Enhancing LLM reasoning in federated settings is nontrivial due to stringent computational, communication, and privacy constraints, especially in healthcare, where clinically consequential decisions require not only accuracy but also interpretable, auditable rationales to meet safety, accountability, and regulatory requirements. Conventional federated fine-tuning largely imitates final answers rather than cultivating step-by-step reasoning, often relying on privacy-sensitive centralized distillation and still incurring substantial communication overhead. We address this gap with FedCoT, a federated reasoning framework that combines lightweight chain-of-thought resampling with a compact discriminator for selection, and client-aware LoRA stacking with weighted classifier aggregation to accommodate heterogeneity while reducing aggregation noise and communication; clients generate candidate chains and supervision locally, and only lightweight modules are aggregated on the server. Experiments on medical reasoning benchmarks show consistent gains under tight resource budgets while keeping data local and respecting privacy, offering an interpretable and resource-efficient solution.

Despite the advanced capabilities of Large Language Models (LLMs), training specialized reasoning models for the medical domain remains a significant challenge due to the scarcity of high-quality, large-scale Chain-of-Thought (CoT) data. Moreover, the intermediate reasoning steps in teacher-generated CoT data can be redundant and noisy, leading models to acquire spurious patterns and resulting in suboptimal performance. To address these issues, we propose MedCoach, a novel framework that introduces a dedicated coach role to guide the student model through question decomposition, thereby smoothing its learning curve in medical reasoning. The framework employs a curriculum-oriented warm-up on simplified sub-questions, facilitating domain adaptation before advancing to complex long-chain reasoning. To ensure the fidelity of the intermediate chain-of-thought signals, we augment this phase with medical knowledge graphs to suppress factual drift and mitigate reasoning noise at a granular level.Subsequently, we introduce a targeted factual perturbation mechanism to foster fine-grained discrimination between valid fact utilization and subtle factual misapplications. Extensive experiments across diverse benchmarks demonstrate notable improvements over existing methods, validating the effectiveness of MedCoach.

Large reasoning models (LRMs) are strong multi-step problem solvers, yet their long reasoning traces may gradually drift into unsafe compliance when exposed to harmful or adversarial prompts. Existing safety alignment methods can reduce harmful outputs; however, they often rely on broad refusal or supervision over full safe reasoning traces, which may weaken the reasoning behavior that makes LRMs useful. We introduce Intent–Risk–Decision (IRD), a lightweight prefix-tuning method that steers only the beginning of the reasoning trace: the model first emits a prompt-intent tag, a risk statement, and a compliance decision, and then continues under its original reasoning policy. Across three LRMs and evaluations covering harmfulness, jailbreak robustness, over-refusal, and reasoning ability, IRD substantially reduces unsafe outputs while preserving reasoning performance. In particular, IRD lowers average harmfulness to 1.2–7.4% across the evaluated backbones and keeps near-lossless reasoning performance, %within 0.2 average points of the base models, showing that targeted early steering can improve safety without supervising complete safe traces or full model fine-tuning. Code and data are available at https://github.com/ReXT9527/IRD.

Recent advances in large reasoning models (LRMs) have made machine unlearning more challenging, as protected facts or unsafe rationales may surface in intermediate chain-of-thought (CoT) traces before the final answer is produced. Existing unlearning objectives typically suppress the target content or redirect internal representations, but they never specify how the post-forgetting trajectory should continue, which can lead to hallucinated substitutes, malformed boundaries, or repetitive outputs. We argue that LRM unlearning should instead learn a natural forgetting trajectory: a coherent non-disclosing CoT followed by a stable refusal-style answer that replace the original disclosure. To this end, we propose Guided Answer-Reasoning Distillation (GUARD), which converts model-generated unsafe disclosures into safe-exit trajectories, aligns a frozen LRM via guidance tokens, and distills the guided behavior into model parameters.To address the lack of metrics for replacement quality beyond leakage, we further introduce Natural Forgetting Reasoning Score (NFRS), which captures structural stability, fluency, and unsupported substitutes in forgotten outputs. Extensive experiments on R-TOFU and a STAR-1-derived harmful-intent setting show that GUARD substantially reduces unsafe and privacy disclosures across two widely adopted distilled LRMs while preserving reasoning utility. Codes are available at https://github.com/zeyu-Yan/GUARD

Developing reliable models for clinical tasks, such as Medical Question Answering (QA), is severely constrained by the limited availability of high-quality, expert-annotated training data. This challenge is exacerbated by stringent privacy requirements and the impracticality of utilizing large open-source corpora or proprietary cloud APIs within resource-limited clinical settings. To address these obstacles, we introduce SAGE (Semantic Anchor-Guided Evolution), a novel data synthesis framework that enables small, locally deployed models to generate high-quality medical training data. SAGE leverages lightweight, publicly available taxonomies such as MeSH as semantic anchors, imposing a structured prior to effectively guide and ground the data generation process. At its core, SAGE iteratively interleaves atomic (individual concept-based) and associative (relation-based) synthesis, bootstrapping training data from minimal seeds. This approach eliminates the need for large collections of medical documents or reliance on external APIs, providing a practical solution for on-premises data creation. Extensive experiments across multiple medical question-answering benchmarks demonstrate that models fine-tuned with SAGE-synthesized data consistently outperform those trained using self-derived or conventional document-based paradigms, highlighting tangible improvements in data efficiency and resource utilization for medical LLM development.

Multi-task learning in the medical domain remains a significant challenge for large language models (LLMs), plagued by conflicting task objectives and substantial computational overhead associated with full fine-tuning. Although parameter-efficient tuning methods help reduce these resource demands, their effectiveness is limited by two critical issues: (i) rigid allocation of expert capacity across layers, and (ii) lack of regularization mechanisms to partition shared components from task-specific signals. To address these limitations, we propose LAK-LoRA, a novel architecture that enables adaptive, depth-aware scaling of expert capacity, allowing more efficient resource allocation throughout the model. Furthermore, to resolve conflicting task objectives, LAK-LoRA divides experts into two distinct groups dedicated to modeling general and task-specific features, respectively. This separation is implemented via a feature separator module, which disentangles representations into shared and task-specific subspaces before assigning them to the appropriate experts. To foster diversity and minimize representational redundancy, orthogonality constraints are imposed among specialized experts. Extensive experiments on benchmark medical datasets demonstrate that LAK-LoRA achieves the best average performance.

Jailbreaks against multimodal large language models (MLLMs) typically operate through image-text inputs. After each target response, the attacker must decide how the new feedback should shape the next image-text query: which clues should be stated in the text, which should appear as visual evidence, and how the text should refer to the image.Existing multimodal attacks typically rely on fixed image-text pairs or candidate search, limiting their ability to adapt across turns.Text-only multi-turn attacks can adapt to prior responses, but they do not decide how response-derived clues should be expressed across image and text.In this paper, we formulate this missing step as trace-to-query realization and propose  ̲Memory- ̲Augmented  ̲Visual  ̲Querying (MAViQ), a black-box framework that converts response-derived clues into visual and textual constraints, refines the query after observing the generated image, and propagates feedback to subsequent turns.Extensive experiments on two benchmarks across six target MLLMs show that MAViQ achieves average attack success rates (ASRs) of 70.60% on MM-SafetyBench and 82.13% on JailbreakV-mini, outperforming the strongest baseline by 5.90% and 8.32%, respectively. Our code is available at https://github.com/jixingtian/Maviq.

The safety mechanisms of large language models (LLMs) exhibit notable fragility, as even fine-tuning on datasets without harmful content may still undermine their safety capabilities. Meanwhile, existing safety alignment methods predominantly rely on the fine-tuning process, which inadvertently leads to the increased complexity and computational resources required. To address these issues, we introduce LSSF, a novel safety re-alignment framework with Low-Rank Safety Subspace Fusison. Our proposed method exploits the low-rank characteristics of safety information in LLMs by constructing a low-rank projection matrix to extract the principal components of safety vectors. Notably, this projection matrix represents the low-rank safety subspace of the LLMs, which we have observed to remain stable during fine-tuning process and is isolated from the model’s general capabilities. These principal components are used to effectively restore safety alignment when combined with fine-tuned LLMs through linear arithmetic. Additionally, to account for the varying encoding densities of safety information across different layers of LLMs, we propose a novel metric called safety singular value entropy. This metric quantifies the encoding density and allows for the dynamic computation of the safety-critical rank for each safety vector. Extensive experiments demonstrate that our proposed post-hoc alignment method can effectively restore the safety alignment of fine-tuned models with minimal impact on their performance on downstream tasks.

By inducing privacy attacks on NLP models, attackers can obtain sensitive information such as training data and model parameters, etc. Although researchers have studied, in-depth, several kinds of attacks in NLP models, they are non-systematic analyses. It lacks a comprehensive understanding of the impact caused by the attacks. For example, we must consider which scenarios can apply to which attacks, what the common factors are that affect the performance of different attacks, the nature of the relationships between different attacks, and the influence of various datasets and models on the effectiveness of the attacks, etc. Therefore, we need a benchmark to holistically assess the privacy risks faced by NLP models. In this paper, we present a privacy attack and defense evaluation benchmark in the field of NLP, which includes the conventional/small models and large language models (LLMs). This benchmark supports a variety of models, datasets, and protocols, along with standardized modules for comprehensive evaluation of attacks and defense strategies. Based on the above framework, we present a study on the association between auxiliary data from different domains and the strength of privacy attacks. And we provide an improved attack method in this scenario with the help of Knowledge Distillation (KD). Furthermore, we propose a chained framework for privacy attacks. Allowing a practitioner to chain multiple attacks to achieve a higher-level attack objective. Based on this, we provide some defense and enhanced attack strategies. The code for reproducing the results can be found at https://anonymous.4open.science/r/nlp_doctor-AF48

Text-to-image synthesis for the Chinese language poses unique challenges due to its large vocabulary size, and intricate character relationships. While existing diffusion models have shown promise in generating images from textual descriptions, they often neglect domain-specific contexts and lack robustness in handling the Chinese language. This paper introduces PAI-Diffusion, a comprehensive framework that addresses these limitations. PAI-Diffusion incorporates both general and domain-specific Chinese diffusion models, enabling the generation of contextually relevant images. It explores the potential of using LoRA and ControlNet for fine-grained image style transfer and image editing, empowering users with enhanced control over image generation. Moreover, PAI-Diffusion seamlessly integrates with Alibaba Cloud’s Platform for AI, providing accessible and scalable solutions. All the Chinese diffusion model checkpoints, LoRAs, and ControlNets, including domain-specific ones, are publicly available. A user-friendly Chinese WebUI and the diffusers-api elastic inference toolkit, also open-sourced, further facilitate the easy deployment of PAI-Diffusion models in various local and cloud environments, making it a valuable resource for Chinese text-to-image synthesis.

Recently, Contrastive Visual-Language Pre-training (CLIP) has demonstrated remarkable capability in various Visual Language Understanding (VLU) tasks. Yet, most CLIP-based methods require tasks-specific designs and sufficient training data. In this paper, we introduce a simple yet efficient paradigm for low-resource VLU named XtremeCLIP, which involves very few trainable parameters to improve the generalization ability of the trained models. In our XtremeCLIP framework, we reformulate a series of VLU tasks as a unified open-book affinity-matching problem. Furthermore, to handle the insufficient supervised signals in small datasets, we adopt contrastive learning to utilize the implicit sorting information of ground-truth labels to provide more supervised cues. Extensive experiments over multiple datasets on visual entailment, visual question answering, and image classification show that XtremeCLIP consistently outperforms existing baselines in low-resource settings.

As privacy issues are receiving increasing attention within the Natural Language Processing (NLP) community, numerous methods have been proposed to sanitize texts subject to differential privacy. However, the state-of-the-art text sanitization mechanisms based on a relaxed notion of metric local differential privacy (MLDP) do not apply to non-metric semantic similarity measures and cannot achieve good privacy-utility trade-offs. To address these limitations, we propose a novel Customized Text sanitization (CusText) mechanism based on the original 𝜖-differential privacy (DP) definition, which is compatible with any similarity measure.Moreover, CusText assigns each input token a customized output set to provide more advanced privacy protection at the token level.Extensive experiments on several benchmark datasets show that CusText achieves a better trade-off between privacy and utility than existing mechanisms.The code is available at https://github.com/sai4july/CusText.