Clinical note generation aims to produce free-text summaries of a patient's condition and diagnostic process, with discharge instructions being a representative long-form example. While recent LLM-based methods pre-trained on general clinical corpora show promise in clinical text generation, they fall short in producing long-form notes from limited patient information. In this paper, we propose ReinRAG, a reinforced reasoning augmented generation (RAG) for long-form discharge instructions based on pre-admission information. ReinRAG retrieves reasoning paths from a medical knowledge graph to provide explicit semantic guidance to the LLM. To bridge the information gap, we propose group-based retriever optimization (GRO) which improves retrieval quality with group-normalized rewards, encouraging reasoning leaps for deeper inference by the LLM. Comprehensive experiments on the real-world dataset show that ReinRAG outperforms baselines in both clinical efficacy and natural language generation metrics. Further analysis reveals that ReinRAG fills semantic gaps in sparse input scenarios, and retrieved reasoning paths help LLMs avoid clinical misinterpretation by focusing on key evidence and following coherent reasoning.

To enhance reasoning capabilities, previous works have explored incorporating special-purpose tokens into the training process. These strategies strengthen the learning mechanism of transformer-based large language models (LLMs). Building on prior research, in which inserting dummy tokens consecutively just before reasoning steps can enhance effectiveness, we introduce a novel approach termed Dynamic Inserting Tokens Training (DIT). Our method identifies positions within sequences where model confidence is lowest according to token log-likelihood. Strategically inserting [PAUSE] tokens on these positions bolsters the model's predictive capabilities for subsequent tokens. Experimental results across diverse datasets and models, from the 2.7B model to the 8B model, demonstrate that DIT consistently outperforms traditional fine-tuning and previous token insertion methods. With this simple yet effective method, we achieve accuracy gains of up to 4.7%p on GSM8K, 3.23%p on AQUA-RAT, and pass@1 improvements of up to 3.4%p on MBPP datasets. Our work shows a model-based, dynamic approach rather than a heuristic one, thereby broadening the scope of research in reasoning.

Multimodal foundation models have demonstrated impressive capabilities across diverse tasks. However, their potential as plug-and-play solutions for missing modality reconstruction remains underexplored. To bridge this gap, we identify and formalize three potential paradigms for missing modality reconstruction, and perform a comprehensive evaluation across these paradigms, covering 42 model variants in terms of reconstruction accuracy and adaptability to downstream tasks. Our analysis reveals that current foundation models often fall short in two critical aspects: (i) fine-grained semantic extraction from the available modalities, and (ii) robust validation of generated modalities. These limitations lead to suboptimal and, at times, misaligned generations. To address these challenges, we propose an agentic framework tailored for missing modality reconstruction. This framework dynamically formulates modality-aware mining strategies based on the input context, facilitating the extraction of richer and more discriminative semantic features. In addition, we introduce a self-refinement mechanism, which iteratively verifies and enhances the quality of generated modalities through internal feedback. Experimental results show that our method reduces FID for missing image reconstruction by at least 14\% and MER for missing text reconstruction by at least 10\% compared to baselines. Code are released at: https://github.com/Guanzhou-Ke/AFM2.

Large language models (LLMs) are increasingly becoming valuable to corporate data management due to their ability to process text from various document formats and facilitate user interactions through natural language queries. However, LLMs must consider the sensitivity of information when communicating with employees, especially given access restrictions. Simple filtering based on user clearance levels can pose both performance and privacy challenges. To address this, we propose the concept of sensitivity awareness (SA), which enables LLMs to adhere to predefined access rights rules. In addition, we developed a benchmarking environment called ACCESS DENIED INC to evaluate SA. Our experimental findings reveal significant variations in model behavior, particularly in managing unauthorized data requests while effectively addressing legitimate queries. This work establishes a foundation for benchmarking sensitivity-aware language models and provides insights to enhance privacy-centric AI systems in corporate environments.

Large Language Models (LLMs) have achieved remarkable performance on a wide range of Natural Language Processing (NLP) benchmarks, often surpassing human-level accuracy. However, their reliability in high-stakes domains such as medicine, particularly in low-resource languages, remains underexplored. In this work, we introduce PersianMedQA, a large-scale dataset of 20,785 expert-validated multiple-choice Persian medical questions from 14 years of Iranian national medical exams, spanning 23 medical specialties and designed to evaluate LLMs in both Persian and English. We benchmark 40 state-of-the-art models, including general-purpose, Persian fine-tuned, and medical LLMs, in zero-shot and chain-of-thought (CoT) settings. Our results show that closed-source general models (e.g., GPT-4.1) consistently outperform all other categories, achieving 83.09% accuracy in Persian and 80.7% in English, while Persian fine-tuned models such as Dorna underperform significantly (e.g., 34.9% in Persian), often struggling with both instruction-following and domain reasoning. We also analyze the impact of translation, showing that while English performance is generally higher, 3-10% of questions can only be answered correctly in Persian due to cultural and clinical contextual cues that are lost in translation. Finally, we demonstrate that model size alone is insufficient for robust performance without strong domain or language adaptation. PersianMedQA provides a foundation for evaluating bilingual and culturally grounded medical reasoning in LLMs. The PersianMedQA dataset is available: https://huggingface.co/datasets/MohammadJRanjbar/PersianMedQA .

The growing popularity of social deduction games has created an increasing need for intelligent frameworks where humans can collaborate with AI agents, particularly in post-pandemic contexts with heightened psychological and social pressures. Social deduction games like Werewolf, traditionally played through verbal communication, present an ideal application for Large Language Models (LLMs) given their advanced reasoning and conversational capabilities. Prior studies have shown that LLMs can outperform humans in Werewolf games, but their reliance on external modules introduces latency that left their contribution in academic domain only, and omit such game should be user-facing. We propose \textbf{Verbal Werewolf}, a novel LLM-based Werewolf game system that optimizes two parallel pipelines: gameplay powered by state-of-the-art LLMs and a fine-tuned Text-to-Speech (TTS) module that brings text output to life. Our system operates in near real-time without external decision-making modules, leveraging the enhanced reasoning capabilities of modern LLMs like DeepSeek V3 to create a more engaging and anthropomorphic gaming experience that significantly improves user engagement compared to existing text-only frameworks.

Addressing intricate real-world problems necessitates in-depth information seeking and multi-step reasoning. Recent progress in agentic systems, exemplified by Deep Research, underscores the potential for autonomous multi-step research. In this work, we present a cohesive paradigm for building end-to-end agentic information seeking agents from a data-centric and training-stage perspective. Our approach consists of four key stages: (1) browsing data construction, (2) trajectories sampling, (3) supervised fine-tuning for effective cold start, and (4) reinforcement learning for enhanced generalisation. We instantiate this framework in a web agent based on the ReAct, WebDancer. Empirical evaluations on the challenging information seeking benchmarks, GAIA and WebWalkerQA, demonstrate the strong performance of WebDancer, achieving considerable results and highlighting the efficacy of our training paradigm. Further analysis of agent training provides valuable insights and actionable, systematic pathways for developing more capable agentic models. The codes and demo will be released in https://github.com/Alibaba-NLP/WebAgent.

Training large language models (LLMs) for complex reasoning via Reinforcement Learning with Verifiable Rewards (RLVR) is effective but limited by reliance on costly, domain-specific supervision. We explore Reinforcement Learning from Internal Feedback (RLIF), a framework that enables LLMs to learn from intrinsic signals without external rewards or labeled data. We propose Intuitor, an RLIF method that uses a model's own confidence, termed self-certainty, as its sole reward signal. Intuitor replaces external rewards in Group Relative Policy Optimization (GRPO) with self-certainty scores, enabling fully unsupervised learning. Experiments demonstrate that Intuitor matches GRPO's performance on mathematical benchmarks while achieving superior generalization to out-of-domain tasks like code generation, without requiring gold solutions or test cases. Our findings show that intrinsic model signals can drive effective learning across domains, offering a scalable alternative to RLVR for autonomous AI systems where verifiable rewards are unavailable. Code is available at https://github.com/sunblaze-ucb/Intuitor

We demonstrate, for the first time, fully quantized training (FQT) of large language models (LLMs) using predominantly 4-bit floating-point (FP4) precision for weights, activations, and gradients on datasets up to 200 billion tokens. We extensively investigate key design choices for FP4, including block sizes, scaling formats, and rounding methods. Our analysis shows that the NVFP4 format, where each block of 16 FP4 values (E2M1) shares a scale represented in E4M3, provides optimal results. We use stochastic rounding for backward and update passes and round-to-nearest for the forward pass to enhance stability. Additionally, we identify a theoretical and empirical threshold for effective quantized training: when the gradient norm falls below approximately $\sqrt{3}$ times the quantization noise, quantized training becomes less effective. Leveraging these insights, we successfully train a 7-billion-parameter model on 256 Intel Gaudi2 accelerators. The resulting FP4-trained model achieves downstream task performance comparable to a standard BF16 baseline, confirming that FP4 training is a practical and highly efficient approach for large-scale LLM training. A reference implementation is supplied in https://github.com/Anonymous1252022/fp4-all-the-way .

Using LLMs in a production environment presents security challenges that include vulnerabilities to jailbreaks and prompt injections, which can result in harmful outputs for humans or the enterprise. The challenge is amplified when working within a specific domain, as topics generally accepted for LLMs to address may be irrelevant to that field. These problems can be mitigated, for example, by fine-tuning large language models with domain-specific and security-focused data. However, these alone are insufficient, as jailbreak techniques evolve. Additionally, API-accessed models do not offer the flexibility needed to tailor behavior to industry-specific objectives, and in-context learning is not always sufficient or reliable. In response to these challenges, we introduce Archias, an expert model adept at distinguishing between in-domain and out-of-domain communications. Archias classifies user inquiries into several categories: in-domain (specifically for the automotive industry), malicious questions, price injections, prompt injections, and out-of-domain examples. Our methodology integrates outputs from the expert model (Archias) into prompts, which are then processed by the LLM to generate responses. This method increases the model's ability to understand the user's intention and give appropriate answers. Archias can be adjusted, fine-tuned, and used for many different purposes due to its small size. Therefore, it can be easily customized to the needs of any industry. To validate our approach, we created a benchmark dataset for the automotive industry. Furthermore, in the interest of advancing research and development, we release our benchmark dataset to the community.