Large Language Models (LLMs) are increasingly being used in education, yet their correctness alone does not capture the quality, reliability, or pedagogical validity of their problem-solving behavior, especially in mathematics, where multistep logic, symbolic reasoning, and conceptual clarity are critical. Conventional evaluation methods largely focus on final answer accuracy and overlook the reasoning process. To address this gap, we introduce a novel interpretability framework for analyzing LLM-generated solutions using undergraduate calculus problems as a representative domain. Our approach combines reasoning flow extraction and decomposing solutions into semantically labeled operations and concepts with prompt ablation analysis to assess input salience and output stability. Using structured metrics such as reasoning complexity, phrase sensitivity, and robustness, we evaluated the model behavior on real Calculus I to III university exams. Our findings revealed that LLMs often produce syntactically fluent yet conceptually flawed solutions, with reasoning patterns sensitive to prompt phrasing and input variation. This framework enables fine-grained diagnosis of reasoning failures, supports curriculum alignment, and informs the design of interpretable AI-assisted feedback tools. This is the first study to offer a structured, quantitative, and pedagogically grounded framework for interpreting LLM reasoning in mathematics education, laying the foundation for the transparent and responsible deployment of AI in STEM learning environments.

The proliferation of Large Language Model (LLM) architectures presents a fundamental challenge: valuable, task-specific behaviors learned through fine-tuning methods like Low-Rank Adaptation (LoRA) are effectively trapped within their source model's architecture, herein referred to architectural lock-in. Existing transfer methods attempt to bridge this gap by aligning the static weight spaces of models, a brittle and indirect approach that relies on tenuous correlations between parameter geometries. This paper introduces a fundamentally different and more direct paradigm: the Cartridge Activation Space Transfer (CAST), a novel framework that liberates LoRA-encoded behaviors by learning a direct, nonlinear mapping between the activation manifolds, the geometric structures formed by the model's internal neuron activations, of two distinct LLM architectures. CAST treats a pre-trained LoRA as a frozen "behavioral kernel." It learns a set of lightweight, bidirectional projection heads that translate the target model's activation stream into the source model's latent space, apply the frozen kernel, and project the result back. This process, trained on a general text corpus without any task-specific data, effectively decouples the learned skill from the source architecture. We demonstrate that CAST enables true "zero-shot" translation of any standard LoRA adapter. Our experiments, including transfers between heterogeneous model families like Llama-2 and Mistral, show that CAST-translated adapters achieve 85-95\% of the performance of a LoRA fully retrained on the target model, quantitatively outperforming current weight-space transfer techniques and establishing a new state-of-the-art in model interoperability.

Large language models (LLMs) are entering legal workflows, yet we lack a jurisdiction-specific framework to assess their baseline competence therein. We use India's public legal examinations as a transparent proxy. Our multi-year benchmark assembles objective screens from top national and state exams and evaluates open and frontier LLMs under real-world exam conditions. To probe beyond multiple-choice questions, we also include a lawyer-graded, paired-blinded study of long-form answers from the Supreme Court's Advocate-on-Record exam. This is, to our knowledge, the first exam-grounded, India-specific yardstick for LLM court-readiness released with datasets and protocols. Our work shows that while frontier systems consistently clear historical cutoffs and often match or exceed recent top-scorer bands on objective exams, none surpasses the human topper on long-form reasoning. Grader notes converge on three reliability failure modes: procedural or format compliance, authority or citation discipline, and forum-appropriate voice and structure. These findings delineate where LLMs can assist (checks, cross-statute consistency, statute and precedent lookups) and where human leadership remains essential: forum-specific drafting and filing, procedural and relief strategy, reconciling authorities and exceptions, and ethical, accountable judgment.

Transformer-based large language models (LLMs) have achieved remarkable success, yet their standard attention mechanism incurs quadratic computation and memory costs with respect to sequence length, posing a major bottleneck for long-context training. Prior work tackles this challenge along two directions: (1) kernel-level optimizations, which accelerate dense and sparse attention operators; and (2) module-level strategies, often referred to as distributed attention or context parallel training, which scale attention across multiple devices. However, systematic evaluation still remains limited: operator-level comparisons are often incomplete, while context parallel strategies are typically framework-specific, with unclear performance analysis across contexts. To address these gaps, we propose a unified benchmark that integrates representative attention kernels and context parallel mechanisms with a modular and extensible interface for evaluation. The benchmark evaluates methods along two critical dimensions: (1) attention mask patterns, which strongly affect efficiency, scalability, and usability, and (2) sequence length and distributed scale, which determine performance under extreme long-context training. Through comprehensive experiments on the cluster of up to 96 GPUs, our benchmark enables reproducible comparisons, highlights method-specific trade-offs, and provides practical guidance for designing and deploying attention mechanisms in long-context LLM training.

Large Language Models (LLMs) are increasingly integrated into real-world applications, raising concerns about privacy, security and the need to remove undesirable knowledge. Machine Unlearning has emerged as a promising solution, yet faces two key challenges: (1) practical unlearning needs are often continuous and heterogeneous, and (2) they involve decentralized, sensitive data with asymmetric access. These factors result in inter-domain and intra-domain interference, which further amplifies the dilemma of unbalanced forgetting and retaining performance. In response, we propose a federated unlearning approach for LLMs that is scalable and privacy preserving. Our method decouples unlearning and retention via task-specific adapter learning and employs a hierarchical merging strategy to mitigate conflicting objectives and enables robust, adaptable unlearning updates. Comprehensive experiments on benchmarks of WMDP, MUSE, and TOFU showed that our approach effectively handles heterogeneous unlearning requests while maintaining strong LLM utility compared with baseline methods.

The exponential increase in scientific literature and online information necessitates efficient methods for extracting knowledge from textual data. Natural language processing (NLP) plays a crucial role in addressing this challenge, particularly in text classification tasks. While large language models (LLMs) have achieved remarkable success in NLP, their accuracy can suffer in domain-specific contexts due to specialized vocabulary, unique grammatical structures, and imbalanced data distributions. In this systematic literature review (SLR), we investigate the utilization of pre-trained language models (PLMs) for domain-specific text classification. We systematically review 41 articles published between 2018 and January 2024, adhering to the PRISMA statement (preferred reporting items for systematic reviews and meta-analyses). This review methodology involved rigorous inclusion criteria and a multi-step selection process employing AI-powered tools. We delve into the evolution of text classification techniques and differentiate between traditional and modern approaches. We emphasize transformer-based models and explore the challenges and considerations associated with using LLMs for domain-specific text classification. Furthermore, we categorize existing research based on various PLMs and propose a taxonomy of techniques used in the field. To validate our findings, we conducted a comparative experiment involving BERT, SciBERT, and BioBERT in biomedical sentence classification. Finally, we present a comparative study on the performance of LLMs in text classification tasks across different domains. In addition, we examine recent advancements in PLMs for domain-specific text classification and offer insights into future directions and limitations in this rapidly evolving domain.

With the rapid evolution of large language models (LLMs), the demand for automated, high-performance system kernels has emerged as a key enabler for accelerating development and deployment. We introduce TritonRL, a domain-specialized LLM for Triton kernel generation, trained with a novel training framework that enables robust and automated kernel synthesis. Unlike general-purpose programming languages, Triton kernel generation faces unique challenges due to data scarcity and incomplete evaluation criteria, vulnerable to reward hacking. Our approach addresses these challenges end-to-end by distilling Triton-specific knowledge through supervised fine-tuning on curated datasets, and further improving code quality via reinforcement learning (RL) with robust, verifiable rewards and hierarchical reward assignment. Our RL framework robustly detects reward hacking and guides both reasoning traces and code tokens through fine-grained verification and hierarchical reward decomposition, enabling the model to generate high-quality Triton kernels that can truly replace existing modules. With robust and fine-grained evaluation, our experiments on KernelBench demonstrate that TritonRL achieves state-of-the-art correctness and speedup, surpassing all other Triton-specific models and underscoring the effectiveness of our RL-based training paradigm.

Preprint repositories become central infrastructures for scholarly communication. Their expansion transforms how research is circulated and evaluated before journal publication. Generative large language models (LLMs) introduce a further potential disruption by altering how manuscripts are written. While speculation abounds, systematic evidence of whether and how LLMs reshape scientific publishing remains limited. This paper addresses the gap through a large-scale analysis of more than 2.1 million preprints spanning 2016--2025 (115 months) across four major repositories (i.e., arXiv, bioRxiv, medRxiv, SocArXiv). We introduce a multi-level analytical framework that integrates interrupted time-series models, collaboration and productivity metrics, linguistic profiling, and topic modeling to assess changes in volume, authorship, style, and disciplinary orientation. Our findings reveal that LLMs have accelerated submission and revision cycles, modestly increased linguistic complexity, and disproportionately expanded AI-related topics, while computationally intensive fields benefit more than others. These results show that LLMs act less as universal disruptors than as selective catalysts, amplifying existing strengths and widening disciplinary divides. By documenting these dynamics, the paper provides the first empirical foundation for evaluating the influence of generative AI on academic publishing and highlights the need for governance frameworks that preserve trust, fairness, and accountability in an AI-enabled research ecosystem.

Large language models (LLMs) achieve strong benchmark performance, yet user experiences remain inconsistent due to diverse preferences in style, tone, and reasoning mode. Nevertheless, existing alignment techniques such as reinforcement learning from human feedback (RLHF) or Direct Preference Optimization (DPO) largely optimize toward population-level averages and overlook individual variation. Naive personalization strategies like per-user fine-tuning are computationally prohibitive, and in-context approaches that prepend raw user signals often suffer from inefficiency and noise. To address these challenges, we propose POPI, a general framework that introduces a preference inference model to distill heterogeneous user signals into concise natural language summaries. These summaries act as transparent, compact, and transferable personalization representations that condition a shared generation model to produce personalized responses. POPI jointly optimizes both preference inference and personalized generation under a unified objective using reinforcement learning, ensuring summaries maximally encode useful preference information. Extensive experiments across four personalization benchmarks demonstrate that POPI consistently improves personalization accuracy while reducing context overhead by a large margin. Moreover, optimized summaries seamlessly transfer to frozen off-the-shelf LLMs, enabling plug-and-play personalization without weight updates.

Emotions guide human decisions, but whether large language models (LLMs) use emotion similarly remains unknown. We tested this using altruistic third-party punishment, where an observer incurs a personal cost to enforce fairness, a hallmark of human morality and often driven by negative emotion. In a large-scale comparison of 4,068 LLM agents with 1,159 adults across 796,100 decisions, LLMs used emotion to guide punishment, sometimes even more strongly than humans did: Unfairness elicited stronger negative emotion that led to more punishment; punishing unfairness produced more positive emotion than accepting; and critically, prompting self-reports of emotion causally increased punishment. However, mechanisms diverged: LLMs prioritized emotion over cost, enforcing norms in an almost all-or-none manner with reduced cost sensitivity, whereas humans balanced fairness and cost. Notably, reasoning models (o3-mini, DeepSeek-R1) were more cost-sensitive and closer to human behavior than foundation models (GPT-3.5, DeepSeek-V3), yet remained heavily emotion-driven. These findings provide the first causal evidence of emotion-guided moral decisions in LLMs and reveal deficits in cost calibration and nuanced fairness judgements, reminiscent of early-stage human responses. We propose that LLMs progress along a trajectory paralleling human development; future models should integrate emotion with context-sensitive reasoning to achieve human-like emotional intelligence.