Large Language Models (LLMs) are fast becoming indispensable tools for software developers, assisting or even partnering with them in crafting complex programs. The advantages are evident -- LLMs can significantly reduce development time, generate well-organized and comprehensible code, and occasionally suggest innovative ideas that developers might not conceive on their own. However, despite their strengths, LLMs will often introduce significant errors and present incorrect code with persuasive confidence, potentially misleading developers into accepting flawed solutions. In order to bring LLMs into the software development cycle in a more reliable manner, we propose a methodology for combining them with ``traditional'' software engineering techniques in a structured way, with the goal of streamlining the development process, reducing errors, and enabling users to verify crucial program properties with increased confidence. Specifically, we focus on the Scenario-Based Programming (SBP) paradigm -- an event-driven, scenario-based approach for software engineering -- to allow human developers to pour their expert knowledge into the LLM, as well as to inspect and verify its outputs. To evaluate our methodology, we conducted a significant case study, and used it to design and implement the Connect4 game. By combining LLMs and SBP we were able to create a highly-capable agent, which could defeat various strong existing agents. Further, in some cases, we were able to formally verify the correctness of our agent. Finally, our experience reveals interesting insights regarding the ease-of-use of our proposed approach. The full code of our case-study will be made publicly available with the final version of this paper.

West Lafayette, Indiana, USA Abstract--Zero-shot V oice Cloning (VC) and T ext-to-Speech (TTS) methods have advanced rapidly, enabling the generation of highly realistic synthetic speech and raising serious concerns about their misuse. While numerous detectors have been developed for English and Chinese, Spanish--spoken by over 600 million people worldwide--remains underrepresented in speech forensics. T o address this gap, we introduce HISPASpoof, the first large-scale Spanish dataset designed for synthetic speech detection and attribution. It includes real speech from public corpora across six accents and synthetic speech generated with six zero-shot TTS systems. We evaluate five representative methods, showing that detectors trained on English fail to generalize to Spanish, while training on HISPASpoof substantially improves detection. We also evaluate synthetic speech attribution performance on HISPASpoof, i.e., identifying the generation method of synthetic speech. HISPASpoof thus provides a critical benchmark for advancing reliable and inclusive speech forensics in Spanish. The rapid advancement of speech synthesis techniques has significantly transformed the area of audio generation and speech forensics. Recent Text-to-Speech (TTS) and V oice Cloning (VC) methods [1], [2], [3], [4], [5], [6] are now capable of producing highly realistic synthetic voices that closely mimic the spectral, prosodic, and linguistic traits of real human speech [7], [8], [9], [10].

Large language models (LLMs) excel in general-domain applications, yet their performance often degrades in specialized tasks requiring domain-specific knowledge. E-commerce is particularly challenging, as its data are noisy, heterogeneous, multilingual, and highly dynamic. We present Compass-v3, a vertical-domain Mixture-of-Experts (MoE) model with 245B total parameters and 71B active per token, designed for Southeast Asian e-commerce. Compass-v3 adopts fewer but larger experts, combined with hardware-efficient optimizations-such as intra-node expert parallelism and a customized memcpy operator-to maximize GPU utilization. The model is trained on 12T tokens of curated multilingual corpora and large-scale synthetic e-commerce instructions using a mixed-training strategy. To enhance alignment, we propose Optimal-Transport Direct Preference Optimization (OTPO), which captures token-level distinctions and improves instruction adherence in commerce-specific scenarios. Extensive evaluations demonstrate that Compass-v3 delivers state-of-the-art e-commerce performance, surpassing DeepSeek-V3.1, GPT-4 series, and Qwen3-235B. Moreover, Compass-v3 demonstrates strong multilingual capability across low-resource Southeast Asian languages (Indonesian, Thai, Filipino, Vietnamese, Malay, Taglog) and Portuguese while sustaining competitive performance on general benchmarks. It has already been widely applied in Shopee's industrial-scale e-commerce platform and is gradually replacing OpenAI's traffic, now accounting for over 70\% of total LLM usage, highlighting its dual strengths in specialized commerce expertise and broad linguistic competence.

Abstract--As on-device large language model (LLM) systems become increasingly prevalent, federated fine-tuning enables advanced language understanding and generation directly on edge devices; however, it also involves processing sensitive, user-specific data, raising significant privacy concerns within the federated learning framework. T o address these challenges, we propose DP-FedLoRA, a privacy-enhanced federated fine-tuning framework that integrates LoRA-based adaptation with differential privacy in a communication-efficient setting. Each client locally clips and perturbs its LoRA matrices using Gaussian noise to satisfy (ϵ, δ)-differential privacy. We further provide a theoretical analysis demonstrating the unbiased nature of the updates and deriving bounds on the variance introduced by noise, offering practical guidance for privacy-budget calibration. Experimental results across mainstream benchmarks show that DP-FedLoRA delivers competitive performance while offering strong privacy guarantees, paving the way for scalable and privacy-preserving LLM deployment in on-device environments. The increasing deployment of on-device large language models (LLMs) has brought powerful language understanding and generation capabilities directly to edge devices [1], [2].

The cold-start user issue further compromises the effectiveness of recommender systems in limiting access to the historical behavioral information. It is an effective pipeline to optimize instructional prompts on a few-shot large language model (LLM) used in recommender tasks. We introduce a context-conditioned prompt formulation method P(u,\ Ds)\ \rightarrow\ R\widehat, where u is a cold-start user profile, Ds is a curated support set, and R\widehat is the predicted ranked list of items. Based on systematic experimentation with transformer-based autoregressive LLMs (BioGPT, LLaMA-2, GPT-4), we provide empirical evidence that optimal exemplar injection and instruction structuring can significantly improve the precision@k and NDCG scores of such models in low-data settings. The pipeline uses token-level alignments and embedding space regularization with a greater semantic fidelity. Our findings not only show that timely composition is not merely syntactic but also functional as it is in direct control of attention scales and decoder conduct through inference. This paper shows that prompt-based adaptation may be considered one of the ways to address cold-start recommendation issues in LLM-based pipelines.

This research investigates the effectiveness of alignment techniques, Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and a combined SFT+DPO approach on improving the safety and helpfulness of the OPT-350M language model. Utilizing the Anthropic Helpful-Harmless RLHF dataset, we train and evaluate four models: the base OPT350M, an SFT model, a DPO model, and a model trained with both SFT and DPO. We introduce three key evaluation metrics: Harmlessness Rate (HmR), Helpfulness Rate (HpR), and a Combined Alignment Score (CAS), all derived from reward model outputs. The results show that while SFT outperforms DPO, The combined SFT+DPO model outperforms all others across all metrics, demonstrating the complementary nature of these techniques. Our findings also highlight challenges posed by noisy data, limited GPU resources, and training constraints. This study offers a comprehensive view of how fine-tuning strategies affect model alignment and provides a foundation for more robust alignment pipelines in future work.

Despite strong performance in visual understanding and language-based reasoning, Vision-Language Models (VLMs) struggle with tasks requiring integrated perception and symbolic computation. We study this limitation through visual equation solving, where mathematical equations are embedded in images, variables are represented by object icons, and coefficients must be inferred by counting. While VLMs perform well on textual equations, they fail on visually grounded counterparts. To understand this gap, we decompose the task into coefficient counting and variable recognition, and find that counting is the primary bottleneck, even when recognition is accurate. We also observe that composing recognition and reasoning introduces additional errors, highlighting challenges in multi-step visual reasoning. Finally, as equation complexity increases, symbolic reasoning itself becomes a limiting factor. These findings reveal key weaknesses in current VLMs and point toward future improvements in visually grounded mathematical reasoning.

With the rapid advancement of multimodal retrieval and its application in LLMs and multimodal LLMs, increasingly complex retrieval tasks have emerged. Existing methods predominantly rely on task-specific fine-tuning of vision-language models and are limited to single-modality queries or documents. In this paper, we propose ReT-2, a unified retrieval model that supports multimodal queries, composed of both images and text, and searches across multimodal document collections where text and images coexist. ReT-2 leverages multi-layer representations and a recurrent Transformer architecture with LSTM-inspired gating mechanisms to dynamically integrate information across layers and modalities, capturing fine-grained visual and textual details. We evaluate ReT-2 on the challenging M2KR and M-BEIR benchmarks across different retrieval configurations. Results demonstrate that ReT-2 consistently achieves state-of-the-art performance across diverse settings, while offering faster inference and reduced memory usage compared to prior approaches. When integrated into retrieval-augmented generation pipelines, ReT-2 also improves downstream performance on Encyclopedic-VQA and InfoSeek datasets. Our source code and trained models are publicly available at: https://github.com/aimagelab/ReT-2

Abstract--The prevalence of Large Language Models (LLMs) is having an growing impact on the climate due to the substantial energy required for their deployment and use. T o create awareness for developers who are implementing LLMs in their products, there is a strong need to collect more information about the energy efficiency of LLMs. While existing research has evaluated the energy efficiency of various models, these benchmarks often fall short of representing realistic production scenarios. In this paper, we introduce the LLM Efficiency Benchmark, designed to simulate real-world usage conditions. We examine how factors such as model size, architecture, and concurrent request volume affect inference energy efficiency. Our findings demonstrate that it is possible to create energy efficiency benchmarks that better reflect practical deployment conditions, providing valuable insights for developers aiming to build more sustainable AI systems. Large Language Models (LLMs) have seen a significant rise in popularity in recent years. They are increasingly integrated into everyday applications, such as Google's AI-generated summaries for search results, OpenAI's GPT -4o, and the growing adoption of AI agents across various platforms.

Providing students with flexible and timely academic support is a challenge at most colleges and universities, leaving many students without help outside scheduled hours. Large language models (LLMs) are promising for bridging this gap, but interactions between students and LLMs are rarely overseen by educators. We developed and studied an LLM-powered course assistant deployed across multiple computer science courses to characterize real-world use and understand pedagogical implications. By Spring 2024, our system had been deployed to approximately 2,000 students across six courses at three institutions. Analysis of the interaction data shows that usage remains strong in the evenings and nights and is higher in introductory courses, indicating that our system helps address temporal support gaps and novice learner needs. We sampled 200 conversations per course for manual annotation: most sampled responses were judged correct and helpful, with a small share unhelpful or erroneous; few responses included dedicated examples. We also examined an inquiry-based learning strategy: only around 11% of sampled conversations contained LLM-generated follow-up questions, which were often ignored by students in advanced courses. A Bloom's taxonomy analysis reveals that current LLM capabilities are limited in generating higher-order cognitive questions. These patterns suggest opportunities for pedagogically oriented LLM-based educational systems and greater educator involvement in configuring prompts, content, and policies.