arXiv cs.CL
·
18 hours ago
NAVER LABS implemented a speech-to-text instruction-following system for the IWSLT 2026 shared task using SeamlessM4T-v2-large for speech encoding and Qwen3-4B-Instruct as the language model. The system uses a three-stage pipeline with projector alignment, text-only LoRA pre-training, and multimodal merging, supported by 100,000 synthetic instruction-following examples across ten task types. The model achieved COMET 0.781 on English-to-Chinese speech translation and BERTScore-F1 0.346 on English spoken question answering.
arXiv cs.CL
·
18 hours ago
Researchers introduced ISE, a three-stage method for generating training data for OS agents that creates structured user intents, simulates multi-turn interactions, and executes tool calls in live environments. The dataset contains 43,956 unique intents and 23,132 trajectories averaging 8.12 user turns each, with generated data improving a Qwen3-8B model's performance on ClawEval from 19.3 to 37.7 pass@1. This approach allows smaller models to outperform zero-shot GPT-4o on agent tool-use tasks through better training data grounded in actual execution outcomes.
arXiv cs.CL
·
18 hours ago
Researchers developed a seven-dimensional taxonomy to classify cancer misinformation in Reddit discussions and evaluated large language models on their ability to detect these false claims across communities. Cancer misinformation comprised approximately 6% of Reddit cancer discussions, with notable variation across different cancer types and communities. The taxonomy and annotated dataset enable more nuanced detection of cancer misinformation narratives involving unsupported treatments and distrust of conventional medicine.
arXiv cs.CL
·
18 hours ago
Researchers introduced CANDI-QA, a dataset for evaluating large language models on question-answering tasks in specialized domains like medicine and finance, featuring expert-curated pairs split into direct factual queries and multi-hop reasoning tasks. The evaluation tested over ten language models ranging from open-source to proprietary systems, with MTSS-Net presented as a baseline neuro-symbolic framework combining neural retrieval with rule-based reasoning. The benchmark reveals that current LLMs struggle with contextual alignment in niche domains without enhanced contextual or symbolic integration, providing a tool to advance development of context-aware models for high-stakes applications.
arXiv cs.CL
·
18 hours ago
Researchers developed point-in-time language models trained only on text available up to specific calendar dates to eliminate lookahead bias for financial and social science applications. Models with up to 4 billion parameters trained on 1 trillion chronologically filtered tokens achieved performance approaching comparable unrestricted models like Gemma-3-4B and LLaMA-7B across common reasoning and language understanding benchmarks. The release of the training pipeline and dataset enables reproducible temporal validity research that was previously compromised by models inadvertently using future information.
arXiv cs.AI
·
18 hours ago
MedDiffuseMix is a diffusion-based image augmentation method that uses saliency maps to preserve diagnostically important regions while mixing lower-importance areas to improve medical image classification. The method was evaluated on four public datasets including chest radiographs and histopathology images, showing improvements in accuracy and F1-score over standard augmentation and baseline diffusion approaches. By constraining augmentation to clinically irrelevant regions, the framework maintains diagnostic evidence while increasing training data diversity for improved classifier performance.
arXiv cs.AI
·
18 hours ago
Researchers introduced MedRealMM, a benchmark dataset for evaluating large language models on Chinese online medical consultation using real patient-doctor interactions with images. The dataset contains 5,620 multimodal cases across 64 clinical departments, with physician-refined rubrics that assess both positive clinical behaviors and safety issues. Testing of 19 models found that current frontier models fall short of physician performance, particularly in avoiding unsafe or unsupported responses despite meeting clinical criteria in other areas.
arXiv cs.AI
·
18 hours ago
Researchers released Human4K, a dataset containing over six million 4K images from eight synchronized cameras and motion capture data for training 3D human reconstruction models. The dataset includes recordings of 11 subjects performing complex full-body motions with precise SMPL-X annotations covering hands, feet, and occluded body parts. Models trained on Human4K show improved performance on standard benchmarks, particularly for reconstructing hands, feet, and depth-ambiguous limb configurations.
