We introduce KodCode, a synthetic dataset that addresses the persistent challenge of acquiring high-quality, verifiable training data across diverse difficulties and domains for training Large Language Models for coding. Existing code-focused resources typically fail to ensure either the breadth of coverage (e.g., spanning simple coding tasks to advanced algorithmic problems) or verifiable correctness (e.g., unit tests). In contrast, KodCode comprises question-solution-test triplets that are systematically validated via a self-verification procedure. Our pipeline begins by synthesizing a broad range of coding questions, then generates solutions and test cases with additional attempts allocated to challenging problems. Finally, post-training data synthesis is done by rewriting questions into diverse formats and generating responses under a test-based reject sampling procedure from a reasoning model (DeepSeek R1). This pipeline yields a large-scale, robust and diverse coding dataset. KodCode is suitable for supervised fine-tuning and the paired unit tests also provide great potential for RL tuning. Fine-tuning experiments on coding benchmarks (HumanEval(+), MBPP(+), BigCodeBench, and LiveCodeBench) demonstrate that KodCode-tuned models achieve state-of-the-art performance, surpassing models like Qwen2.5-Coder-32B-Instruct and DeepSeek-R1-Distill-Llama-70B.

Recent advancements in Large Language Model (LLM)-based Natural Language Generation evaluation have largely focused on single-example prompting, resulting in significant token overhead and computational inefficiencies. In this work, we introduce BatchGEMBA-MQM, a framework that integrates batched prompting with the GEMBA-MQM metric for machine translation evaluation. Our approach aggregates multiple translation examples into a single prompt, reducing token usage by 2-4 times (depending on the batch size) relative to single-example prompting. Furthermore, we propose a batching-aware prompt compression model that achieves an additional token reduction of 13-15% on average while also showing ability to help mitigate batching-induced quality degradation. Evaluations across several LLMs (GPT-4o, GPT-4o-mini, Mistral Small, Phi4, and CommandR7B) and varying batch sizes reveal that while batching generally negatively affects quality (but sometimes not substantially), prompt compression does not degrade further, and in some cases, recovers quality loss. For instance, GPT-4o retains over 90% of its baseline performance at a batch size of 4 when compression is applied, compared to a 44.6% drop without compression. We plan to release our code and trained models at https://github.com/NL2G/batchgemba to support future research in this domain.

Biomedical visual question answering (VQA) has been widely studied and has demonstrated significant application value and potential in fields such as assistive medical diagnosis. Despite their success, current biomedical VQA models perform multimodal information interaction only at the model level within large language models (LLMs), leading to suboptimal multimodal semantic alignment when dealing with complex tasks. To address this issue, we propose BioD2C: a novel Dual-level Semantic C onsistency C onstraint Framework for Biomedical VQA, which achieves dual-level semantic interaction alignment at both the model and feature levels, enabling the model to adaptively learn visual features based on the question. Specifically, we firstly integrate textual features into visual features via an image-text fusion mechanism as feature-level semantic interaction, obtaining visual features conditioned on the given text; and then introduce a text-queue-based cross-modal soft semantic loss function to further align the image semantics with the question semantics. Specifically, in this work, we establish a new dataset, BioVGQ, to address inherent biases in prior datasets by filtering manually-altered images and aligning question-answer pairs with multimodal context, and train our model on this dataset. Extensive experimental results demonstrate that BioD2C achieves state-of-the-art (SOTA) performance across multiple downstream datasets, showcasing its robustness, generalizabil-ity, and potential to advance biomedical VQA research. The source code of this work and the BioVGQ dataset can be accessed through code and dataset. Keywords: Biomedical VQA Dual Interaction Semantic Alignment 1 Introduction Biomedical VQA aims to design and develop systems capable of understanding biomedical images and generating relevant answers based on given textual instructions.

Generating high-quality question-answer pairs for specialized technical domains remains challenging, with existing approaches facing a tradeoff between leveraging expert examples and achieving topical diversity. We present ExpertGenQA, a protocol that combines few-shot learning with structured topic and style categorization to generate comprehensive domain-specific QA pairs. Using U.S. Federal Railroad Administration documents as a test bed, we demonstrate that ExpertGenQA achieves twice the efficiency of baseline few-shot approaches while maintaining $94.4\%$ topic coverage. Through systematic evaluation, we show that current LLM-based judges and reward models exhibit strong bias toward superficial writing styles rather than content quality. Our analysis using Bloom's Taxonomy reveals that ExpertGenQA better preserves the cognitive complexity distribution of expert-written questions compared to template-based approaches. When used to train retrieval models, our generated queries improve top-1 accuracy by $13.02\%$ over baseline performance, demonstrating their effectiveness for downstream applications in technical domains.

Bilateral markets, such as those for government bonds, involve decentralized and opaque transactions between market makers (MMs) and clients, posing significant challenges for traditional modeling approaches. To address these complexities, we introduce TRIBE an agent-based model augmented with a large language model (LLM) to simulate human-like decision-making in trading environments. TRIBE leverages publicly available data and stylized facts to capture realistic trading dynamics, integrating human biases like risk aversion and ambiguity sensitivity into the decision-making processes of agents. Our research yields three key contributions: first, we demonstrate that integrating LLMs into agent-based models to enhance client agency is feasible and enriches the simulation of agent behaviors in complex markets; second, we find that even slight trade aversion encoded within the LLM leads to a complete cessation of trading activity, highlighting the sensitivity of market dynamics to agents' risk profiles; third, we show that incorporating human-like variability shifts power dynamics towards clients and can disproportionately affect the entire system, often resulting in systemic agent collapse across simulations. These findings underscore the emergent properties that arise when introducing stochastic, human-like decision processes, revealing new system behaviors that enhance the realism and complexity of artificial societies.

Although these approaches can improve models' Large Language Models (LLMs) face challenges for on-device memory efficiency, they inevitably impact the generality inference due to high memory demands. Traditional methods performance and still suffer in extreme resource-constrained to reduce memory usage often compromise performance scenarios [4, 9, 12]. Furthermore, these methods lack the flexibility and lack adaptability. We propose FlexInfer, an optimized to vary memory budgets or deployment constraints, offloading framework for on-device inference, addressing requiring adjusting the hyper-parameters, such as quantization these issues with techniques like asynchronous prefetching, or sparsity levels, offering limited choices, and imposing balanced memory locking, and flexible tensor preservation.

Recent advancements in speech large language models (SpeechLLMs) have attracted considerable attention. Nonetheless, current methods exhibit suboptimal performance in adhering to speech instructions. Notably, the intelligence of models significantly diminishes when processing speech-form input as compared to direct text-form input. Prior work has attempted to mitigate this semantic inconsistency between speech and text representations through techniques such as representation and behavior alignment, which involve the meticulous design of data pairs during the post-training phase. In this paper, we introduce a simple and scalable training method called InSerter, which stands for Interleaved Speech-Text Representation Pre-training. InSerter is designed to pre-train large-scale unsupervised speech-text sequences, where the speech is synthesized from randomly selected segments of an extensive text corpus using text-to-speech conversion. Consequently, the model acquires the ability to generate textual continuations corresponding to the provided speech segments, obviating the need for intensive data design endeavors. To systematically evaluate speech instruction-following capabilities, we introduce SpeechInstructBench, the first comprehensive benchmark specifically designed for speech-oriented instruction-following tasks. Our proposed InSerter achieves SOTA performance in SpeechInstructBench and demonstrates superior or competitive results across diverse speech processing tasks.

Abstract-- We present LLM-Glasses, a wearable navigation system designed to assist visually impaired individuals by combining haptic feedback, YOLO-World object detection, and GPT-4o-driven reasoning. The system delivers real-time tactile guidance via temple-mounted actuators, enabling intuitive and independent navigation. Three user studies were conducted to evaluate its effectiveness: (1) a haptic pattern recognition study achieving an 81.3% average recognition rate across 13 distinct patterns, (2) a VICON-based navigation study in which participants successfully followed predefined paths in open spaces, and (3) an LLM-guided video evaluation demonstrating 91.8% accuracy in open scenarios, 84.6% with static obstacles, and 81.5% with dynamic obstacles. These results demonstrate the system's reliability in controlled environments, with ongoing work focusing on refining its responsiveness and adaptability to diverse real-world scenarios. LLM-Glasses showcases the potential of combining generative AI with haptic interfaces to empower visually impaired individuals with intuitive and effective mobility solutions.

This paper presents Matrix, an advanced AI-powered framework designed for real-time 3D object generation in Augmented Reality (AR) environments. By integrating a cutting-edge text-to-3D generative AI model, multilingual speech-to-text translation, and large language models (LLMs), the system enables seamless user interactions through spoken commands. The framework processes speech inputs, generates 3D objects, and provides object recommendations based on contextual understanding, enhancing AR experiences. A key feature of this framework is its ability to optimize 3D models by reducing mesh complexity, resulting in significantly smaller file sizes and faster processing on resource-constrained AR devices. Our approach addresses the challenges of high GPU usage, large model output sizes, and real-time system responsiveness, ensuring a smoother user experience. Moreover, the system is equipped with a pre-generated object repository, further reducing GPU load and improving efficiency. We demonstrate the practical applications of this framework in various fields such as education, design, and accessibility, and discuss future enhancements including image-to-3D conversion, environmental object detection, and multimodal support. The open-source nature of the framework promotes ongoing innovation and its utility across diverse industries.

In the data era, the integration of multiple data types, known as multimodality, has become a key area of interest in the research community. This interest is driven by the goal to develop cutting-edge multimodal models capable of serving as adaptable reasoning engines across a wide range of modalities and domains. Despite the fervent development efforts, the challenge of optimally representing different forms of data within a single unified latent space--a crucial step for enabling effective multimodal reasoning--has not been fully addressed. To bridge this gap, we introduce AlignXpert, an optimization algorithm inspired by Kernel CCA crafted to maximize the similarities between N modalities while imposing some other constraints. This work demonstrates the impact on improving data representation for a variety of reasoning tasks, such as retrieval and classification, underlining the pivotal importance of data representation.