Measuring and Detecting Harmful AI Sycophancy

Sycophantic responses are becoming pervasive in large language models (LLMs), and prior work has pointed out that some of them could be harmful. This paper focuses on one harmful sycophancy: preference-induced stance reversal sycophancy (PSRS), where a model reverses an initial stance merely to align with a user's stated preference. While existing research mainly measures how sycophantic a model is, we go further and ask whether PSRS can also be detected automatically from a single response. To investigate this at scale, we introduce CAP (Contrastive Anchor Probing), a framework for collecting labeled PSRS data. Applying CAP to 17 open- and closed-source LLMs, we collect 290,460 labeled responses across 12 everyday-advice domains. We organize our study around three research questions. (1) How often does PSRS occur? (2) How well can it be detected? (3) How does detection generalize to unseen models? We first reveal that PSRS rates range from 5% to 56% across LLMs, with more capable models being less sycophantic. Next, we show that detecting PSRS is feasible from the response text alone, and detectors need to learn subtle PSRS patterns from the training data. Because new LLMs appear rapidly, detectors inevitably encounter unseen models, making cross-model generalization an important framework goal. We demonstrate that detection performance drops on unseen models and propose an initial approach to address this challenge. We will release our dataset and code to support future research.

TextCloak: Thwarting Unauthorized LLM Exploitation via RL-Driven Unlearnable Text

The rapid development of Large Language Models (LLMs) has led to significant advances across a wide range of language tasks, while simultaneously raising growing concerns about unauthorized data exploitation and privacy leakage. Unlearnable examples (UEs) offer a promising defense by introducing carefully designed perturbations into data such that models trained on them exhibit degraded utility. However, existing methods for text protection are primarily designed for classification tasks (e.g., sentiment analysis) in discriminative language models and often rely on injecting class-specific linguistic cues, which limits their effectiveness in the open-ended generation settings of LLMs. In this work, we propose TextCloak, an RL-driven framework for protecting textual data against unauthorized LLM exploitation. TextCloak employs a generative policy that transforms batches of clean text into unlearnable examples while preserving semantic fidelity and linguistic naturalness. To optimize the policy, we introduce GRPO-UE, which rewards generated unlearnable text based on the downstream degradation they induce in fine-tuned surrogate LLMs and updates the generator parameters via group-relative policy optimization. This bi-level optimization enables the generator to discover generalizable protective patterns beyond class-specific cues. Comprehensive experiments on six publicly available datasets and nine state-of-the-art LLMs demonstrate that TextCloak consistently impairs unauthorized fine-tuning while maintaining text utility for legitimate use. Further analyses establish its transferability and robustness across model architectures, training configurations, and adaptive attacks, highlighting its broad applicability as a practical defense against unauthorized LLM exploitation.

CAMO: Causality-Guided Adversarial Multimodal Domain Generalization for Crisis Classification

Crisis classification in social media aims to extract actionable disaster-related information from multimodal posts, which is a crucial task for enhancing situational awareness and facilitating timely emergency responses. However, the wide variation in crisis types makes achieving generalizable performance across unseen disasters a persistent challenge. Existing approaches primarily leverage deep learning to fuse textual and visual cues for crisis classification, achieving numerically plausible results under in-domain settings. However, they exhibit poor generalization across unseen crisis types because they 1. do not disentangle spurious and causal features, resulting in performance degradation under domain shift, and 2. fail to align heterogeneous modality representations within a shared space, which hinders the direct adaptation of established single-modality domain generalization (DG) techniques to the multimodal setting. To address these issues, we introduce a causality-guided multimodal domain generalization (MMDG) framework that combines adversarial disentanglement with unified representation learning for crisis classification. The adversarial objective encourages the model to disentangle and focus on domain-invariant causal features, leading to more generalizable classifications grounded in stable causal mechanisms. The unified representation aligns features from different modalities within a shared latent space, enabling single-modality DG strategies to be seamlessly extended to multimodal learning. Experiments on the different datasets demonstrate that our approach achieves the best performance in unseen disaster scenarios.

Are Today's LLMs Ready to Explain Well-Being Concepts?

Well-being encompasses mental, physical, and social dimensions essential to personal growth and informed life decisions. As individuals increasingly consult Large Language Models (LLMs) to understand well-being, a key challenge emerges: Can LLMs generate explanations that are not only accurate but also tailored to diverse audiences? High-quality explanations require both factual correctness and the ability to meet the expectations of users with varying expertise. In this work, we construct a large-scale dataset comprising 43,880 explanations of 2,194 well-being concepts, generated by ten diverse LLMs. We introduce a principle-guided LLM-as-a-judge evaluation framework, employing dual judges to assess explanation quality. Furthermore, we show that fine-tuning an open-source LLM using Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO) can significantly enhance the quality of generated explanations. Our results reveal: (1) The proposed LLM judges align well with human evaluations; (2) explanation quality varies significantly across models, audiences, and categories; and (3) DPO- and SFT-finetuned models outperform their larger counterparts, demonstrating the effectiveness of preference-based learning for specialized explanation tasks.

Preference Leakage: A Contamination Problem in LLM-as-a-judge

Large Language Models (LLMs) as judges and LLM-based data synthesis have emerged as two fundamental LLM-driven data annotation methods in model development. While their combination significantly enhances the efficiency of model training and evaluation, little attention has been given to the potential contamination brought by this new model development paradigm. In this work, we expose preference leakage, a contamination problem in LLM-as-a-judge caused by the relatedness between the synthetic data generators and LLM-based evaluators. To study this issue, we first define three common relatednesses between data generator LLM and judge LLM: being the same model, having an inheritance relationship, and belonging to the same model family. Through extensive experiments, we empirically confirm the bias of judges towards their related student models caused by preference leakage across multiple LLM baselines and benchmarks. Further analysis suggests that preference leakage is a pervasive issue that is harder to detect compared to previously identified biases in LLM-as-a-judge scenarios. All of these findings imply that preference leakage is a widespread and challenging problem in the area of LLM-as-a-judge. We release all codes and data at: https://github.com/David-Li0406/Preference-Leakage.

In-Group Love, Out-Group Hate: A Framework to Measure Affective Polarization via Contentious Online Discussions

Affective polarization, the emotional divide between ideological groups marked by in-group love and out-group hate, has intensified in the United States, driving contentious issues like masking and lockdowns during the COVID-19 pandemic. Despite its societal impact, existing models of opinion change fail to account for emotional dynamics nor offer methods to quantify affective polarization robustly and in real-time. In this paper, we introduce a discrete choice model that captures decision-making within affectively polarized social networks and propose a statistical inference method estimate key parameters -- in-group love and out-group hate -- from social media data. Through empirical validation from online discussions about the COVID-19 pandemic, we demonstrate that our approach accurately captures real-world polarization dynamics and explains the rapid emergence of a partisan gap in attitudes towards masking and lockdowns. This framework allows for tracking affective polarization across contentious issues has broad implications for fostering constructive online dialogues in digital spaces.

Assessing the Impact of Conspiracy Theories Using Large Language Models

Measuring the relative impact of CTs is important for prioritizing responses and allocating resources effectively, especially during crises. However, assessing the actual impact of CTs on the public poses unique challenges. It requires not only the collection of CT-specific knowledge but also diverse information from social, psychological, and cultural dimensions. Recent advancements in large language models (LLMs) suggest their potential utility in this context, not only due to their extensive knowledge from large training corpora but also because they can be harnessed for complex reasoning. In this work, we develop datasets of popular CTs with human-annotated impacts. Borrowing insights from human impact assessment processes, we then design tailored strategies to leverage LLMs for performing human-like CT impact assessments. Through rigorous experiments, we textit{discover that an impact assessment mode using multi-step reasoning to analyze more CT-related evidence critically produces accurate results; and most LLMs demonstrate strong bias, such as assigning higher impacts to CTs presented earlier in the prompt, while generating less accurate impact assessments for emotionally charged and verbose CTs.

Training Noise Token Pruning

In the present work we present Training Noise Token (TNT) Pruning for vision transformers. Our method relaxes the discrete token dropping condition to continuous additive noise, providing smooth optimization in training, while retaining discrete dropping computational gains in deployment settings. We provide theoretical connections to Rate-Distortion literature, and empirical evaluations on the ImageNet dataset using ViT and DeiT architectures demonstrating TNT's advantages over previous pruning methods.

From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge

Assessment and evaluation have long been critical challenges in artificial intelligence (AI) and natural language processing (NLP). However, traditional methods, whether matching-based or embedding-based, often fall short of judging subtle attributes and delivering satisfactory results. Recent advancements in Large Language Models (LLMs) inspire the "LLM-as-a-judge" paradigm, where LLMs are leveraged to perform scoring, ranking, or selection across various tasks and applications. This paper provides a comprehensive survey of LLM-based judgment and assessment, offering an in-depth overview to advance this emerging field. We begin by giving detailed definitions from both input and output perspectives. Then we introduce a comprehensive taxonomy to explore LLM-as-a-judge from three dimensions: what to judge, how to judge and where to judge. Finally, we compile benchmarks for evaluating LLM-as-a-judge and highlight key challenges and promising directions, aiming to provide valuable insights and inspire future research in this promising research area. Paper list and more resources about LLM-as-a-judge can be found at \url{https://github.com/llm-as-a-judge/Awesome-LLM-as-a-judge} and \url{https://llm-as-a-judge.github.io}.

LRQ-Fact: LLM-Generated Relevant Questions for Multimodal Fact-Checking

Human fact-checkers have specialized domain knowledge that allows them to formulate precise questions to verify information accuracy. However, this expert-driven approach is labor-intensive and is not scalable, especially when dealing with complex multimodal misinformation. In this paper, we propose a fully-automated framework, LRQ-Fact, for multimodal fact-checking. Firstly, the framework leverages Vision-Language Models (VLMs) and Large Language Models (LLMs) to generate comprehensive questions and answers for probing multimodal content. Next, a rule-based decision-maker module evaluates both the original content and the generated questions and answers to assess the overall veracity. Extensive experiments on two benchmarks show that LRQ-Fact improves detection accuracy for multimodal misinformation. Moreover, we evaluate its generalizability across different model backbones, offering valuable insights for further refinement.