Community Question Answering (CQA) platforms can be deemed as important knowledge bases in community, but effectively leveraging historical interactions and domain knowledge in real-time remains a challenge. Existing methods often underutilize external knowledge, fail to incorporate dynamic historical QA context, or lack memory mechanisms suited for industrial deployment. We propose ComRAG, a retrieval-augmented generation framework for real-time industrial CQA that integrates static knowledge with dynamic historical QA pairs via a centroid-based memory mechanism designed for retrieval, generation, and efficient storage. Evaluated on three industrial CQA datasets, ComRAG consistently outperforms all baselines--achieving up to 25.9% improvement in vector similarity, reducing latency by 8.7% to 23.3%, and lowering chunk growth from 20.23% to 2.06% over iterations.

The push to compress and impart the proficiency of Large Language Models (LLMs) into more deployable and efficient Small Language Models (SLMs) has benefited from improvements in knowledge distillation (KD) techniques. These techniques allow a smaller student model to learn from a more capable and larger teacher model's responses. However, distillation often revolves around the student model merely copying the teacher's in-distribution responses, limiting its generalisability. This limitation is amplified on reasoning tasks and can be computationally expensive. In this study, we propose AdvDistill, a reward-guided dataset distillation framework. We utilise multiple generations (responses) from a teacher for each prompt and assign rewards based on rule-based verifiers. These varying and normally distributed rewards serve as weights when training student models. Our methods and their subsequent behavioural analysis demonstrate a significant improvement in student model performance for mathematical and complex reasoning tasks, showcasing the efficacy and benefits of incorporating a rewarding mechanism in dataset distillation processes.

Large Language Models (LLMs) have demonstrated remarkable capabilities at solving complex reasoning tasks with Chain-of-Thought (CoT) prompting, but their decision-making processes remain somewhat blackbox. We introduce textbfinverse reasoning, a novel paradigm enabling LLMs to decompose and explain their own reasoning chains post-hoc. Our approach, used in SAGE-nano, a 4-billion-parameter reasoning model, employs a metacognitive structure that reflects back via attention processes to identify major decision points and generate explanations of reasoning choices. While typical CoT approaches are directed towards forward reasoning generation, inverse reasoning provides insight into why specific reasoning chains were selected over others. Through thorough testing of logical reasoning puzzles, math problems and ethical dilemmas from AQUA-RAT, CommonsenseQA, and customized benchmarks, we demonstrate that SAGE-nano is at the cutting edge both on reasoning accuracy (74.6% on AQUA-RAT) and explanation quality (92.1% human preference score) for its task, and offers performance almost on par with models like Claude-3.5 Sonnet or GPT-4o. Our contributions are: (i) the first rigorous framework for LLM self-reflection via inverse reasoning, (ii) a novel metalearning framework to reverse the attention flow, (iii) comprehensive evaluation frameworks for reasoning transparency, and (iv) evidence that increasing reasoning using inverse reasoning improves interpretability along with reasoning performance. Our work creates new avenues for transparent AI systems and closes significant gaps in AI safety, education, and scientific discovery.

Vision Language Models (VLMs) hold great promise for streamlining labour-intensive medical imaging workflows, yet systematic security evaluations in clinical settings remain scarce. We introduce VSF--Med, an end-to-end vulnerability-scoring framework for medical VLMs that unites three novel components: (i) a rich library of sophisticated text-prompt attack templates targeting emerging threat vectors; (ii) imperceptible visual perturbations calibrated by structural similarity (SSIM) thresholds to preserve clinical realism; and (iii) an eight-dimensional rubric evaluated by two independent judge LLMs, whose raw scores are consolidated via z-score normalization to yield a 0--32 composite risk metric. Built entirely on publicly available datasets and accompanied by open-source code, VSF--Med synthesizes over 30,000 adversarial variants from 5,000 radiology images and enables reproducible benchmarking of any medical VLM with a single command. Our consolidated analysis reports mean z-score shifts of $0.90σ$ for persistence-of-attack-effects, $0.74σ$ for prompt-injection effectiveness, and $0.63σ$ for safety-bypass success across state-of-the-art VLMs. Notably, Llama-3.2-11B-Vision-Instruct exhibits a peak vulnerability increase of $1.29σ$ for persistence-of-attack-effects, while GPT-4o shows increases of $0.69σ$ for that same vector and $0.28σ$ for prompt-injection attacks.

With the rise of large language models, the paradigm of training foundation models with massive parameter counts on vast datasets has been adopted in multiple domains to achieve remarkable success. Time series foundation models represent a significant extension of this paradigm, demonstrating exceptional expressive power, generalization, and cross-domain transferability. However, this gives rise to a fundamental paradox: time series data reflect distinct dynamical systems, making cross-domain transfer intuitively implausible, yet this is contradicted by the models' empirical success. To resolve this paradox, this paper investigates, from both theoretical and experimental perspectives, the representation learning mechanisms and generalization capabilities of patch-based time series foundation models. We argue that such models are not merely applying a new architecture but are fundamentally generalizing the representation paradigm of language models by extending deterministic vector-based representations to latent probabilistic distributional forms. Our theoretical analysis supports this framework by demonstrating that continuous time-series patches can be faithfully quantized into a discrete vocabulary whose key statistical properties are highly consistent with those of natural language. This generalization allows time series models to inherit the robust representation and transfer abilities of large language models, thereby explaining their superior performance in temporal tasks. Ultimately, our work provides a rigorous theoretical cornerstone for understanding, evaluating, and improving the safety and reliability of large-scale time series foundation models.

Standard Verification Rule Format (SVRF) is essential for semiconductor applications like Design Rule Check (DRC), Layout Versus Schematic (LVS), and Optical Proximity Correction (OPC) and it faces challenges as advancing nodes create complex design rules that renders traditional SVRF development ineffective and highlight an expertise gap. This paper introduces a novel methodology integrating Abstract Syntax Tree (AST) embedding and Retrieval-Augmented Generation (RAG) for enhanced SVRF code synthesis, ensuring semantic accuracy and error minimization through structural validation with domain-specific insights for precise code generation. We evaluate different T5-based models and propose an innovative SVRF-specific scoring framework that complements standard metrics like BLEU and ROUGE-L. In our approach, AST provides rigorous structural validation, while RAG infuses relevant domain knowledge, effectively enhancing the code generation workflow. Testing on a comprehensive benchmark of 740 DRC rule implementations, our methodology demonstrates up to a 40\% improvement in code generation accuracy compared to basic text-based fine-tuning process. This fusion of industry expertise with advanced coding strategies not only optimizes SVRF development under limited dataset constraints but also creates a more intuitive and efficient coding environment. Consequently, users can rapidly iterate through design cycles, reduce manual error correction, and significantly improve overall productivity.

Mixture-of-Experts (MoE) enables efficient scaling of large language models by activating only a subset of experts per input token. However, deploying MoE-based models incurs significant memory overhead due to the need to retain all experts in memory. While structured pruning is promising to reduce memory costs, existing methods often show suboptimal performance and unstable degradation in three dimensions: model architectures, calibration data sources, and calibration sample sizes. This paper proposes Mixture-of-Novices-and-Experts (MoNE), a novel expert pruning method that replaces redundant experts with lightweight novices to achieve effective and robust model compression. MoNE evaluates expert redundancy based on two metrics: access frequency and output variance. Experts exhibiting low usage and stable outputs are pruned and replaced with lightweight novices-unbiased estimations of their original outputs-minimizing performance degradation. Extensive experiments demonstrate that MoNE consistently outperforms baseline methods with minimal accuracy degradation across the three dimensions, confirming its effectiveness and robustness. Notably, it improves the average zero shot accuracy across nine downstream tasks by up to 2.71 under 25\% pruning ratio and 3.61 under 50\% pruning. The code is available at https://github.com/zxgx/mode-pd.

Modern firms face a flood of dense, unstructured reports. Turning these documents into usable insights takes heavy effort and is far from agile when quick answers are needed. VTS-AI tackles this gap. It integrates Visual Thinking Strategies, which emphasize evidence-based observation, linking, and thinking, into AI agents, so the agents can extract business insights from unstructured text, tables, and images at scale. The system works in three tiers (micro, meso, macro). It tags issues, links them to source pages, and rolls them into clear action levers stored in a searchable YAML file. In tests on an 18-page business report, VTS-AI matched the speed of a one-shot ChatGPT prompt yet produced richer findings: page locations, verbatim excerpts, severity scores, and causal links. Analysts can accept or adjust these outputs in the same IDE, keeping human judgment in the loop. Early results show VTS-AI spots the direction of key metrics and flags where deeper number-crunching is needed. Next steps include mapping narrative tags to financial ratios, adding finance-tuned language models through a Model-Context Protocol, and building a Risk & Safety Layer to stress-test models and secure data. These upgrades aim to make VTS-AI a production-ready, audit-friendly tool for rapid business analysis.

Recent efforts to combine low-rank adaptation (LoRA) with mixture-of-experts (MoE) for adapting large language models (LLMs) to multiple tasks still exhibit prevailing limitations: they either swap entire attention/feed-forward layers for switch experts or bolt on parallel expert branches, diluting parameter efficiency and task fidelity. We propose the LoRA-Mixer, a modular and lightweight MoE framework that integrates LoRA experts. Our core innovation lies in replacing the projection matrices of the attention module's input/output linear layers with dynamically routed, task-specific LoRA experts. This design ensures seamless compatibility with diverse foundation models, including transformers and state space models (SSMs), by leveraging their inherent linear projection structures. The framework supports two operational paradigms: (1) joint optimization of LoRA experts and routing mechanisms via a novel hard-soft routing strategy, or (2) direct deployment of pre-trained, frozen LoRA modules sourced from external repositories. To enable robust router training with limited data while ensuring stable routing decisions and maximizing expert reuse, we introduce an adaptive Specialization Balance Loss (SBL) that jointly optimizes expert balance and task-specific alignment. Extensive experiments on seven benchmark datasets, including MedQA, CoLA, SST-2, GSM8K, ARC-E, ARC-C, and HumanEval, demonstrate the effectiveness of LoRA-Mixer. On datasets such as GSM8K, HumanEval, and MedQA, LoRA-Mixer achieves significant improvements of 7.61%, 4.88%, and 3.08% over the base models, respectively. Compared with state-of-the-art methods, LoRA-Mixer achieves additional improvements of 1.09%, 1.45%, and 1.68%, respectively, using only 48% of the parameters, demonstrating its efficiency and strong performance.

Evaluating creative writing generated by large language models (LLMs) remains challenging because open-ended narratives lack ground truths. Without performant automated evaluation methods, off-the-shelf (OTS) language models are employed as zero-shot judges, yet their reliability is unclear in this context. In pursuit of robust evaluation for creative writing, we introduce LitBench, the first standardized benchmark and paired dataset for creative writing verification, comprising a held-out test set of 2,480 debiased, human-labeled story comparisons drawn from Reddit and a 43,827-pair training corpus of human preference labels. Using LitBench, we (i) benchmark zero-shot LLM judges, (ii) train Bradley Terry and generative reward models, and (iii) conduct an online human study to validate reward model rankings on newly LLM-generated stories. Our benchmark identifies Claude-3.7-Sonnet as the strongest off-the-shelf judge, reaching 73% agreement with human preferences; among trained reward models, Bradley-Terry and Generative reward models both attain an accuracy of 78%, outperforming all off-the-shelf judges. An online human study further confirms that our trained reward models consistently align with human preferences in novel LLM-generated stories. We release LitBench and reward models at https://huggingface.co/collections/SAA-Lab/litbench-68267b5da3aafe58f9e43461, providing a vetted resource for reliable, automated evaluation and optimization of creative writing systems.