Abstract--General-purpose large language models (LLMs) have demonstrated remarkable generative and reasoning capabilities but remain limited in healthcare and caregiving applications due to two key deficiencies: factual unreliability and a lack of empathetic communication. These shortcomings pose significant risks in sensitive contexts where users, particularly nonprofessionals and caregivers, seek medically relevant guidance or emotional reassurance. T o address these challenges, we introduce a Direct Preference Optimization (DPO)-based alignment framework designed to improve factual correctness, semantic coherence, and human-centric qualities such as empathy, politeness, and simplicity in caregiver-patient dialogues. Our approach fine-tunes domain-adapted Large Language Models (LLMs) using pairwise preference data, where preferred responses reflect supportive and accessible communication styles while rejected ones represent prescriptive or overly technical tones. Empirical evaluations across multiple open and proprietary LLMs show that our DPO-tuned models achieve higher semantic alignment, improved factual accuracy, and stronger human-centric evaluation scores compared to baseline and commercial alternatives such as Google's medical dialogue systems. These improvements demonstrate that preference-based alignment offers a scalable and transparent pathway toward developing trustworthy, empathetic, and clinically informed AI assistants for caregiver and healthcare communication. Caring for individuals with chronic or neuro-degenerative conditions such as Alzheimer's disease and dementia requires not only clinical coordination but also constant emotional resilience. Family caregivers and care partners often become the primary interpreters of medical information, navigating complex treatment decisions, behavioral changes, and communication challenges on a daily basis. LLMs have rapidly become integrated into everyday life. They can explain complex ideas in plain language, adjust to a user's tone, and offer a sense of understanding that static websites cannot. For caregivers seeking clear, kind, and quick answers, these systems can feel like an always-available companion in moments of doubt or stress.

Large Language Models (LLMs) have attracted significant attention for classification tasks, offering a flexible alternative to trusted classical machine learning models like LightGBM through zero-shot prompting. However, their reliability for structured tabular data remains unclear, particularly in high-stakes applications like financial risk assessment. Our study systematically evaluates LLMs and generates their SHAP values on financial classification tasks. Our analysis shows a divergence between LLMs self-explanation of feature impact and their SHAP values, as well as notable differences between LLMs and LightGBM SHAP values. These findings highlight the limitations of LLMs as standalone classifiers for structured financial modeling, but also instill optimism that improved explainability mechanisms coupled with few-shot prompting will make LLMs usable in risk-sensitive domains.

Activation steering offers a promising approach to controlling the behavior of Large Language Models by directly manipulating their internal activations. However, most existing methods struggle to jointly steer multiple attributes, often resulting in interference and undesirable trade-offs. To address this challenge, we propose Multi-Subspace Representation Steering (MSRS), a novel framework for effective multi-attribute steering via subspace representation fine-tuning. MSRS reduces inter-attribute interference by allocating orthogonal subspaces to each attribute, isolating their influence within the model's representation space. MSRS also incorporates a hybrid subspace composition strategy: it combines attribute-specific subspaces for unique steering directions with a shared subspace for common steering directions. A dynamic weighting function learns to efficiently integrate these components for precise control. During inference, MSRS introduces a token-level steering mechanism that dynamically identifies and intervenes on the most semantically relevant tokens, enabling fine-grained behavioral modulation. Experimental results show that MSRS significantly reduces attribute conflicts, surpasses existing methods across a range of attributes, and generalizes effectively to diverse downstream tasks. These models often exhibit undesirable behaviors, including toxicity, bias, or factual inaccuracies, rooted in the complex and opaque representations learned during training (Y ang et al., 2024c; Wang et al., 2025a; Zhang et al., 2025b). Effectively controlling these behaviors without compromising model performance remains an open research problem (Jiao et al., 2025). Recently, activation steering methods offer a promising avenue for behavior adjustment by manipulating model activations post-training (Im & Li, 2025). Compared to fine-tuning, they offer lightweight control without the need for retraining or access to model weights, enabling scalable adaptation to diverse downstream tasks. These approaches derive an activation steering vector from the difference between the activations of positive and negative samples, applying it during inference to guide outputs toward desired properties without altering model parameters (Rimsky et al., 1 The left output contains biases and falsehoods. MSRS reduces attribute conflicts by separating steering spaces and using a shared subspace to capture common properties, enabling better integration.

Language models (LMs) are often adapted through supervised fine-tuning (SFT) to specialize their capabilities for downstream tasks. However, in typical scenarios where the fine-tuning data is limited, e.g., compared to pre-training, SFT can lead LMs to overfit, causing them to rely on spurious patterns within the target task or to compromise other broadly useful capabilities as a side effect of narrow specialization. In this paper, we propose Learning-from-the-Undesirable (LfU), a simple yet effective regularization scheme for SFT to mitigate overfitting issues when fine-tuning LMs with limited data. Specifically, we aim to regularize the fine-tuning process to favor solutions that are resilient to "undesirable" model updates, e.g., gradient ascent steps that steer the model toward undesirable behaviors. To this end, we propose a novel form of consistency regularization that directly aligns internal representations of the model with those after an undesirable update. By leveraging representation-level data augmentation through undesirable updates, LfU effectively promotes generalization under limited data. Our experiments on diverse LM downstream tasks show that LfU serves as an effective prior that enhances adaptability while preserving pretrained knowledge. For example, our LM from LfU achieves a 16.8% average improvement on math tasks compared to vanilla SFT on the same dataset, where the latter even leads to degraded performance on those tasks. Furthermore, LfU exhibits improved robustness to prompt variations, e.g., yielding a 92.1% lower standard deviation in output performances compared to SFT, highlighting its versatile effects.

Understanding human attitudes, preferences, and behaviors through social surveys is essential for academic research and policymaking. Y et traditional surveys face persistent challenges, including fixed-question formats, high costs, limited adaptability, and difficulties ensuring cross-cultural equivalence. While recent studies explore large language models (LLMs) to simulate survey responses, most are limited to structured questions, overlook the entire survey process, and risks under-representing marginalized groups due to training data biases. We introduce AlignSurvey, the first benchmark that systematically replicates and evaluates the full social survey pipeline using LLMs. It defines four tasks aligned with key survey stages: social role modeling, semi-structured interview modeling, attitude stance modeling and survey response modeling. It also provides task-specific evaluation metrics to assess alignment fidelity, consistency, and fairness at both individual and group levels, with a focus on demographic diversity. To support AlignSurvey, we construct a multi-tiered dataset architecture: (i) the Social Foundation Corpus, a cross-national resource with 44K+ interview dialogues and 400K+ structured survey records; and (ii) a suite of Entire-Pipeline Survey Datasets, including the expert-annotated AlignSurvey-Expert (ASE) and two nationally representative surveys for cross-cultural evaluation. We release the SurveyLM family, obtained through two-stage fine-tuning of open-source LLMs, and offer reference models for evaluating domain-specific alignment. All datasets, models, and tools are available at github and huggingface to support transparent and socially responsible research.

Reinforcement learning with verifiable rewards (RLVR) has been shown to enhance the reasoning capabilities of large language models (LLMs), enabling the development of large reasoning models (LRMs). However, LRMs such as DeepSeek-R1 and OpenAI o1 often generate verbose responses containing redundant or irrelevant reasoning step-a phenomenon known as overthinking-which substantially increases computational costs. Prior efforts to mitigate this issue commonly incorporate length penalties into the reward function, but we find they frequently suffer from two critical issues: length collapse and training collapse, resulting in sub-optimal performance. To address them, we propose a pipeline for training a Conciseness Reward Model (CRM) that scores the conciseness of reasoning path. Additionally, we introduce a novel reward formulation named Conciseness Reward Function (CRF) with explicit dependency between the outcome reward and conciseness score, thereby fostering both more effective and more efficient reasoning. From a theoretical standpoint, we demonstrate the superiority of the new reward from the perspective of variance reduction and improved convergence properties. Besides, on the practical side, extensive experiments on five mathematical benchmark datasets demonstrate the method's effectiveness and token efficiency, which achieves an 8.1% accuracy improvement and a 19.9% reduction in response token length on Qwen2.5-7B. Furthermore, the method generalizes well to other LLMs including Llama and Mistral. The implementation code and datasets are publicly available for reproduction: https://anonymous.4open.science/r/CRM.

Large language models (LLMs) have achieved remarkable progress across diverse tasks, yet their internal mechanisms remain largely opaque. In this work, we address a fundamental question: to what extent can the original input text be recovered from a single last-token representation within an LLM? We propose Rep2Text, a novel framework for decoding full text from last-token representations. Rep2Text employs a trainable adapter that projects a target model's internal representations into the embedding space of a decoding language model, which then autoregressively reconstructs the input text. Experiments on various model combinations (Llama-3.1-8B, Gemma-7B, Mistral-7B-v0.1, Llama-3.2-3B) demonstrate that, on average, over half of the information in 16-token sequences can be recovered from this compressed representation while maintaining strong semantic integrity and coherence. Furthermore, our analysis reveals an information bottleneck effect: longer sequences exhibit decreased token-level recovery while preserving strong semantic integrity. Besides, our framework also demonstrates robust generalization to out-of-distribution medical data.

Abstract--Large Language Models are fundamental actors in the modern IT landscape dominated by AI solutions. However, security threats associated with them might prevent their reliable adoption in critical application scenarios such as government organizations and medical institutions. For this reason, commercial LLMs typically undergo a sophisticated censoring mechanism to eliminate any harmful output they could possibly produce. These mechanisms maintain the integrity of LLM alignment by guaranteeing that the models respond safely and ethically. In response to this, attacks on LLMs are a significant threat to such protections, and many previous approaches have already demonstrated their effectiveness across diverse domains. Existing LLM attacks mostly adopt a generate-and-test strategy to craft malicious input. T o improve the comprehension of censoring mechanisms and design a targeted attack, we propose an Explainable-AI solution that comparatively analyzes the behavior of censored and uncensored models to derive unique exploitable alignment patterns. Then, we propose XBreaking, a novel approach that exploits these unique patterns to break the security and alignment constraints of LLMs by targeted noise injection. Our thorough experimental campaign returns important insights about the censoring mechanisms and demonstrates the effectiveness and performance of our approach. Nowadays, Large Language Models (LLMs, for short) represent the most promising and relevant advancement in the field of Artificial Intelligence. These complex deep learning models are trained on massive datasets that cover almost all aspects of people's daily lives, thus granting them the capability of generating, understanding, and processing human language. For this reason, their integration as support tools is becoming pervasive with applications spanning from text editor and proofreading to virtual assistant and personalized text generation. However, the diffusion of this technology, especially in critical domains such as government organizations and medical institutions, imposes the assessment of their security and privacy characteristics.

Modern approaches to enhancing Large Language Models' factual accuracy and knowledge utilization face a fundamental trade-off: non-parametric retrieval-augmented generation (RAG) provides flexible access to external knowledge but suffers from high inference latency and shallow integration, while parametric fine-tuning methods like LoRA risk catastrophic forgetting and degraded general capabilities. In this work, we propose MLP Memory, a lightweight parametric module that learns to internalize retrieval patterns without explicit document access. By pretraining an MLP to imitate a $k$NN retriever's behavior on the entire pretraining dataset, we create a differentiable memory component that captures the benefits of retrieval-based knowledge access in a fully parametric form. Our architecture integrates this pretrained MLP Memory with Transformer decoders through simple probability interpolation, yielding 17.5\% and 24.1\% scaling gains on WikiText-103 and Web datasets, respectively. It further achieves 12.3\% relative improvement on five question-answering benchmarks and 5.2 points absolute gain across nine general NLP tasks, while reducing hallucinations by up to 10 points on HaluEval. Moreover, MLP Memory delivers 2.5$\times$ faster inference than RAG with superior accuracy. Our findings show that learning retrieval patterns parametrically bridges the gap between efficient inference and effective knowledge access, offering a practical alternative to both RAG and fine-tuning approaches.