Risk and Quality (R&Q) assurance in highly regulated industries requires constant navigation of complex regulatory frameworks, with employees handling numerous daily queries demanding accurate policy interpretation. Traditional methods relying on specialized experts create operational bottlenecks and limit scalability. We present a novel Retrieval Augmented Generation (RAG) system leveraging Large Language Models (LLMs), hybrid search and relevance boosting to enhance R&Q query processing. Evaluated on 124 expert-annotated real-world queries, our actively deployed system demonstrates substantial improvements over traditional RAG approaches. Additionally, we perform an extensive hyperparameter analysis to compare and evaluate multiple configuration setups, delivering valuable insights to practitioners.

This thesis advances the computational understanding and manipulation of text styles through three interconnected pillars: (1) Text Style Transfer (TST), which alters stylistic properties (e.g., sentiment, formality) while preserving content; (2)Authorship Attribution (AA), identifying the author of a text via stylistic fingerprints; and (3) Authorship V erification (A V), determining whether two texts share the same authorship. We address critical challenges in these areas by leveraging parameter-efficient adaptation of large language models (LLMs), contrastive disentanglement of stylistic features, and instruction-based fine-tuning for explainable verification. First, for TST, we conduct a comprehensive survey and reproducibility study of 19 state-of-the-art algorithms, establishing benchmarks across diverse datasets. Building on these insights, we introduce LLM-Adapters, a unified framework for parameter-efficient fine-tuning (PEFT) that enables cost-effective adaptation of LLMs for style-centric tasks. This culminates in Adapter-TST, a novel architecture that models multiple stylistic attributes (e.g., sentiment, tense) using lightweight neural adapters. Adapter-TST achieves superior performance in multi-attribute transfer and compositional editing while reducing computational costs by 80% compared to full fine-tuning. For AA, we propose ContrastDistAA, a contrastive learning framework that disentangles content and style features to address performance degradation under topic shifts. Our method advances both individual-level attribution and regional linguistic analysis, achieving state-of-the-art accuracy by isolating culturally influenced stylistic patterns.

The performance of large language models (LLMs) is closely tied to their training data, which can include copyrighted material or private information, raising legal and ethical concerns. Additionally, LLMs face criticism for dataset contamination and internalizing biases. To address these issues, the Pre-Training Data Detection (PDD) task was proposed to identify if specific data was included in an LLM's pre-training corpus. However, existing PDD methods often rely on superficial features like prediction confidence and loss, resulting in mediocre performance. To improve this, we introduce NA-PDD, a novel algorithm analyzing differential neuron activation patterns between training and non-training data in LLMs. This is based on the observation that these data types activate different neurons during LLM inference. We also introduce CCNewsPDD, a temporally unbiased benchmark employing rigorous data transformations to ensure consistent time distributions between training and non-training data. Our experiments demonstrate that NA-PDD significantly outperforms existing methods across three benchmarks and multiple LLMs.

Within the context of Natural Language Processing (NLP), fairness evaluation is often associated with the assessment of bias and reduction of associated harm. In this regard, the evaluation is usually carried out by using a benchmark dataset, for a task such as Question Answering, created for the measurement of bias in the model's predictions along various dimensions, including gender identity. In our work, we evaluate gender bias in German Large Language Models (LLMs) using the Bias Benchmark for Question Answering by Parrish et al. (2022) as a reference. Specifically, the templates in the gender identity subset of this English dataset were machine translated into German. The errors in the machine translated templates were then manually reviewed and corrected with the help of a language expert. We find that manual revision of the translation is crucial when creating datasets for gender bias evaluation because of the limitations of machine translation from English to a language such as German with grammatical gender. Our final dataset is comprised of two subsets: Subset-I, which consists of group terms related to gender identity, and Subset-II, where group terms are replaced with proper names. We evaluate several LLMs used for German NLP on this newly created dataset and report the accuracy and bias scores. The results show that all models exhibit bias, both along and against existing social stereotypes.

Scaling LLM vocabulary is often used to reduce input sequence length and alleviate attention's quadratic cost. Yet, current LLM architectures impose a critical bottleneck to this procedure: the output projection layer scales linearly with vocabulary size, rendering substantial expansion impractical. We propose SpeLLM, a method that decouples input and output vocabularies by predicting character-level strings through multiple output heads. In SpeLLM, each of the $k$ linear heads predicts a single character simultaneously, enabling the model to represent a much larger output space using smaller, independent linear heads. We present a self-distillation approach for converting a standard LLM to a SpeLLM. Our experiments with four pre-trained LLMs show their SpeLLM variants achieve competitive performance on downstream tasks while reducing runtime by 5.1% on average across models. Our approach provides a potential avenue for reducing LLM costs, while increasing support for underrepresented languages and domains.

Perovskite solar cells (PSCs) have rapidly emerged as a leading contender in next-generation photovoltaic technologies, owing to their exceptional power conversion efficiencies and advantageous material properties. Despite these advances, challenges such as long-term stability, environmental sustainability, and scalable manufacturing continue to hinder their commercialization. Precursor additive engineering has shown promise in addressing these issues by enhancing both the performance and durability of PSCs. However, the explosive growth of scientific literature and the complex interplay of materials, processes, and device architectures make it increasingly difficult for researchers to efficiently access, organize, and utilize domain knowledge in this rapidly evolving field. To address this gap, we introduce Perovskite-R1, a specialized large language model (LLM) with advanced reasoning capabilities tailored for the discovery and design of PSC precursor additives. By systematically mining and curating 1,232 high-quality scientific publications and integrating a comprehensive library of 33,269 candidate materials, we constructed a domain-specific instruction-tuning dataset using automated question-answer generation and chain-of-thought reasoning. Fine-tuning the QwQ-32B model on this dataset resulted in Perovskite-R1, which can intelligently synthesize literature insights and generate innovative and practical solutions for defect passivation and the selection of precursor additives. Experimental validation of several model-proposed strategies confirms their effectiveness in improving material stability and performance. Our work demonstrates the potential of domain-adapted LLMs in accelerating materials discovery and provides a closed-loop framework for intelligent, data-driven advancements in perovskite photovoltaic research.

The emergence of deep research systems presents significant capabilities in problem-solving, extending from basic queries to sophisticated research tasks. However, existing benchmarks primarily evaluate these systems as agents for web retrieval and report generation, overlooking their potential to discover novel insights on the frontiers of scientific research. To address this gap, we introduce ResearcherBench, the first benchmark focused on evaluating the capabilities of these advanced, agentic systems - which we refer to as Deep AI Research Systems (DARS) - on frontier AI scientific questions. We compiled a dataset of 65 research questions expertly selected from real-world scientific scenarios such as laboratory discussions and interviews, spanning 35 different AI subjects and categorized into three types: technical details, literature review, and open consulting. Our dual evaluation framework combines rubric assessment, which uses expert-designed criteria to evaluate insight quality, with factual assessment, which measures citation accuracy (faithfulness) and coverage (groundedness). We evaluated several leading commercial DARS and baseline systems. Results show that OpenAI Deep Research and Gemini Deep Research significantly outperform other systems, with particular strength in open-ended consulting questions. Such capabilities represent a meaningful step toward AI self-improvement, aligning with the vision of ASI for AI. We open-source ResearcherBench to provide a standardized platform for promoting the development of next-generation AI research assistants, hoping to foster a new perspective in AI research evaluation for a novel pattern of scientific collaboration: https://github.com/GAIR-NLP/ResearcherBench.

--Although vision transformers (ViT) have shown remarkable success in various vision tasks, their computationally expensive self-attention hinder their deployment on resource-constrained devices. T oken reduction, which discards less important tokens during forward propagation, has been proposed to enhance the efficiency of transformer models. However, existing methods handle unimportant tokens irreversibly, preventing their reuse in subsequent blocks. Considering that transformers focus on different information among blocks, tokens reduced in early blocks might be useful later . Furthermore, to adapt transformer models for resource-constrained devices, it is crucial to strike a balance between model performance and computational overhead. T o address these challenges, in this paper, we introduce a novel T oken Freezing and Reusing (T oFe) framework, where we identify important tokens at each stage and temporarily freeze the unimportant ones, allowing their lagged reusing at a later stage. Specifically, we design a prediction module for token identification and an approximate module for recovery of the frozen tokens. By jointly optimizing with the backbone through computation budget-aware end-to-end training, T oFe can adaptively process the necessary tokens at each block, thereby reducing computational cost while maintaining performance. Extensive experiments demonstrate that T oFe reduces the computational cost of L V-ViT model by 50% with less than 2% drop in T op-1 accuracy, achieving a better trade-off between performance and complexity compared to state-of-the-art methods. Large-scale pre-trained vision transformer (ViT) models [37] have achieved remarkable progress in the field of vision tasks.

Fisheye cameras introduce significant distortion and pose unique challenges to object detection models trained on conventional datasets. In this work, we propose a data-centric pipeline that systematically improves detection performance by focusing on the key question of identifying the blind spots of the model. Through detailed error analysis, we identify critical edge-cases such as confusing class pairs, peripheral distortions, and underrepresented contexts. Then we directly address them through edge-case synthesis. W e fine-tuned an image generative model and guided it with carefully crafted prompts to produce images that replicate real-world failure modes. These synthetic images are pseudo-labeled using a high-quality detector and integrated into training. Our approach results in consistent performance gains, highlighting how deeply understanding data and selectively fixing its weaknesses can be impactful in specialized domains like fisheye object detection.

Large Language Models (LLMs) are increasingly used in circuit design tasks and have typically undergone multiple rounds of training. Both the trained models and their associated training data are considered confidential intellectual property (IP) and must be protected from exposure. Confidential Computing offers a promising solution to protect data and models through Trusted Execution Environments (TEEs). However, existing TEE implementations are not designed to support the resource-intensive nature of LLMs efficiently. In this work, we first present a comprehensive evaluation of the LLMs within a TEE-enabled confidential computing environment, specifically utilizing Intel Trust Domain Extensions (TDX). We constructed experiments on three environments: TEE-based, CPU-only, and CPU-GPU hybrid implementations, and evaluated their performance in terms of tokens per second. Our first observation is that distilled models, i.e., DeepSeek, surpass other models in performance due to their smaller parameters, making them suitable for resource-constrained devices. Also, in the quantized models such as 4-bit quantization (Q4) and 8-bit quantization (Q8), we observed a performance gain of up to 3x compared to FP16 models. Our findings indicate that for fewer parameter sets, such as DeepSeek-r1-1.5B, the TDX implementation outperforms the CPU version in executing computations within a secure environment. We further validate the results using a testbench designed for SoC design tasks. These validations demonstrate the potential of efficiently deploying lightweight LLMs on resource-constrained systems for semiconductor CAD applications.