Distilling large reasoning models (LRMs) has become essential for making their Long-CoT reasoning capabilities practical, as full-scale inference remains computationally prohibitive. Existing curation-based approaches, which select complete reasoning traces post-hoc, overlook the collaborative potential of heterogeneous teachers and fail to adapt exploration dynamically, often leading to redundant sampling and missed opportunities for complementary reasoning. To address this, we introduce CoRD, a collaborative multi-teacher decoding framework that performs step-wise reasoning synthesis guided by predictive perplexity–based scoring and beam search. This approach enables heterogeneous LRMs to jointly construct coherent reasoning trajectories while maintaining diverse, high-potential hypotheses efficiently. Experiments show that CoRD generates higher-quality reasoning data and achieves student performance approaching teacher-level results, demonstrating that fine-grained collaboration among diverse LRMs yields structured, efficient, and robust reasoning distillation. The dataset and model are available at https://github.com/DISL-Lab/CoRD

With recent advances in diffusion models, text-to-image (T2I) generation has achieved remarkable progress, producing high-quality images from simple text prompts. However, as these models have become more expressive, concerns regarding the generation of unsafe or inappropriate content have grown. To address this issue, we propose the Safety-Adjustable Framework for Ethics (Safe), a novel training-free, plug-and-play prompt optimization framework. Leveraging the reasoning capabilities of large language models (LLMs), Safe optimizes prompts to meet concept-specific safety requirements while preserving their original semantics. The framework operates in three stages: (1) generating an image from the original prompt using a pretrained T2I model, (2) analyzing the generated image with a vision-language model (VLM), and (3) conducting DAG-based reasoning over both the prompt and the image description to produce refined prompt. In addition, Safe uses a search module to incorporate external knowledge such as cultural norms and legal constraints, enabling adaptation to diverse user contexts. Extensive experiments demonstrate that Safe effectively complements existing safety methods and achieves strong performance in user-centric safety alignment.

Large language models (LLMs) are increasingly used in scientific writing, yet their citation behavior remains poorly understood. We introduce CiteSpectrum, a controlled dataset for analyzing citations across five levels of AI involvement, from LLM-Only generation to Human-Authored papers. Using citation-behavior features across fidelity, usage, and topology, we show that AI-generated papers differ from human-written papers in how citations are validated, grounded, used, and structurally organized. Even when retrieval and human guidance reduce ghost citations, AI-generated papers contain more unsupported claims, rely on front-section evidence, use citations more affirmatively, and form broader but less interconnected reference networks. Moreover, citation behavior alone enables a lightweight classifier to outperform strong text-based detectors, achieving 0.880 balanced accuracy and 0.944 AUROC. These results suggest that citation behavior provides a strong, interpretable signal of AI involvement beyond conventional textual patterns. The dataset is available at: https://github.com/DISL-Lab/CiteSpectrum-EMNLP26.

Documents that users query mix tables with prose, but the QA systems built for them do not: retrieve-and-read pipelines flatten tables, while text-to-SQL pipelines ignore passages. Corporate filings, scientific reports, and Wikipedia articles all sit on this fault line, and a single user question routinely crosses it. We present ATR (AgenticTableRAG), an agentic framework for hybrid table-text question answering. ATR decomposes each question into ordered sub-queries and dispatches each to one of four execution primitives: text reading, cell retrieval, SQL execution, or a hybrid of cell grounding and SQL. When the retrieved evidence does not support the predicted answer, the router is re-invoked with the failed-route history and dispatches a different primitive. Value grounding composed with constrained SQL execution rules out column and value hallucinations by construction. The routing policy is distilled into a lightweight student model that transfers across benchmarks without additional inference-time LLM calls. A single ATR configuration leads on token F1 across three benchmarks without per-dataset tuning.