Multi-objective preference alignment often faces alignment conflicts, where optimizing for one objective (e.g., helpfulness) degrades performance on others (e.g., harmlessness). While prior work focuses on algorithmic solutions, the intrinsic conflict within data and its theoretical impact on training remain underexplored. To bridge this gap, we introduce the principle of Reward Consistency (RC), a theory-grounded criterion that approximates the alignment conflicts via reward models. We prove that a sample mitigates conflicts if and only if it satisfies RC, thereby ensuring improvement across all objectives during optimization. Building on this, we propose Reward Consistency Sampling (RCS), an automated framework for constructing pairwise data that adheres to RC, supplemented by a relaxation strategy to enhance flexibility. Extensive experiments show that RCS brings significant and consistent performance gains, achieving an average improvement of 23.07% in both harmlessness and helpfulness during simultaneous optimization comparde to the vanilla dataset. Our data-centric approach is complementary to existing alignment algorithms and effective in both sequential and simultaneous optimization scenarios.

Accurate comprehension and controllable generation of emotion and rhetoric are pivotal for enhancing the reasoning capabilities of large language models (LLMs). Existing studies mostly rely on external optimizations, lacking in-depth exploration of internal representation mechanisms, thus failing to achieve fine-grained steering at the neuron level. A handful of works on neurons are confined to emotions, neglecting rhetoric neurons and their intrinsic connections. Traditional neuron masking also exhibits counterintuitive phenomena, making reliable verification of neuron functionality infeasible. To address these issues, we systematically investigate the neurons representation mechanisms and inherent associations of 6 emotion categories and 4 core rhetorical devices. We propose a neuron identification framework that integrates multi-dimensional screening, and design an adaptive masking method incorporating dynamic filtering, attenuation masking, and feedback optimization, enabling reliable functional validation of neuron functionality. Through neuron regulation, we achieve directed induction of non-target sentences and enhancement of emotion tasks via rhetoric neurons. Experiments on 5 commonly used datasets validate the effectiveness of our method, providing a novel paradigm for the fine-grained steering of emotion and rhetoric expressions in LLMs.

Traditional industrial agents rely on modular pipelines, including Router, Retriever, Planner, Executor, Responder, Reviewer and so on, which inevitably fracture into a labyrinth of ad-hoc patches, leading to cascading errors and high latency. We propose OneModel, an applicable paradigm shift from external workflows to internalized knowledge representation. Unlike modular systems that slice fluid user intents into static steps, OneModel consolidates complex business logic and SOPs directly into the model’s parameters.Through Continual Pre-training (CPT) and logic-compilation SFT, we transform fragmented business rules into the model’s intuitive reasoning within a unified attention space. Deployed in our global financial service system, OneModel effectively breaks the impossible triangle of latency, accuracy, and complexity. Online A/B testing demonstrates end-to-end latency reduction of more than 50% (18.7s → 8s) while the Intelligent Resolution Rate (IRR) jumps from 64.3% to 83.3%. The results demonstrate our paradigm OneModel effectively replaces brittle engineering logic with internalized cognitive intuition, offering a scalable and future-proof blueprint for transitioning industrial agents from complex, error-prone workflow to unified model architectures.

Reinforcement Learning with Human Feedback (RLHF) is the key to the success of large language models (LLMs) in recent years. In this work, we first introduce the concepts of knowledge breadth and knowledge depth, which measure the comprehensiveness and depth of an LLM or knowledge source respectively. We reveal that the imbalance in the number of prompts and responses can lead to a potential disparity in breadth and depth learning within alignment tuning datasets by showing that even a simple uniform method for balancing the number of instructions and responses can lead to significant improvements. Building on this, we further propose Balanced Preference Optimization (BPO), designed to dynamically augment the knowledge depth of each sample. BPO is motivated by the observation that the usefulness of knowledge varies across samples, necessitating tailored learning of knowledge depth. To achieve this, we introduce gradient-based clustering, estimating the knowledge informativeness and usefulness of each augmented sample based on the model’s optimization direction. Our experimental results across various benchmarks demonstrate that BPO outperforms other baseline methods in alignment tuning while maintaining training efficiency. Furthermore, we conduct a detailed analysis of each component of BPO, providing guidelines for future research in preference data optimization.

In light of the widespread deployment of Large Language Models (LLMs), the responsibility for safeguarding and regulating LLM-generated content has taken on heightened significance. Recent advancements in LLM-based moderation methods, e.g., LlamaGuard, have demonstrated remarkable promise in identifying safety risks associated with both inputs and outputs in human-AI interactions. However, integrating LLM-based safeguards into a chatbot system requires an additional inference stage involving a moderation LLM with billions of parameters, which significantly increases computational costs and reduces overall efficiency. In this paper, we demonstrate that simply learning a classification head on the last-layer hidden states of the dialogue model provides a strong capability to identify harmful contents. The classification head, referred to as ShieldHead, serves as an auxiliary branch paralleled with next-token-prediction LM head, enabling the detection of potential risks in past text sequences. Additionally, a label disambiguation technique is employed to supervise ShieldHead with both token-level and sentence-level labels, which further enhances its performance. ShieldHead exhibits remarkable efficiency during inference, providing real-time moderation results alongside token-wise streaming output during the chatbot system’s decoding phase. Extensive experimental results demonstrate the superiority of the proposed framework: a state-of-the-art performance on the XSTest and SafeRLHF datasets while running at a speed about 300× faster (<1ms) than previous LLM-based moderation models with ** 99%** less parameters of LlamaGuard.

Key Information Extraction (KIE) is a challenging multimodal task aimed at extracting structured value entities from visually rich documents. Despite recent advancements, two major challenges remain. First, existing datasets typically feature fixed layouts and a limited set of entity categories, while current methods are based on a full-shot setting that is difficult to apply in real-world scenarios, where new entity categories frequently emerge. Secondly, current methods often treat key entities simply as parts of the OCR-parsed context, neglecting the positive impact of the relationships between key-value entities. To address the first challenge, we introduce a new large-scale, human-annotated dataset, Complex Layout document for Key Information Extraction (CLEX). Comprising 5,860 images with 1,162 entity categories, CLEX is larger and more complex than existing datasets. It also primarily focuses on the zero-shot and few-shot KIE tasks, which are more aligned with real-world applications. To tackle the second challenge, we propose the Parallel Pointer-based Network (P²Net). This model frames KIE as a pointer-based classification task and effectively leverages implicit relationships between key-value entities to enhance extraction. Its parallel extraction mechanism enables simultaneous and efficient extraction of multiple results. Experiments on widely-used datasets, including SROIE, CORD, and the newly introduced CLEX, demonstrate that P²Net outperforms existing state-of-the-art methods (including GPT-4V) while maintaining fast inference speeds.

Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet they often refuse to answer legitimate queries—a phenomenon known as overrefusal. Overrefusal typically stems from over-conservative safety alignment, causing models to treat many reasonable prompts as potentially risky. To systematically understand this issue, we probe and leverage the models’ safety decision boundaries to analyze and mitigate overrefusal. Our findings reveal that overrefusal is closely tied to misalignment at these boundary regions, where models struggle to distinguish subtle differences between benign and harmful content. Building on these insights, we present RASS, an automated framework for prompt generation and selection that strategically targets overrefusal prompts near the safety boundary. By harnessing steering vectors in the representation space, RASS efficiently identifies and curates boundary-aligned prompts, enabling more effective and targeted mitigation of overrefusal. This approach not only provides a more precise and interpretable view of model safety decisions but also seamlessly extends to multilingual scenarios. We have explored the safety decision boundaries of various LLMs and construct the MORBench evaluation set to facilitate robust assessment of model safety and helpfulness across multiple languages. Code and datasets are available at https://github.com/Master-PLC/RASS.

Large Language Models (LLMs) excel in various natural language processing tasks but remain vulnerable to generating harmful content or being exploited for malicious purposes. Although safety alignment datasets have been introduced to mitigate such risks through supervised fine-tuning (SFT), these datasets often lack comprehensive risk coverage. Most existing datasets focus primarily on lexical diversity while neglecting other critical dimensions. To address this limitation, we propose a novel analysis framework to systematically measure the risk coverage of alignment datasets across three essential dimensions: Lexical Diversity, Malicious Intent, and Jailbreak Tactics. We further introduce TRIDENT, an automated pipeline that leverages persona-based, zero-shot LLM generation to produce diverse and comprehensive instructions spanning these dimensions. Each harmful instruction is paired with an ethically aligned response, resulting in two datasets: TRIDENT-Core, comprising 26,311 examples, and TRIDENT-Edge, with 18,773 examples. Fine-tuning Llama 3.1-8B on TRIDENT-Edge demonstrates substantial improvements, achieving an average 14.29% reduction in Harm Score, and a 20% decrease in Attack Success Rate compared to the best-performing baseline model fine-tuned on the WildBreak dataset.

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.