Safety alignment approaches in large language models (LLMs) often lead to the over-refusal of benign queries, significantly diminishing their utility in sensitive scenarios. To address this challenge, we introduce FalseReject, a comprehensive resource containing 16k seemingly toxic queries accompanied by structured responses across 44 safety-related categories. We propose a graph-informed adversarial multi-agent interaction framework to generate diverse and complex prompts, while structuring responses with explicit reasoning to aid models in accurately distinguishing safe from unsafe contexts. FalseReject includes training datasets tailored for both standard instruction-tuned models and reasoning-oriented models, as well as a human-annotated benchmark test set. Our extensive benchmarking on 29 state-of-the-art (SOTA) LLMs reveals persistent over-refusal challenges. Empirical results demonstrate that supervised finetuning with FalseReject substantially reduces unnecessary refusals without compromising overall safety or general language capabilities.

Social simulation is transforming traditional social science research by modeling human behavior through interactions between virtual individuals and their environments. With recent advances in large language models (LLMs), this approach has shown growing potential in capturing individual differences and predicting group behaviors. However, existing methods face alignment challenges related to the environment, target users, interaction mechanisms, and behavioral patterns. To this end, we introduce SocioVerse, an LLM-agent-driven world model for social simulation. Our framework features four powerful alignment components and a user pool of 10 million real individuals. To validate its effectiveness, we conducted large-scale simulation experiments across three distinct domains: politics, news, and economics. Results demonstrate that SocioVerse can reflect large-scale population dynamics while ensuring diversity, credibility, and representativeness through standardized procedures and minimal manual adjustments.

Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities.

Production-grade LLM systems require robust adherence to dozens or even hundreds of instructions simultaneously. However, the instruction-following capabilities of LLMs at high instruction densities have not yet been characterized, as existing benchmarks only evaluate models on tasks with a single or few instructions. We introduce IFScale, a simple benchmark of 500 keyword-inclusion instructions for a business report writing task to measure how instruction-following performance degrades as instruction density increases. We evaluate 20 state-of-the-art models across seven major providers and find that even the best frontier models only achieve 68% accuracy at the max density of 500 instructions. Our analysis reveals model size and reasoning capability to correlate with 3 distinct performance degradation patterns, bias towards earlier instructions, and distinct categories of instruction-following errors. Our insights can help inform design of instruction-dense prompts in real-world applications and highlight important performance-latency tradeoffs. We open-source the benchmark and all results for further analysis at https://distylai.github.io/IFScale.

This paper presents the development of HKGAI-V1, a foundational sovereign large language model (LLM), developed as part of an initiative to establish value-aligned AI infrastructure specifically tailored for Hong Kong. Addressing the region's unique multilingual environment (Cantonese, Mandarin, and English), its distinct socio-legal context under the "one country, two systems" framework, and specific local cultural and value considerations, the model is built upon the DeepSeek architecture and systematically aligned with regional norms through a multifaceted full parameter fine-tuning process. It is further integrated with a retrieval-augmented generation (RAG) system to ensure timely and factually grounded information access. The core contribution lies in the design and implementation of a comprehensive, region-specific AI alignment and safety framework, demonstrated through two key achievements: 1) The successful development of HKGAI-V1 itself - which outper-forms general-purpose models in handling Hong Kong-specific culturally sensitive queries, and embodies a "governance-embedded" approach to digital sovereignty - empowers Hong Kong to exercise control over AI applications in critical sectors including public services, legal systems, and edu-cation. 2) The development of the proprietary Adversarial HK Value Benchmark, a rigorous tool for evaluating model alignment with local ethical and legal stand-ards under challenging conditions. By documenting these achievements, the paper provides not only a technological artifact but also a replicable blueprint for developing advanced, regionally focused AI systems deeply rooted in their local identities.

Large language models are increasingly used to support high-stakes decisions, potentially influencing who is granted bail or receives a loan. Naive chain-of-thought sampling can improve average decision accuracy, but has also been shown to amplify unfair bias. To address this challenge and enable the trustworthy use of reasoning models in high-stakes decision-making, we propose a framework for training a generalizable Fairness Reward Model (FRM). Our model assigns a fairness score to LLM reasoning, enabling the system to down-weight biased trajectories and favor equitable ones when aggregating decisions across reasoning chains. We show that a single Fairness Reward Model, trained on weakly supervised, LLM-annotated examples of biased versus unbiased reasoning, transfers across tasks, domains, and model families without additional fine-tuning. Applied to real-world decision-making tasks including recidivism prediction and social media moderation, we show that our approach consistently improves fairness while matching, or even surpassing, baseline accuracy.

The proliferation of hate speech has inflicted significant societal harm, with its intensity and directionality closely tied to specific targets and arguments. In recent years, numerous machine learning-based methods have been developed to detect hateful comments on online platforms automatically. However, research on Chinese hate speech detection lags behind, and interpretability studies face two major challenges: first, the scarcity of span-level fine-grained annotated datasets limits models' deep semantic understanding of hate speech; second, insufficient research on identifying and interpreting coded hate speech restricts model explainability in complex real-world scenarios. To address these, we make the following contributions: (1) We introduce the Span-level Target-Aware Toxicity Extraction dataset (STATE ToxiCN), the first span-level Chinese hate speech dataset, and evaluate the hate semantic understanding of existing models using it. (2) We conduct the first comprehensive study on Chinese coded hate terms, LLMs' ability to interpret hate semantics. (3) We propose a method to integrate an annotated lexicon into models, significantly enhancing hate speech detection performance. Our work provides valuable resources and insights to advance the interpretability of Chinese hate speech detection research.

Within a legal framework, fairness in datasets and models is typically assessed by dividing observations into predefined groups and then computing fairness measures (e.g., Disparate Impact or Equality of Odds with respect to gender). However, when sensitive attributes such as skin color are continuous, dividing into default groups may overlook or obscure the discrimination experienced by certain minority subpopulations. To address this limitation, we propose a fairness-based grouping approach for continuous (possibly multidimensional) sensitive attributes. By grouping data according to observed levels of discrimination, our method identifies the partition that maximizes a novel criterion based on inter-group variance in discrimination, thereby isolating the most critical subgroups. We validate the proposed approach using multiple synthetic datasets and demonstrate its robustness under changing population distributions - revealing how discrimination is manifested within the space of sensitive attributes. Furthermore, we examine a specialized setting of monotonic fairness for the case of skin color. Our empirical results on both CelebA and FFHQ, leveraging the skin tone as predicted by an industrial proprietary algorithm, show that the proposed segmentation uncovers more nuanced patterns of discrimination than previously reported, and that these findings remain stable across datasets for a given model. Finally, we leverage our grouping model for debiasing purpose, aiming at predicting fair scores with group-by-group post-processing. The results demonstrate that our approach improves fairness while having minimal impact on accuracy, thus confirming our partition method and opening the door for industrial deployment.

The optimization of industrial processes remains a critical challenge, particularly when no mathematical formulation of objective functions or constraints is available. This study addresses this issue by proposing a surrogate-based, data-driven methodology for optimizing complex real-world manufacturing systems using only historical process data. Machine learning models are employed to approximate system behavior and construct surrogate models, which are integrated into a tailored metaheuristic approach: Data-Driven Differential Evolution with Multi-Level Penalty Functions and Surrogate Models, an adapted version of Differential Evolution suited to the characteristics of the studied process. The methodology is applied to an extrusion process in the tire manufacturing industry, with the goal of optimizing initialization parameters to reduce waste and production time. Results show that the surrogate-based optimization approach outperforms historical best configurations, achieving a 65\% reduction in initialization and setup time, while also significantly minimizing material waste. These findings highlight the potential of combining data-driven modeling and metaheuristic optimization for industrial processes where explicit formulations are unavailable.

Vision-language models (VLMs) are increasingly applied to identify unsafe or inappropriate images due to their internal ethical standards and powerful reasoning abilities. However, it is still unclear whether they can recognize various unsafe concepts when presented in different modalities, such as text and images. To address this, we first compile the UnsafeConcepts dataset, featuring 75 unsafe concepts, i.e., ``Swastika,'' ``Sexual Harassment,'' and ``Assaults,'' along with associated 1.5K images. We then conduct a systematic evaluation of VLMs' perception (concept recognition) and alignment (ethical reasoning) capabilities. We assess eight popular VLMs and find that, although most VLMs accurately perceive unsafe concepts, they sometimes mistakenly classify these concepts as safe. We also identify a consistent modality gap among open-source VLMs in distinguishing between visual and textual unsafe concepts. To bridge this gap, we introduce a simplified reinforcement learning (RL)-based approach using proximal policy optimization (PPO) to strengthen the ability to identify unsafe concepts from images. Our approach uses reward scores based directly on VLM responses, bypassing the need for collecting human-annotated preference data to train a new reward model. Experimental results show that our approach effectively enhances VLM alignment on images while preserving general capabilities. It outperforms baselines such as supervised fine-tuning (SFT) and direct preference optimization (DPO). We hope our dataset, evaluation findings, and proposed alignment solution contribute to the community's efforts in advancing safe VLMs.