-based numerical reasoning, and cross-modal reasoning for multimodal understanding. To transfer these capabilities to small models, we propose curriculum-based agentic training that progresses from single-ability learning to adaptive orchestration through supervision from data-grounded reasoning traces and executable interaction trajectories. Across MMLongBench-Doc, DocBench, FinRAGBench-V, and MADQA, DocPilot achieves strong gains with frontier VLM backbones, while DocPilot-4B delivers competitive performance against substantially larger models, demonstrating strong effectiveness and generalization across diverse long-document reasoning settings. The code is available at https://github.com/HKUSTDial/DocPilot.

Abstract: Recent advancements in topic modeling have split into two main paths: efficient embedding-based methods that offer scalability but are often outperformed on standard clustering metrics, and Large Language Model (LLM) approaches that achieve state-of-the-art clustering performance but at the cost of prohibitive computational expense and poor scalability. This paper introduces TopicMerge, a novel multi-stage pipeline that bridges these approaches to achieve both state-of-the-art clustering performance and computational efficiency. TopicMerge first employs an efficient clustering pipeline to deliberately over-segment a document corpus into a large set of fine-grained, semantically pure micro-clusters, which are then given initial semantic labels. It then introduces a novel, iterative merging algorithm where an LLM, guided by local neighborhood context, arbitrates on consolidating candidate pairs. These approved merges are then selectively executed and semantically relabeled by the LLM, hierarchically building broader, more coherent topics. Extensive experiments on the Bills and Wiki datasets show that TopicMerge achieves state-of-the-art clustering performance while reducing required LLM calls by over 92% compared to document-level methods, offering a practical and scalable solution.

Abstract: We present PatMT2 (Patent Machine Translation, v2), a two-stage LoRA pipeline that adapts the Hy-MT2-7B dedicated translation model to four-language (Chinese/English/Japanese/Korean) patent translation. On six translation directions, the Stage-2 checkpoint reaches a macro-average BLEU of 43.39 (BF16) to 43.57 (AWQ INT4), improving over the Hy-MT2-7B base by up to +8.8 BLEU and also exceeding Qwen3-32B and several external systems on the same average. We then measure the cost of adaptation: probing 38 languages absent from the fine-tuning data, the model still produces correct, fluent output on 34, while four low-resource languages collapse into distinct, reproducible failure modes. We treat these as diagnostic cases rather than retention rates, and report their mechanisms without attributing them to a single cause. Paired bootstrap tests confirm the translation gains over all baselines (p < 0.001).

Abstract: Large language models are increasingly deployed in morally and legally sensitive contexts, yet little is known about how they internally represent distinctions between moral reasoning styles. We ask whether a model's differential response to rule-framed and consequence-framed moral questions reflects distinct internal structure or merely surface lexical sensitivity. We construct a 2$\times$2 factorial dataset crossing reasoning style (rule-based vs. consequence-based framing) with judgment polarity (permissive vs. prohibitive), drawn from three established moral reasoning benchmarks and evaluated on a cleanroom test set of novel phrasings that includes conflict pairs where the two framings predict opposite labels. Using linear probing and activation steering, we identify two orthogonal directions in the residual stream: Dir-A tracks reasoning style and Dir-B tracks judgment polarity, each near-blind to the other factor (cross-correlation $|r| < 0.04$). This double dissociation replicates across three instruction-tuned models from different developers at comparable proportional depth, and it is not knife-edged in layer: the directions are recoverable and causally effective throughout an extended band of the network. Dir-A can causally change judgments, and in conflicting scenarios, it selectively corrects errors in rule-based judgments without changing consequence-based judgments, suggesting that it does not simply make the model more likely to accept everything. Generalization beyond the training templates is partial: the direction transfers across benchmark sources, but its advantage over lexical baselines on novel phrasings is modest. We therefore frame the result as moral reasoning style sensitivity, a geometrically dissociable and causally effective structure, rather than internalized normative reasoning.

Abstract: Legal language understanding requires reasoning over semi-structured regulatory documents, yet standard pre-training objectives do not explicitly target this structural capability. Because current legal language models rely solely on masked language modeling, they treat statutory text as flat token sequences and ignore the underlying hierarchical constraints. To bridge this gap, we present a multi-task pre-training approach for German traffic law (StVO). Our approach combines three complementary objectives: Next Sentence Prediction (NSP), Knowledge Graph (KG) Guided Clustering (CLU), and Knowledge Graph (KG) Aware Paragraph Classification (CLS). This approach explicitly encodes structured legal knowledge directly into the model weights. This encoding enables zero-shot legal reasoning at inference time and removes the need for discrete graph traversal. We evaluate our approach across nine Transformer-based encoders and 324 experimental runs. On synthetic StVO question answering, our models demonstrate consistent relative improvements between 10.3% and 155.0% over baselines. Furthermore, relative to the established LexGLUE baselines (Chalkidis et al., 2022), our models improve over prior results on four of seven legal benchmarks. It also achieves consistent zero-shot transfer gains between 3.4% and 15.8% on Austrian driving regulations. These experimental results indicate that our training approach generalizes successfully beyond the initial training legal domain.

Abstract: Many works approach continual learning through the lens of infinite context windows. As an agent puts more observation into context (concretely the KV cache), compacting said context is akin to direct memory manipulation, without affecting the base model's weights. Many works pose KV compaction as an optimization problem: learn a smaller set of KV vectors that matches the behavior of the full KV cache. While this preserves base model behavior, optimizing through a frozen base model results in a highly nontrivial optimization problem with a brittle and flat loss landscape. In this paper, we characterize what makes these optimization problems difficult and demonstrate that a heavily simplified Perceiver-based architecture not only matches performance of a full Perceiver transformer in continuous context compaction, but outperforms baselines on compaction utility. Results are presented on MCQ tasks across Finance, Legal, Gutenberg, and Code.

Abstract: We introduce InvestigationWorlds, an agentic environment for legal investigation. We build on an underused artifact of U.S. civil litigation: the summary judgment motion. This motion relies upon a record composed of real evidence exhibits, and results in a court-adopted hypothesis that is treated as ground truth for the purposes of deciding the motion. Each environment is built from a real U.S. Federal Court case retrieved from Public Access to Court Electronic Records (PACER) and augmented by an attorney-validated generation pipeline that synthesizes role-tagged documents around the original record. The resulting corpus admits multiple coherent factual readings, only one of which matches the court-adopted hypothesis. Evaluating on 100 cases, we find agents often commit to incorrect hypotheses despite retrieving relevant evidence, struggling to distinguish the court-adopted hypothesis from alternative hypotheses.

Abstract: Patent search requires processing documents routinely exceeding tens of thousands of tokens. Most neural retrieval approaches operate on truncated inputs, limiting their effectiveness. Graph-based retrieval addresses this by representing each patent as a structured invention graph, but constructing these graphs relies on brittle rule-based parsers. We present the neural parser, which adapts biaffine attention from dependency parsing to predict invention graphs directly from patent text. Our local biaffine attention restricts pairwise scoring to a sliding window, reducing complexity from $O(n^2)$ to $O(n \cdot w)$. Since local and global scoring share the same weights, the model trains on short sequences and deploys on documents exceeding 40,000 tokens without retraining. Distilled from 1 million rule-parsed documents, it surpasses its teacher at 3$\times$ lower inference cost: neural graphs improve citation recall by 0.5% on short queries and 1.1% on full documents in a downstream Graph Transformer retrieval system.

Abstract: Rule-based reasoning, as in eligibility checks and contract reviews, requires language models to assess evidence against individual conditions and combine their judgments under explicit rules. Errors in evidence assessment can leave a decision unchanged, but misinterpreting or overlooking decision-critical evidence can reverse it. Identifying such evidence allows more capable models to focus on checking the corresponding condition judgments, supporting accurate and safe decisions. Recognizing the evidence's criticality requires understanding how evidence affects a condition judgment and how that judgment affects the decision. To achieve the goal, we propose a INTERvention-based imPACT learning framework (InterPact), which enables counterfactual verification of evidence criticality in rule-governed decisions. Specifically, its evidence intervention constructor generates training pairs for a propagation verifier by editing case facts with a frozen language model while holding rules and non-target conditions fixed. Human-reviewed labels record the resulting condition and decision changes, while complete state-to-decision mappings supervise consequences beyond the observed edit. During training, the verifier weights learned conditional decision predictions by evidence-based condition probabilities through a fixed composition operation, propagating decision-change supervision into the base model. At inference, the trained base model directly judges criticality from the original case and target evidence, without human or stronger-model supervision. On single-case evidence criticality verification over adapted rule-governed decision cases, InterPact achieves 68.28% accuracy, outperforming all six baselines. These results support learned decision sensitivity as a basis for prioritizing evidence checks.

Abstract: Certo is a small non-generative decision model (Qwen3-4B): it scores candidate actions from their text and returns a probability, instead of generating an answer. The accurate design reads the state, the rules, and each candidate together (a joint scorer), so cost grows with the menu. Independent encoding lets each candidate be encoded once and reused across states (about 5x cheaper at 77 candidates), but separates state from candidate. We ask how much rule-sensitivity survives that move, and whether it can be trained back. Four experiments on Certo: (1) the tested conversion to cacheable scoring loses rule-sensitivity (recall@1 1.00 -> 0.24) while the joint scorer holds 1.00, and a shortlist+rerank rescue fails; (2) targeted counterfactual supervision restores strong performance on held-out synthetic rule tasks (paraphrase, counterfactual, composition; reproducible across seeds), though we do not isolate whether predictions depend on the supplied rule; (3) on real rules the added benefit is not established -- after fixing a truncation confound, the joint scorer wins significantly on the short tier (0.861 vs 0.500) and directionally on the hard tier (0.655 vs 0.483, n=29); (4) a matched cross-domain real-prose mixture did not help and reduced contract accuracy (-9.3, -16.2 points). A cacheable encoder can be made rule-sensitive on its training distribution, but transfer to unseen-source real rules is not established; the joint scorer keeps an edge at the cost of caching.

Abstract: Large Language Models (LLMs) struggle with region-specific factual knowledge, particularly in financial regulation. While benchmarks such as CFinBench, provide broad coverage of financial knowledge in other regions, no comparable resource exists for European financial regulation. We close this gap by proposing an end-to-end pipeline for evaluating and improving LLMs on European financial regulatory question answering. Our dataset is grounded in the official Q&A corpora of the European Banking Authority (EBA) and the European Securities and Markets Authority (ESMA). We evaluate candidate answers with a pointwise LLM-as-a-Judge protocol using three judges from distinct model families, retain only unanimously judged pairs, and sharpen the rejected side through a taxonomy of regulatory failure modes : law swaps, article swaps, hallucinated citations, and hallucinated text. By fine-tuning on the resulting preference pairs exposes a divergence at the core of our findings: supervised fine-tuning attains the highest judge score of any method yet got the lowest rule-based Citation F1. SFT is producing longer answers with three times as many citations, most of them unsupported. DPO and GRPO offer the better trade-off with more concise answers and the highest citation F1 among fine-tuned models. Standard LLM-as-a-Judge evaluation rewards citation density as evidence of grounding and cannot detect when those citations are fabricated. Thus, we detect a blind spot that matters wherever answers must be verifiable. We release the benchmark and evaluation stack at https://github.com/auliakharis/FinRegQA-EU.

Abstract: Retrieval-augmented generation is widely used in professional writing, yet whether retrieval grounds revision or merely injects templates is rarely tested where "correct" has a definable meaning. Patent claim amendment supplies that signal: the examiner names the attacked limitation and cites prior art, providing per-case ground truth. We release three artifacts: (i) a corpus of 7,385 USPTO prosecution cases with XML-aligned pre/post claims, rejection, and cited prior art; (ii) a seven-probe battery comparing random and structural-match retrieval as two policies under a fixed prompt scaffold; (iii) a deterministic five-channel metric (C1-C3 and C5 in main, C4 supplementary) requiring no LLM evaluation. Across 9,600 pre-registered calls on four frontier LLMs (Claude Sonnet 4, Claude Haiku 4.5, GPT-5.4, GPT-4o-mini), no tested model exhibits detectable classical prior-injection behavior; retrieval effects are small and direction-inconsistent between random and structural retrieval, and the null is unchanged under a dense (semantic) retriever, across retrieval depths k in {1,3,5,10}, and under a paraphrase-sensitive grounding metric. Revision locality reveals a model-specific difference that the template channel misses. The four-cell taxonomy, which we treat as exploratory, leaves the prior-injector cell unoccupied.

Abstract: Legal document summarisation in low-resource languages presents significant challenges due to the scarcity of annotated data and the complexity of domain-specific terminology. This paper presents SinBrief, a hybrid abstractive summarisation framework for Sinhala legal documents that does not require human-annotated training data. The proposed framework combines domain-aware word graph construction with neural sentence scoring to generate abstractive summaries from Sinhala legal text. Five sentence scoring models are evaluated within the framework: mBert, Llama 3.1, Falcon 7B, Laser, and a continually pre-trained Llama model domain-adapted to Sinhala legal text. The framework is evaluated on a Sinhala legal corpus using reference-free metrics, including Coverage, Density, Compression Ratio, SummaC, and Self-BertScore. Experimental results demonstrate that SinBrief produces summaries with lower lexical overlap than extractive baselines while maintaining factual consistency, demonstrating the viability of hybrid, largely annotation-free abstractive summarisation for low-resource legal NLP tasks.

Abstract: Retrieval-augmented generation (RAG) is often used to address outdated knowledge by providing external evidence. But retrieval helps only when that evidence is still valid. We identify a temporal alignment failure, stale-document poisoning, in which outdated evidence makes a model wrong despite answering correctly without retrieval. We construct a benchmark of 317 verified knowledge reversals across medicine, law, software, and platform policy, grounded in dated official sources. Across 12 models, recent medical reversals are harder than long-established ones. More importantly, outdated retrieval flips 30% of Llama and 37% of Qwen answers even without instructions to trust the document; explicit follow instructions raise these rates to 66% and 75%. Across four open models and four domains, poisoning ranges from 17-91%, while matched up-to-date evidence is followed in 97-100% of trials. To isolate temporal applicability, we keep the historical evidence unchanged across 50 reversals and vary only the evaluation date. A clear pattern emerges: dates alone produce only modest adaptation, but when models are explicitly told when the old evidence stops applying, the larger models switch to the appropriate answer almost perfectly. Causal interventions confirm that this validity information directly shapes the final decision. The same internal components also support broader comparison tasks, suggesting that temporal applicability can recruit a general reasoning mechanism used for other comparisons. Finally, a fixed recency-aware hybrid re-ranker reduces poisoning by 4.6-10.0 points when dates are accurate, with gains that depend on reliable temporal metadata. Reliable RAG therefore requires selective trust: models must determine not only what retrieved evidence says, but whether it still applies.

Abstract: Financial institutions operate under dense, frequently amended rulebooks, and answering a compliance question correctly requires not only fluency but verifiable grounding in the authoritative text. Large language models are attractive for this task, yet the models that firms can realistically deploy on-premise are compact ones, and compact models hallucinate obligations. We study whether a carefully domain-adapted retrieval-augmented generation pipeline closes that gap. Our retriever is built in three stages on top of LegalBERT: entailment tuning that recasts question--passage matching as premise--hypothesis reconstruction, contrastive tuning with in-batch negatives, and score-level fusion with BM25. Our generator is a compact model (2B--12B parameters) served under 4-bit quantization, either prompted or adapted with retrieval-aware fine-tuning (RAFT) through LoRA. On ObliQA, a question-answering benchmark built from the Abu Dhabi Global Market rulebooks, the staged retriever raises Recall@10 from 0.256 to 0.774 and outperforms BM25 (0.678) and E5-large-v2 (0.758), the strongest general-purpose dense encoder we tested. RAFT-LoRA then improves the composite RePASs answer-quality score for every model we could adapt, with the largest gain on the weakest one. However, the adapted models do not transfer to Australian case-law questions, and a closed-book model that receives no passages at all scores within 0.011 RePASs of the full pipeline while producing answers that cite nothing and misstate obligations. The retrieval gain is therefore measured directly, the generation gain is a gain in RePASs rather than demonstrated grounding, and grounding itself requires an evaluation protocol that RePASs does not provide.

Abstract: Negation does not carry a uniform interpretation across domains. In legal, regulatory, and medical reasoning, the intended interpretation depends on the reading in force -- open- versus closed-world, two- versus three-valued, and credulous versus skeptical. We study which reading of negation large language models adopt by default and whether they can override that preference when a different reading is explicitly specified. To this end, we introduce NAFBench, a procedural generator of solver-certified instances spanning four semantic viewpoints: SLDNF, well-founded semantics (WFS), and credulous and skeptical reasoning under stable-model semantics. The generator emits ground normal logic programs with controlled depth, width, and cycle structure. Each program is solved under all four viewpoints using SWI-Prolog, a well-founded semantics solver, and clingo, yielding up to four divergent labels. The programs are then verbalized into natural language under multiple framings and rule orderings that leave the answer invariant. The results expose a consistent gap. Across open-source models, following a specified negation semantics remains unsolved: the strongest models score 59--74% across the four semantic viewpoints, while the weakest score 31--67%. All models are order-sensitive on more than half of logically identical rule shufflings, while the two weaker models frequently overcommit on well-founded "undefined." Two frontier models reach 100% on the main fixed-complexity evaluation set, and a third, o4-mini, is near-perfect, falling only to 81% on well-founded "undefined." Delegating reasoning to a solver, fine-tuning on certified traces, or forcing an explicit three-valued verdict each partly closes the gap.

Abstract: Legal contracts are structurally complex documents in which contradictions may emerge across distant and interconnected provisions. Although large language models (LLMs) improve legal language understanding, contradiction analysis remains a human-centered and evidence-grounded review task. We present ContraVis, a visual analytics system for human-in-the-loop contradiction analysis in legal contracts. The system models contracts as typed paragraph graphs that combine explicit contractual references with semantic relationships between paragraphs. This graph plays a dual role: it conditions LLM reasoning and serves as the interactive representation the analyst explores, keeping model context and human inspection aligned across coordinated views. In a controlled comparison, graph-conditioned reasoning recovered more injected contradictions than standalone LLM analysis as contract length grew, while surfacing additional candidates for analyst validation. A formative study with contract-domain lawyers indicated that in-context evidence comparison supported contradiction validation, and we distill design implications for evidence-grounded, LLM-assisted document review.

Abstract: Large language models are increasingly applied to high-risk domains such as law, yet complex legal reasoning remains limited by two structural challenges. First, existing RAG and GraphRAG methods emphasize lexical or semantic similarity while overlooking normative relations among legal provisions. Second, vanilla Chain-of-Thought prompting may generate plausible rationales without enforcing the normative structure of legal reasoning. To deal with the bottleneck of pipelines in the legal reasoning domain, we propose LEGO, a dual-module framework that synergizes Legal Expert GraphRAG and expert Chain-of-thought for complex legal reasoning. ExpertGraphRAG uses an expert-annotated civil code graph encoding these normative relations with a greedy normative-coverage retrieval algorithm to dynamically extract instance-specific provision subgraphs, while ExpertCoT organizes the retrieved provisions and case facts into structured Provision-Fact-Conclusion reasoning. With a Qwen3-8B backbone, LEGO achieves 40.53% exact-match accuracy on LawExamQA_Civil, outperforming the evaluated RAG and CoT baselines and performing comparably to the evaluated larger models, while remaining robust on multi-hop questions. It also achieves the best results among the evaluated baselines on the open-ended benchmarks. Ablation studies confirm the individual and complementary contributions of both modules, demonstrating LEGO's effectiveness in improving LLMs' complex legal reasoning ability. Code and dataset can be found in the link: https://github.com/BLK-WHT/LEGO

Abstract: Judicial reasoning remains challenging for large language models (LLMs) to analyze. This paper contributes a sentence-level benchmark for evaluating the ability of LLMs to classify interpretive canons as articulated by Larenz in the tradition of Savigny. Our contributions are threefold. First, we operationalize this conception of interpretation as classification criteria. Second, we provide a dataset of decisions of the German Federal Constitutional Court annotated at the sentence level. Third, we report baseline evaluations of four LLMs from three model families under expert hand-written prompts, compared against prompts optimized with Genetic-Pareto (GEPA). Mean F1 over the seven binary subtasks clusters between 70.4 and 79.2 across models, with grammatical interpretation usually the easiest canon to identify and systematic interpretation usually the hardest; under the tested configuration, GEPA-optimized prompts do not systematically outperform the hand-written ones, suggesting that the expert prompts provide a meaningful baseline.

Abstract: In this paper, we argue that acceptance and rejection are classical speech acts that express two opposite polar positions toward propositional content and can be systematically represented within a classically outlined bilateral formal model. Acceptance occupies the positive pole, represented by positive A, whereas rejection occupies the negative pole, represented by minus A. Their classical relationship is mediated by negation because acceptance of A corresponds to rejection of negation A, while rejection of A corresponds to acceptance of negation A, represented by positive A iff minus negation A and minus A iff positive negation A. Coherency and incoherency are introduced as supplementary notions to clarify this relationship and its connection with reflexivity. A further result of the manuscript is the methodical distinction between rejection and negation namely, rejection expresses negative polarity toward A, whereas negation operates upon propositional content.

Abstract: Know Your Client (KYC) onboarding in capital markets requires analysts to manually classify documents, extract structured data from heterogeneous sources, and validate compliance against complex regulatory policies. This process requires significant analyst time per client, with end-to-end onboarding often stretching to multiple weeks due to sequential handoffs. In this work, we analyze an on-boarding process and find that it comprises repeatable components well-suited to AI automation. We therefore propose a restructured workflow to be amenable to automation: we consolidate the traditional four-party process into two parties that share most of the work and can be automated together, eliminating intermediate handoffs that compound delays. To automate the remaining steps, we introduce Spectra, an AI-assisted document processing platform that combines a structured rules engine with LLM-based classification, extraction, and validation agents. The rules engine encodes compliance policy as a queryable database, enabling focused context injection that reduces token usage while improving extraction precision. Rather than a single monolithic prompt, the system decomposes document processing into isolated, auditable stages, each optimized independently and traceable to specific policy clauses. In evaluation on real KYC documents, Spectra achieves 100% classification accuracy and 89.4% extraction accuracy. Human review burden dropped by 96%.

Abstract: Large language models have shown strong performance across a range of legal tasks, but existing benchmarks rarely evaluate the ability to take and defend a legal position, reason under incomplete information, or synthesize multiple statutory provisions. This gap is particularly pronounced for Canadian law, which remains underrepresented in legal NLP. We introduce GRACE (Grounded Reasoning Adversarial Canadian LEgal examples), a dataset of 1,915 question-reasoning-answer instances grounded in Canadian federal legislation. GRACE covers three reasoning modes: adversarial advocacy, uncertainty, and applied reasoning. We develop a pipeline that partitions raw statutory text, generates scenario-based questions and reasoning, and filters examples through model-free citation verification and LLM-based quality auditing. As a proof of concept, we fine-tune CLeAR-4B (Canadian Legal Adversarial Reasoning), a lightweight model for grounded legal reasoning, and evaluate it against the unmodified Qwen3-4B base model in open- and closed-book settings. CLeAR-4B substantially improves agreement with teacher outputs and statutory citation behavior when the relevant act text is provided, while its grounding degrades sharply when the statute is withheld. These results suggest that GRACE can support the development of lightweight legal models that reason more effectively from supplied statutory text.

Abstract: RAG systems commonly retrieve a fixed number of documents (top-k) to ground generation, but this static approach is brittle: simple queries suffer over-retrieval (adding noise and cost) while complex queries are under-retrieved, causing recall failures that cascade into incorrect answers. Motivated by the question of how many documents must be retrieved to answer an arbitrary query reliably, we propose a practical, general framework for query-adaptive retrieval depth. Offline, we estimate per-query retrieval difficulty by measuring NDCG under the default retriever and deriving a query-specific "saturation" point k* from the NDCG-k curve. Because computing these signals online is expensive, we cluster a large set of queries in embedding space and summarize each cluster with a recommended retrieval depth that targets high coverage (e.g., \textasciitilde{}95\%) using a mean-plus-variance rule. At runtime, the system assigns an incoming query to a cluster and selects the corresponding top-k in constant time. Compared with post-retrieval confidence methods that rely on clustering retrieved documents, our approach is pre-retrieval and query-centric, making it robust in heterogeneous, case-like corpora and applicable across domains such as legal, healthcare, finance, and enterprise search. Finally, this framework has been tested in full-traffic queries that improved $F_1$ by over 36\% while reducing token usage by 14\% on low-complexity clusters without accuracy loss.

Abstract: In high-stakes domains such as legal practice, a language-model answer is only useful to the extent that a reader can verify each claim against the source the system cites. Current grounded-generation pipelines score the answer as a whole, so a correct conclusion can rest on fabricated or loosely matched citations and still score well. Closing this gap requires both a system built for per-claim verification and an evaluation that measures it. We introduce GANDR (Grounded ANswer DRafter), a two-agent system in which a Drafter writes an answer in a structured legal-reasoning format and a separate Critic, with the same view as a human verifier, audits each claim against its cited source and emits a per-claim audit trace on every round. We pair it with a strict correctness criterion requiring every citation to resolve to a passage the retriever returned. On a 185-item legal benchmark where all six systems share one backbone, one retrieval surface, and one citation instruction, GANDR ranks first on every primary metric, reaching 70.8% strict accuracy and leading the strongest baseline by 11.3 points (p<0.01). Reverting the protocol-anchored commit rule lowers strict accuracy by 22.7 points, and the strict lead stays positive on three further backbones, at +3.2 to +6.5 points. This lead traces to the Drafter configuration and the protocol-anchored commit, not to rewriting. Against two law-trained annotators the audit flags under-supported claims at F1 0.84 as a binary detector, while its four-way verdict labels agree only weakly and are advisory. Code is available upon request.

Abstract: With hate speech being ubiquitous online, automatic detection is crucial, in particular when it comes to criminally relevant social media posts. We study a variety of retrieval-based in-context learning (RetICL) strategies for detecting defamatory offences under §§ 185-187 StGB (the subject of GermEval 2026 Subtask 4). Few-shot prompting beats zero-shot, but retrieval-based approaches offer only marginal gains over random demonstrations, and even fall behind an optimised static set of demonstrations. Providing concrete legal knowledge helps, yet model choice outweighs every other system choice. Models over-predict criminal relevance while still missing 26-57% of criminally relevant posts, suiting them for triage rather than autonomous moderation.

Abstract: A dense retriever encodes a question as one vector, but the question arrives one word at a time. We read 2,144 held-out headlines from Thu Vien Phap Luat (Vietnamese legal library) word by word with Nemotron-3-Embed 8B/1B and Qwen3-Embedding 8B/0.6B, encoding 65,444 prefixes against 20,034 articles, plus every prefix of 3,438 sub-questions from 1,112 multi-question headlines and of 168 answers. (i) The gold article becomes rank 1 after a median of 6-7 content words in every encoder, before the interrogative frame is read, and stays there to the end in 78-85% of cases. (ii) In a multi-question headline the lock is inside the first sub-question 94-98% of the time; the second leaves rank unchanged in 89-95%; encoded alone, the second reaches rank 1 in 42-58% vs 91-96% for the first, at the same lock word (95-97% identical). (iii) Numbers, dates and instrument identifiers move the embedding twice as far as content words and four times as far as interrogative words; 72-78% of steps move toward the gold article, and the closing interrogative frame moves against that direction in 95-99% of headlines. (iv) Rank/cosine clustering yields six archetypes (instant, typical, unstable, late, never-locking) that differ by legal area and form (chi-squared p < 1e-8): real-estate and litigation headlines never lock on a number; environmental and accounting headlines do so a third of the time. (v) An answer read word by word retrieves its article after 8-16 words and addresses the sub-questions in order asked in 83-89% of cases. (vi) A word's step keeps a consistent direction across headlines (cosine 0.25-0.33; 0.44-0.60 for numbers); a preceding question rotates that step by about 60 degrees and a greeting by about 30 degrees; steps shrink as i^{-0.8}; and a two-question headline is within 12-17 degrees of a linear mix of its two questions. We call this a context-modulated additive walk.

Abstract: Legal case retrieval (LCR) is an essential tool for not only assisting legal practitioners to efficiently retrieve precedents but also enabling ordinary individuals to find valuable legal case information without relying on expensive professional legal services. Our previous work CaseLink demonstrated the effectiveness of using case to case graph structures to improve retrieval accuracy. However, its high computational cost during inference on large-scale legal databases limits its practical use in real-world settings. The main inefficiency comes from constructing test time graphs and computing pairwise term frequency similarities of cases. This process has O(n^2) complexity for n legal cases, making the runtime prohibitive as the number of candidates grows. For example, the retrieval time for one query on a database (COLIEE2022) with 1,563 candidate cases is more than 500 milliseconds, while the runtime would increase drastically to more than 3,500 seconds for a database (LeCaRDv2) with 55,192 candidate cases. To further enhance the retrieval performance while achieving a significant speed-up, in this extension paper, Cassette framework is proposed with a distillation strategy involving ranking objective and eigen-matching objective for an effective transfer of knowledge from a powerful and well-trained heavy teacher retriever to a lightweight and efficient hybrid student dual encoder. Specifically, the student query encoder is implemented as a multilayer perceptron model designed for fast online processing, whereas the student candidate encoder adopts a GNN architecture, suitable for an offline manner within the case database. Extensive experiments are conducted on three benchmark datasets, and the results verify the effectiveness of the ranking distillation while achieving high efficiency. The code has been released on https://github.com/yanran-tang/Cassette/.

Abstract: Multi-agent LLM pipelines increasingly span organisational boundaries; when a fault surfaces, someone must determine where it entered. The artifact available is rarely a full execution trace: it is the reports each agent filed, and a filed report can state a conclusion alongside its observations. Using a pre-registered, institutionally partitioned pipeline of six agents with process-level information boundaries, balanced defect injection and matched clean twins (345,600 requests per chain model, two models), we first report that our pre-registered hypothesis -- that collective responsibility framing degrades escalation with chain length -- is not supported. The layer nevertheless fails asymmetrically. It originates almost nothing: zero allegations across 7,996 clean episodes where every agent stayed silent. It filters upstream error poorly, naming an innocent party in 34.4% and 62.6% of clean episodes where an agent raised a false alarm. Conditional on no agent proposing the true origin (59.5% of episodes on one chain model), an auditor reading the reports recovers it in 4.1% of cases -- below a uniform guess (20%) and the best fixed-link accuser (31.0%) -- while reaching 60.3% from the raw documentation of the same episodes. Deleting one clause, the field carrying the agents' own conclusion, isolates the cause at constant observations: accuracy rises to 45.2% (+41.2 pp, 95% CI +35.3 to +46.9) and adherence collapses from 94.4% to 3.4%; where the suggestion was correct the same deletion instead costs accuracy, 70.5% to 55.7%. The harm replicates on two frontier auditors in four conditions out of four (+8.5 to +39.0 pp) and in a second domain (+47.7 and +61.1 pp), where the cost disappears. The net effect is governed by upstream reliability together with both conditional magnitudes. An accountability layer needs evidence sufficiently independent of the conclusions it verifies.