Motivated by the need for rigorous and scalable evaluation of large language models, we study contextual preference inference for pairwise comparison functionals of context-dependent preference score functions across domains. Focusing on the contextual Bradley-Terry-Luce model, we develop a semiparametric efficient estimator that automates the debiased estimation through aggregating weighted residual balancing terms across the comparison graph. We show that the efficiency is achieved when the weights are derived from a novel strategy called Fisher random walk. We also propose a computationally feasible method to compute the weights by a potential representation of nuisance weight functions. We show our inference procedure is valid for general score function estimators accommodating the practitioners' need to implement flexible deep learning methods. We extend the procedure to multiple hypothesis testing using a Gaussian multiplier bootstrap that controls familywise error and to distributional shift via a cross-fitted importance-sampling adjustment for target-domain inference. Numerical studies, including language model evaluations under diverse contexts, corroborate the accuracy, efficiency, and practical utility of our method.

In online second-hand marketplaces, multi-turn bargaining is a crucial part of seller-buyer interactions. Large Language Models (LLMs) can act as seller agents, negotiating with buyers on behalf of sellers under given business constraints. A critical ability for such agents is to track and accurately interpret cumulative buyer intents across long negotiations, which directly impacts bargaining effectiveness. We introduce a multi-turn evaluation framework for measuring the bargaining ability of seller agents in e-commerce dialogues. The framework tests whether an agent can extract and track buyer intents. Our contributions are: (1) a large-scale e-commerce bargaining benchmark spanning 622 categories, 9,892 products, and 3,014 tasks; (2) a turn-level evaluation framework grounded in Theory of Mind (ToM) with annotated buyer intents, moving beyond outcome-only metrics; and (3) an automated pipeline that extracts reliable intent from massive dialogue data.

We propose TraceRL, a trajectory-aware reinforcement learning framework for diffusion language models (DLMs) that incorporates preferred inference trajectory into post-training, and is applicable across different architectures. Equipped with a diffusion-based value model that enhances training stability, we demonstrate improved reasoning performance on complex math and coding tasks. Besides, it can also be applied to adapt block-specific models to larger blocks, which improves sampling flexibility. Employing TraceRL, we derive a series of state-of-the-art diffusion language models, namely TraDo. Although smaller than 7B-scale AR models, TraDo-4B-Instruct still consistently outperforms them across complex math reasoning tasks. TraDo-8B-Instruct achieves relative accuracy improvements of 6.1% over Qwen2.5-7B-Instruct and 51.3% over Llama3.1-8B-Instruct on mathematical reasoning benchmarks. Through curriculum learning, we also derive the first long-CoT DLM, outperforming Qwen2.5-7B-Instruct on MATH500 with an 18.1% relative accuracy gain. To facilitate reproducible research and practical applications, we release a comprehensive open-source framework for building, training, and deploying diffusion LLMs across diverse architectures. The framework integrates accelerated KV-cache techniques and inference engines for both inference and reinforcement learning, and includes implementations of various supervised fine-tuning and RL methods for mathematics, coding, and general tasks. Code and Models: https://github.com/Gen-Verse/dLLM-RL

Reinforcement learning (RL) has emerged as a powerful method for improving the reasoning abilities of large language models (LLMs). Outcome-based RL, which rewards policies solely for the correctness of the final answer, yields substantial accuracy gains but also induces a systematic loss in generation diversity. This collapse undermines real-world performance, where diversity is critical for test-time scaling. We analyze this phenomenon by viewing RL post-training as a sampling process and show that, strikingly, RL can reduce effective diversity even on the training set relative to the base model. Our study highlights two central findings: (i) a transfer of diversity degradation, where reduced diversity on solved problems propagates to unsolved ones, and (ii) the tractability of the outcome space, since reasoning tasks admit only a limited set of distinct answers. Motivated by these insights, we propose outcome-based exploration, which assigns exploration bonuses according to final outcomes. We introduce two complementary algorithms: historical exploration, which encourages rarely observed answers via UCB-style bonuses, and batch exploration, which penalizes within-batch repetition to promote test-time diversity. Experiments on standard competition math with Llama and Qwen models demonstrate that both methods improve accuracy while mitigating diversity collapse. On the theoretical side, we formalize the benefit of outcome-based exploration through a new model of outcome-based bandits. Together, these contributions chart a practical path toward RL methods that enhance reasoning without sacrificing the diversity essential for scalable deployment.

Traditional Artificial Intelligence (AI) approaches in cybersecurity exhibit fundamental limitations: inadequate conceptual grounding leading to non-robustness against novel attacks; limited instructibility impeding analyst-guided adaptation; and misalignment with cybersecurity objectives. Neuro-Symbolic (NeSy) AI has emerged with the potential to revolutionize cybersecurity AI. However, there is no systematic understanding of this emerging approach. These hybrid systems address critical cybersecurity challenges by combining neural pattern recognition with symbolic reasoning, enabling enhanced threat understanding while introducing concerning autonomous offensive capabilities that reshape threat landscapes. In this survey, we systematically characterize this field by analyzing 127 publications spanning 2019-July 2025. We introduce a Grounding-Instructibility-Alignment (G-I-A) framework to evaluate these systems, focusing on both cyber defense and cyber offense across network security, malware analysis, and cyber operations. Our analysis shows advantages of multi-agent NeSy architectures and identifies critical implementation challenges including standardization gaps, computational complexity, and human-AI collaboration requirements that constrain deployment. We show that causal reasoning integration is the most transformative advancement, enabling proactive defense beyond correlation-based approaches. Our findings highlight dual-use implications where autonomous systems demonstrate substantial capabilities in zero-day exploitation while achieving significant cost reductions, altering threat dynamics. We provide insights and future research directions, emphasizing the urgent need for community-driven standardization frameworks and responsible development practices that ensure advancement serves defensive cybersecurity objectives while maintaining societal alignment.

We participate in CheckThat! Task 2 English and explore various methods of prompting and in-context learning, including few-shot prompting and fine-tuning with different LLM families, with the goal of extracting check-worthy claims from social media passages. Our best METEOR score is achieved by fine-tuning a FLAN-T5 model. However, we observe that higher-quality claims can sometimes be extracted using other methods, even when their METEOR scores are lower.

Large Language Models (LLMs), trained on extensive datasets using advanced deep learning architectures, have demonstrated remarkable performance across a wide range of language tasks, becoming a cornerstone of modern AI technologies. However, ensuring their trustworthiness remains a critical challenge, as reliability is essential not only for accurate performance but also for upholding ethical, cultural, and social values. Careful alignment of training data and culturally grounded evaluation criteria are vital for developing responsible AI systems. In this study, we introduce the EPT (Evaluation of Persian Trustworthiness) metric, a culturally informed benchmark specifically designed to assess the trustworthiness of LLMs across six key aspects: truthfulness, safety, fairness, robustness, privacy, and ethical alignment. We curated a labeled dataset and evaluated the performance of several leading models - including ChatGPT, Claude, DeepSeek, Gemini, Grok, LLaMA, Mistral, and Qwen - using both automated LLM-based and human assessments. Our results reveal significant deficiencies in the safety dimension, underscoring the urgent need for focused attention on this critical aspect of model behavior. Furthermore, our findings offer valuable insights into the alignment of these models with Persian ethical-cultural values and highlight critical gaps and opportunities for advancing trustworthy and culturally responsible AI. The dataset is publicly available at: https://github.com/Rezamirbagheri110/EPT-Benchmark.

Effective Operation and Maintenance (O&M) is critical to reducing the Levelised Cost of Energy (LCOE) from wind power, yet the unstructured, free-text nature of turbine maintenance logs presents a significant barrier to automated analysis. Our paper addresses this by presenting a novel and reproducible framework for benchmarking Large Language Models (LLMs) on the task of classifying these complex industrial records. To promote transparency and encourage further research, this framework has been made publicly available as an open-source tool. We systematically evaluate a diverse suite of state-of-the-art proprietary and open-source LLMs, providing a foundational assessment of their trade-offs in reliability, operational efficiency, and model calibration. Our results quantify a clear performance hierarchy, identifying top models that exhibit high alignment with a benchmark standard and trustworthy, well-calibrated confidence scores. We also demonstrate that classification performance is highly dependent on the task's semantic ambiguity, with all models showing higher consensus on objective component identification than on interpretive maintenance actions. Given that no model achieves perfect accuracy and that calibration varies dramatically, we conclude that the most effective and responsible near-term application is a Human-in-the-Loop system, where LLMs act as a powerful assistant to accelerate and standardise data labelling for human experts, thereby enhancing O&M data quality and downstream reliability analysis.

As Large Language Models (LLMs) increasingly permeate human life, their security has emerged as a critical concern, particularly their ability to maintain harmless responses to malicious instructions. Although extensive methods have improved LLMs' security, they often lead to conservative, rejection-oriented responses that compromise practical usability. This presents a key challenge: how to advance the Pareto frontier between LLMs' usability and security, rather than necessitate a trade-off between them. To address this, we propose the MoGU framework, in which the intra-layer router dynamically allocates weights by sensing hidden states, thereby balancing the contributions of security-optimized and usability-optimized variants. Despite its initial potential, the MoGU framework faces limitations such as parameter redundancy and performance bottlenecks. To overcome these, we further propose an improved MoGU_v2 framework that establishes a tighter coupling between the routers and hidden states. In MoGU_v2, routers are embedded only in layers encoding highly classifiable security features, and backbone modules are activated during router optimization to enable bidirectional adaptation. MoGU_V2 exhibits strong adaptability and stable improvements across various series of LLMs, including mainstream LLMs serving as brains in various applications, on-device LLMs optimized for resource-constrained scenarios, and reasoning LLMs tailored for user interpretability. Meanwhile, even facing risks introduced by Instruction Fine-tuning, MoGU_v2 can easily restore security without compromising the task performance gains via a simple data-mix strategy. These comprehensive improvements highlight MoGU_V2 as a robust and versatile solution for mitigating security risks in real-world applications.

Retrieval-Augmented Generation (RAG) systems rely heavily on the retrieval stage, particularly the coarse-ranking process. Existing coarse-ranking optimization approaches often struggle to balance domain-specific knowledge learning with query enhencement, resulting in suboptimal retrieval performance. To address this challenge, we propose MoLER, a domain-aware RAG method that uses MoL-Enhanced Reinforcement Learning to optimize retrieval. MoLER has a two-stage pipeline: a continual pre-training (CPT) phase using a Mixture of Losses (MoL) to balance domain-specific knowledge with general language capabilities, and a reinforcement learning (RL) phase leveraging Group Relative Policy Optimization (GRPO) to optimize query and passage generation for maximizing document recall. A key innovation is our Multi-query Single-passage Late Fusion (MSLF) strategy, which reduces computational overhead during RL training while maintaining scalable inference via Multi-query Multi-passage Late Fusion (MMLF). Extensive experiments on benchmark datasets show that MoLER achieves state-of-the-art performance, significantly outperforming baseline methods. MoLER bridges the knowledge gap in RAG systems, enabling robust and scalable retrieval in specialized domains.