In common law systems, legal professionals such as lawyers and judges rely on precedents to build their arguments. As the volume of cases has grown massively over time, effectively retrieving prior cases has become essential. Prior case retrieval (PCR) is an information retrieval (IR) task that aims to automatically identify the most relevant court cases for a specific query from a large pool of potential candidates. While IR methods have seen several paradigm shifts over the last few years, the vast majority of PCR methods continue to rely on traditional IR methods, such as BM25. The state-of-the-art deep learning IR methods have not been successful in PCR due to two key challenges: i. Lengthy legal text limitation; when using the powerful BERT-based transformer models, there is a limit of input text lengths, which inevitably requires to shorten the input via truncation or division with a loss of legal context information. ii. Lack of legal training data; due to data privacy concerns, available PCR datasets are often limited in size, making it difficult to train deep learning-based models effectively. In this research, we address these challenges by leveraging LLM-based text embedders in PCR. LLM-based embedders support longer input lengths, and since we use them in an unsupervised manner, they do not require training data, addressing both challenges simultaneously. In this paper, we evaluate state-of-the-art LLM-based text embedders in four PCR benchmark datasets and show that they outperform BM25 and supervised transformer-based models.

Utilizing CPUs to serve large language models (LLMs) is a resource-friendly alternative to GPU serving. Existing CPU-based solutions ignore workload differences between the prefill and the decode phases of LLM inference, applying a static per-NUMA (Non-Uniform Memory Access) node model partition and utilizing vendor libraries for operator-level execution, which is suboptimal. We propose Sandwich, a hardware-centric CPU-based LLM serving engine that uses different execution plans for the prefill and decode phases and optimizes them separately. We evaluate Sandwich across diverse baselines and datasets on five CPU platforms, including x86 with AVX-2 and AVX-512, as well as ARM with NEON. Sandwich achieves an average 2.01x throughput improvement and 90% satisfactory time-to-first-token (TTFT) and time-per-output-token (TPOT) latencies with up to 3.40x lower requirements in single sequence serving, and significant improvement in Goodput in continuous-batching serving. The GEMM kernels generated by Sandwich outperform representative vendor kernels and other dynamic shape solutions, achieving performance comparable to static compilers with three orders of magnitude less kernel tuning costs.

The cognitive and reasoning abilities of large language models (LLMs) have enabled remarkable progress in natural language processing. However, their performance in interpreting structured data, especially in tabular formats, remains limited. Although benchmarks for English tabular data are widely available, Arabic is still underrepresented because of the limited availability of public resources and its unique language features. To address this gap, we present AraTable, a novel and comprehensive benchmark designed to evaluate the reasoning and understanding capabilities of LLMs when applied to Arabic tabular data. AraTable consists of various evaluation tasks, such as direct question answering, fact verification, and complex reasoning, involving a wide range of Arabic tabular sources. Our methodology follows a hybrid pipeline, where initial content is generated by LLMs and subsequently filtered and verified by human experts to ensure high dataset quality. Initial analyses using AraTable show that, while LLMs perform adequately on simpler tabular tasks such as direct question answering, they continue to face significant cognitive challenges when tasks require deeper reasoning and fact verification. This indicates that there are substantial opportunities for future work to improve performance on complex tabular reasoning tasks. We also propose a fully automated evaluation framework that uses a self-deliberation mechanism and achieves performance nearly identical to that of human judges. This research provides a valuable, publicly available resource and evaluation framework that can help accelerate the development of foundational models for processing and analysing Arabic structured data.

Opinions expressed in online finance-related textual data are having an increasingly profound impact on trading decisions and market movements. This trend highlights the vital role of sentiment analysis as a tool for quantifying the nature and strength of such opinions. With the rapid development of Generative AI (GenAI), supervised fine-tuned (SFT) large language models (LLMs) have become the de facto standard for financial sentiment analysis. However, the SFT paradigm can lead to memorization of the training data and often fails to generalize to unseen samples. This is a critical limitation in financial domains, where models must adapt to previously unobserved events and the nuanced, domain-specific language of finance. To this end, we introduce FinDPO, the first finance-specific LLM framework based on post-training human preference alignment via Direct Preference Optimization (DPO). The proposed FinDPO achieves state-of-the-art performance on standard sentiment classification benchmarks, outperforming existing supervised fine-tuned models by 11% on the average. Uniquely, the FinDPO framework enables the integration of a fine-tuned causal LLM into realistic portfolio strategies through a novel 'logit-to-score' conversion, which transforms discrete sentiment predictions into continuous, rankable sentiment scores (probabilities). In this way, simulations demonstrate that FinDPO is the first sentiment-based approach to maintain substantial positive returns of 67% annually and strong risk-adjusted performance, as indicated by a Sharpe ratio of 2.0, even under realistic transaction costs of 5 basis points (bps).

The evaluation of Large Language Models (LLMs) increasingly relies on other LLMs acting as judges. However, current evaluation paradigms typically yield a single score or ranking, answering which model is better but not why. While essential for benchmarking, these top-level scores obscure the specific, actionable reasons behind a model's performance. To bridge this gap, we introduce CLEAR, an interactive, open-source package for LLM-based error analysis. CLEAR first generates per-instance textual feedback, then it creates a set of system-level error issues, and quantifies the prevalence of each identified issue. Our package also provides users with an interactive dashboard that allows for a comprehensive error analysis through aggregate visualizations, applies interactive filters to isolate specific issues or score ranges, and drills down to the individual instances that exemplify a particular behavioral pattern. We demonstrate CLEAR analysis for RAG and Math benchmarks, and showcase its utility through a user case study.

Large language models (LLMs) have recently demonstrated remarkable progress in reasoning capabilities through reinforcement learning with verifiable rewards (RLVR). By leveraging simple rule-based rewards, RL effectively incentivizes LLMs to produce extended chain-of-thought (CoT) reasoning trajectories, progressively guiding them toward correct answers. However, existing approaches remain largely heuristic and intuition-driven, limiting the development of principled methodologies. In this paper, we present a theoretical characterization of LLM reasoning grounded in information bottleneck (IB) principle, introducing IB-aware reasoning optimization (IBRO), a framework that encourages reasoning trajectories to be both informative about the final correct answer and generalizable across diverse prompts. We derive a practical token-level surrogate objective and propose an efficient approximation, resulting in the lightweight IB regularization method. This technique integrates seamlessly into existing RL-based post-training frameworks without additional computational overhead, requiring only a one-line code modification. Empirically, we validate IB regularization across multiple mathematical reasoning benchmarks and RL algorithms, demonstrating consistent improvements in LLM reasoning performance.

Eye-tracking data reveals valuable insights into users' cognitive states but is difficult to analyze due to its structured, non-linguistic nature. While large language models (LLMs) excel at reasoning over text, they struggle with temporal and numerical data. This paper presents a multimodal human-AI collaborative framework designed to enhance cognitive pattern extraction from eye-tracking signals. The framework includes: (1) a multi-stage pipeline using horizontal and vertical segmentation alongside LLM reasoning to uncover latent gaze patterns; (2) an Expert-Model Co-Scoring Module that integrates expert judgment with LLM output to generate trust scores for behavioral interpretations; and (3) a hybrid anomaly detection module combining LSTM-based temporal modeling with LLM-driven semantic analysis. Our results across several LLMs and prompt strategies show improvements in consistency, interpretability, and performance, with up to 50% accuracy in difficulty prediction tasks. This approach offers a scalable, interpretable solution for cognitive modeling and has broad potential in adaptive learning, human-computer interaction, and educational analytics.

Clinician burnout has motivated the growing adoption of ambient medical scribes in the clinic. In this work, we introduce a custom-built ambient scribe application integrated into the EHR system at Included Health, a personalized all-in-one healthcare company offering telehealth services. The application uses Whisper for transcription and a modular in-context learning pipeline with GPT-4o to automatically generate SOAP notes and patient instructions. Testing on mock visit data shows that the notes generated by the application exceed the quality of expert-written notes as determined by an LLM-as-a-judge. The application has been widely adopted by the clinical practice, with over 540 clinicians at Included Health using the application at least once. 94% (n = 63) of surveyed clinicians report reduced cognitive load during visits and 97% (n = 66) report less documentation burden when using the application. Additionally, we show that post-processing notes with a fine-tuned BART model improves conciseness. These findings highlight the potential for AI systems to ease administrative burdens and support clinicians in delivering efficient, high-quality care.

We propose LingBench++, a linguistically-informed benchmark and reasoning framework designed to evaluate large language models (LLMs) on complex linguistic tasks inspired by the International Linguistics Olympiad (IOL). Unlike prior benchmarks that focus solely on final answer accuracy, LingBench++ provides structured reasoning traces, stepwise evaluation protocols, and rich typological metadata across over 90 low-resource and cross-cultural languages. We further develop a multi-agent architecture integrating grammatical knowledge retrieval, tool-augmented reasoning, and deliberate hypothesis testing. Through systematic comparisons of baseline and our proposed agentic models, we demonstrate that models equipped with external knowledge sources and iterative reasoning outperform single-pass approaches in both accuracy and interpretability. LingBench++ offers a comprehensive foundation for advancing linguistically grounded, culturally informed, and cognitively plausible reasoning in LLMs.

Online news outlets operate predominantly on an advertising-based revenue model, compelling journalists to create headlines that are often scandalous, intriguing, and provocative -- commonly referred to as clickbait. Automatic detection of clickbait headlines is essential for preserving information quality and reader trust in digital media and requires both contextual understanding and world knowledge. For this task, particularly in less-resourced languages, it remains unclear whether fine-tuned methods or in-context learning (ICL) yield better results. In this paper, we compile CLIC, a novel dataset for clickbait detection of Croatian news headlines spanning a 20-year period and encompassing mainstream and fringe outlets. We fine-tune the BERTić model on this task and compare its performance to LLM-based ICL methods with prompts both in Croatian and English. Finally, we analyze the linguistic properties of clickbait. We find that nearly half of the analyzed headlines contain clickbait, and that finetuned models deliver better results than general LLMs.