With the commercialization of short video platforms (SVPs), the demand for compliance auditing of advertising content has grown rapidly. The rise of large vision-language models (VLMs) offers new opportunities for automating ad content moderation. However, short video advertising scenarios present unique challenges due to \textbf{data drift (DD)} and \textbf{label drift (LD)}. DD refers to rapid shifts in data distribution caused by advertisers to evade platform review mechanisms. LD arises from the evolving and increasingly standardized review guidelines of SVPs, which effectively alter the classification boundaries over time. Despite the significance of these phenomena, there is currently a lack of benchmark tools designed to evaluate model performance under such conditions. To address this gap, we propose \textbf{AdDriftBench (ADB)}. The ADB dataset consists of 3,480 short video ads, including 2,280 examples labeled under data drift scenarios, designed to evaluate the generalization capabilities of VLMs under rapidly shifting content distributions. An additional 1,200 examples represent label drift scenarios, aimed at assessing VLMs’ abilities in instruction following and fine-grained semantic understanding under varying auditing standards. Through extensive experiments on 16 open-source VLMs, we find that current models perform moderately in short video advertising contexts, particularly in handling fine-grained semantics and adapting to shifting instructions. Our dataset will be made publicly available.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Generating radiology reports from medical images has garnered sufficient attention in the research community. While existing methods have demonstrated promise, they often tend to generate reports that are factually incomplete and inconsistent, fail to focus on informative regions within an image, and impose strong annotation assumptions, such as bounding box annotations, image level annotations (which can be challenging to obtain) for model training. In this paper, we propose MediVLM, a vision language model (VLM) for radiology report generation from medical images. The proposed model consists of a pre-trained object detector to extract the salient anatomical regions from the images, an image encoder, a text encoder, a module to align the visual and text representations, a cross attention layer to fuse the two representations and finally, a transformer based decoder to generate the final report. MediVLM can generate radiology reports even when no reports are available for training; this is an extremely useful feature, as curating such reports is a labor-intensive task. Further, it computes a severity score (depicting the seriousness of a patient’s medical condition) from the generated radiology reports, which can be used to prioritize patients who need immediate medical attention. Our extensive empirical analyses on three benchmark datasets corroborate the promise and potential of our method against competing baselines. Our code is open-sourcedin our project webpage at: https://sites.google.com/view/medivlm/home
CATCH: A Novel Data Synthesis Framework for High Therapy Fidelity and Memory-Driven Planning Chain of Thought in AI Counseling
PDF ↗Recently, advancements in AI counseling based on large language models have shown significant progress. However, existing studies employ a one-time generation approach to synthesize multi-turn dialogue samples, resulting in low therapy fidelity and failing to capture the decision-making rationale behind each response. In this work, we propose CATCH, a novel data synthesis framework designed to address these challenges. Specifically, to improve therapy fidelity, we introduce the Progressive Dialogue Synthesis strategy, which extracts goals, resources, and solutions from a client’s self-report, organizes them into structured outlines, and then incrementally generates stage-aligned counseling dialogues. To capture decision-making rationale behind each response, we propose the Memory-Driven Dynamic Planning thinking pattern that integrates memory enhancement, global planning, and strategy reasoning; a collaborative multi-agent optimizer then leverages MDP to attach explicit chain-of-thought to each dialogue turn. Extensive experiments and human evaluations demonstrate that CATCH significantly enhances fidelity and logical coherence in AI counseling.
With the rapid development of large language models (LLMs) in math reasoning, the accuracy of models on existing math benchmarks has gradually approached 90% or even higher. More challenging math benchmarks are hence urgently in need to satisfy the increasing evaluation demands. To bridge this gap, we propose HighMATH. Problems in HighMATH are collected according to 3 criteria: problem complexity, knowledge domain diversity and fine-grained annotations. We collect 5,293 problems from Chinese senior high school mathematics exams published in 2024, covering 8 subjects and 7 levels of difficulty, with each problem involving an average of more than 2.4 knowledge points. We conduct a thorough evaluation of latest LLMs on the curated HighMATH, including o1-like models. Evaluation results demonstrate that the accuracy of advanced LLMs on HighMATH is significantly lower than that on previous math reasoning benchmarks. This gap even exceeds 30%. Our results also suggest that properly trained smaller LLMs may have great potential in math reasoning. Our data is available at https://github.com/tjunlp-lab/HighMATH.
With the rapid advancement of large language models (LLMs), natural language processing (NLP) has achieved remarkable progress. Nonetheless, significant challenges remain in handling texts with ambiguity, polysemy, or uncertainty. We introduce the Fuzzy Reasoning Chain (FRC) framework, which integrates LLM semantic priors with continuous fuzzy membership degrees, creating an explicit interaction between probability-based reasoning and fuzzy membership reasoning. This transition allows ambiguous inputs to be gradually transformed into clear and interpretable decisions while capturing conflicting or uncertain signals that traditional probability-based methods cannot. We validate FRC on sentiment analysis tasks, where both theoretical analysis and empirical results show that it ensures stable reasoning and facilitates knowledge transfer across different model scales. These findings indicate that FRC provides a general mechanism for managing subtle and ambiguous expressions with improved interpretability and robustness.
Large Language Models (LLMs) excel in various Natural Language Processing (NLP) tasks but remain vulnerable to misinformation, particularly in multi-turn dialogues where misleading context accumulates. Existing benchmarks, such as TruthfulQA and FEVER, assess factual accuracy in isolated queries but fail to evaluate LLMs’ resilience to misinformation in interactive settings. To address this limitation, we introduce MisinfoBench, a multi-dimensional benchmark designed to assess LLMs’ ability to discern, resist, and reject misinformation. MisinfoBench defines three core dimensions—Discernment, Resistance, and Principled Refusal—across seven evaluation tasks, systematically testing misinformation identification, contextual resistance, and the rejection of coercive false premises. It includes a dataset of 4,962 multi-turn dialogues and 2,000 misinformation-based question-answer pairs, capturing diverse misinformation scenarios. We evaluate 16 LLMs, revealing substantial disparities in misinformation resilience: proprietary models outperform open-source counterparts, while multi-turn dialogues and cross-lingual settings exacerbate misinformation susceptibility. Our findings highlight persistent vulnerabilities in LLMs’ misinformation defenses, emphasizing the need for context-aware training, adversarial robustness, and principled reasoning. MisinfoBench establishes a rigorous standard for evaluating misinformation resilience, advancing the development of more trustworthy AI systems.
Large language models (LLMs) are increasingly used to automate data analysis through executable code generation. Yet, data science tasks often admit multiple statistically valid solutions—for example, different modeling strategies—making it critical to understand the reasoning behind analyses, not just their outcomes. While manual review of LLM-generated code can help ensure statistical soundness, it is labor-intensive and requires expertise. A more scalable approach is to evaluate the underlying workflows—the logical plans guiding code generation. However, it remains unclear how to assess whether an LLM-generated workflow supports reproducible implementations.To address this, we present **AIRepr**, an **A**nalyst–**I**nspector framework for automatically evaluating and improving the **repr**oducibility of LLM-generated data analysis workflows. Our framework is grounded in statistical principles and supports scalable, automated assessment. We introduce two novel reproducibility-enhancing prompting strategies and benchmark them against standard prompting across 15 analyst–inspector LLM pairs and 1,032 tasks from three public benchmarks. Our findings show that workflows with higher reproducibility also yield more accurate analyses, and that reproducibility-enhancing prompts substantially improve both metrics. This work provides a foundation for transparent, reliable, and efficient human–AI collaboration in data science. Our code is publicly available: [https://github.com/Anonymous-2025-Repr/LLM-DS-Reproducibility](https://github.com/Anonymous-2025-Repr/LLM-DS-Reproducibility)
Recently, Large Language Models (LLMs) have shown strong potential in recommendation tasks due to their broad world knowledge and reasoning capabilities. However, applying them to serendipity-oriented recommendation remains challenging, mainly due to a domain gap of LLMs in modeling personalized user behavior and the scarcity of labeled serendipitous interactions. In this paper, we introduce **SOLAR** (**S**erendipity-**O**ptimized **L**anguage model **A**ligned for **R**ecommendation), a two-stage framework that addresses these challenges. To alleviate label scarcity, we adopt a weak supervision strategy: a sequential ID-based recommender generates candidate items, which are then reranked by an LLM acting as a preference judge to produce serendipity-aware pseudo-labels. To bridge the domain gap, we propose a domain-adaptive instruction tuning method (SUN) that aligns LLMs with recommendation tasks. Experiments on three real-world datasets show that **SOLAR** consistently improves both accuracy and serendipity over strong baselines, showing its effectiveness in enabling more diverse, user-centric recommendations. Code and dataset are released at [https://github.com/SOLAR2025ARR/SOLAR](https://github.com/SOLAR2025ARR/SOLAR).
Learning to Describe Implicit Changes: Noise-robust Pre-training for Image Difference Captioning
PDF ↗Image Difference Captioning (IDC) methods have advanced in highlighting subtle differences between similar images, but their performance is often constrained by limited training data. Using Large Multimodal Models (LMMs) to describe changes in image pairs mitigates data limits but adds noise. These change descriptions are often coarse summaries, obscuring fine details and hindering noise detection. In this work, we improve IDC with a noise-robust approach at both data and model levels. We use LMMs with structured prompts to generate fine-grained change descriptions during data curation. We propose a Noise-Aware Modeling and Captioning (NAMC) model with three modules: Noise Identification and Masking (NIM) to reduce noisy correspondences, Masked Image Reconstruction (MIR) to correct over-masking errors, and Fine-grained Description Generation (FDG) to produce coherent change descriptions. Experiments on four IDC benchmarks show that NAMC, pre-trained on our large-scale data, outperforms streamlined architectures and achieves competitive performance with LLM-finetuned methods, offering better inference efficiency.
Multimodal machine translation (MMT) aims to enhance translation quality by integrating visual information. However, existing methods often extract visual features using pre-trained models while learning text features from scratch, leading to representation imbalance. These methods are also prone to being misled by redundant visual information, which results in suboptimal performance. To address these challenges, we propose CAMT, a novel cross-modal VQA-augmented MMT method. CAMT aligns image-source text pairs and image-question text pairs through dual-text contrastive learning, thereby improving semantic consistency across modalities. Additionally, we design an effective strategy for generating question–answer pairs to enhance fine-grained alignment and filter out irrelevant visual noise, while also addressing the scarcity of VQA annotations. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of the proposed CAMT framework, which consistently outperforms state-of-the-art MMT methods across multiple evaluation metrics.
Large language models (LLMs) have shown remarkable performance in various tasks but often fail to handle queries that exceed their knowledge and capabilities, leading to incorrect or fabricated responses. This paper addresses the need for LLMs to recognize and refuse infeasible tasks due to the requests surpassing their capabilities. We conceptualize four main categories of infeasible tasks for LLMs, which cover a broad spectrum of hallucination-related challenges identified in prior literature. We develop and benchmark a new dataset comprising diverse infeasible and feasible tasks to evaluate multiple LLMs’ abilities to decline infeasible tasks. Furthermore, we explore the potential of increasing LLMs’ refusal capabilities with fine-tuning. Experiments validate the effectiveness of our trained models, offering promising directions for refining the operational boundaries of LLMs in real applications.
As NLP models become increasingly integrated into real-world applications, it becomes clear that there is a need to address the fact that models often rely on and generate conflicting information. Conflicts could reflect the complexity of situations, changes that need to be explained and dealt with, difficulties in data annotation, and mistakes in generated outputs. In all cases, disregarding the conflicts in data could result in undesired behaviors of models and undermine NLP models’ reliability and trustworthiness. This survey categorizes these conflicts into three key areas: (1) natural texts on the web, where factual inconsistencies, subjective biases, and multiple perspectives introduce contradictions; (2) human-annotated data, where annotator disagreements, mistakes, and societal biases impact model training; and (3) model interactions, where hallucinations and knowledge conflicts emerge during deployment. While prior work has addressed some of these conflicts in isolation, we unify them under the broader concept of conflicting information, analyze their implications, and discuss mitigation strategies. We highlight key challenges and future directions for developing conflict-aware NLP systems that can reason over and reconcile conflicting information more effectively.
Fine-Tuning Encoder-Decoder Models with Contrastive Learning for In-Context Distractor Generation
PDF ↗Distractor generation is the task of automatically generating plausible yet incorrect options (i.e., distractors) for fill-in-the-blank and multiple-choice questions. In assessment, distractors must be contextually relevant to the given question and answer. Even though recent research works focus on fine-tuning pre-trained encoder-decoder models with data augmentation techniques to generate distractors, these models often fail to capture the full semantic representation of a given question-answer and related distractors. The augmentation methods often rely on expanding the quantity of proposed candidates (i.e., questions or distractors), which can introduce noise into the models without necessarily enhancing their understanding of the deeper semantic relationships between question-answer and related distractors. This paper introduces a novel distractor generation model based on contrastive learning to train the model to recognize essential semantic features necessary to generate in-context distractors. The extensive experiments on two public datasets indicate that contrastive learning introduces a strong baseline model to the distractor generation task. It significantly outperforms recent models, increasing the NDCG@3 score from 24.68 to 32.33 on the MCQ dataset and from 26.66 to 36.68 on the SciQ dataset.
Catch Me If You Can? Not Yet: LLMs Still Struggle to Imitate the Implicit Writing Styles of Everyday Authors
PDF ↗As large language models (LLMs) become increasingly integrated into personal writing tools, a critical question arises: can LLMs faithfully imitate an individual’s writing style from just a few examples? Personal style is often subtle and implicit, making it difficult to specify through prompts yet essential for user-aligned generation. This work presents a comprehensive evaluation of state-of-the-art LLMs’ ability to mimic personal writing styles via in-context learning from a small number of user-authored samples. We introduce an ensemble of complementary metrics—including authorship attribution, authorship verification, style matching, and AI detection—to robustly assess style imitation. Our evaluation spans over 40,000 generations per model across domains such as news, email, forums, and blogs, covering writing samples from more than 400 real-world authors. Results show that while LLMs can approximate user styles in structured formats like news and email, they struggle with nuanced, informal writing in blogs and forums. Further analysis on various prompting strategies such as number of demonstrations reveal key limitations in effective personalization. Our findings highlight a fundamental gap in personalized LLM adaptation and the need for improved techniques to support implicit, style-consistent generation. To aid future research and for reproducibility, we open-source our data and code.
Large Language Models (LLMs) often hallucinate, producing unfaithful or factually incorrect outputs by misrepresenting the provided context or incorrectly recalling internal knowledge. Recent studies have identified specific attention heads within the Transformer architecture, known as retrieval heads, responsible for extracting relevant contextual information. We hypothesise that masking these retrieval heads can induce hallucinations and that contrasting the outputs of the base LLM and the masked LLM can reduce hallucinations. To this end, we propose Decoding by Contrasting Retrieval Heads (DeCoRe), a novel training-free decoding strategy that amplifies information found in the context and model parameters. DeCoRe mitigates potentially hallucinated responses by dynamically contrasting the outputs of the base LLM and the masked LLM, using conditional entropy as a guide. Our extensive experiments confirm that DeCoRe improves performance on tasks requiring high contextual faithfulness, such as summarisation (XSum by 18.6%), instruction following (MemoTrap by 10.9%), and open-book question answering (NQ-Open by 2.4% and NQ-Swap by 5.5%).
We cast nested named entity recognition (NNER) as a sequence labeling task by leveraging prior work that linearizes constituency structures, effectively reducing the complexity of this structured prediction problem to straightforward token classification. By combining these constituency linearizations with pretrained encoders, our method captures nested entities while performing exactly n tagging actions. Our approach achieves competitive performance compared to less efficient systems, and it can be trained using any off-the-shelf sequence labeling library.
We introduce AraSafe, the first large-scale native Arabic safety benchmark for large language models (LLMs), addressing the pressing need for culturally and linguistically representative evaluation resources. The dataset comprises 12K naturally occurring, human-written Arabic prompts containing both harmful and non-harmful content across diverse domains, including linguistics, social studies, and science. Each prompt was independently annotated by two experts into one of nine fine-grained safety categories, including ‘Safe/Not Harmful’, ‘Illegal Activities’, ‘Violence or Harm’, ‘Privacy Violation’, and ‘Hate Speech’. Additionally, to support training classifiers for harmful content and due to the imbalanced representation of harmful content in the natural dataset, we create a synthetic dataset of additional 12K harmful prompts generated by GPT-4o via carefully designed prompt engineering techniques. We benchmark a number of Arabic-centric and multilingual models in the 7 to 13B parameter range, including Jais, AceGPT, Allam, Fanar, Llama-3, Gemma-2, and Qwen3, as well as BERT-based fine-tuned classifier models on detecting harmful prompts. GPT-4o was used as an upper-bound reference baseline. Our evaluation reveals critical safety blind spots in Arabic LLMs and underscores the necessity of localized, culturally grounded benchmarks for building responsible AI systems.
Prompt injection and jailbreak attacks remain a critical vulnerability for deployed large language models (LLMs), allowing adversaries to bypass safety protocols and extract sensitive information. To address this, we present Proxy Barrier (ProB), a lightweight defense that interposes a proxy LLM between the user and the target model. The proxy LLM is tasked solely to repeat the user input, and any failure indicates the presence of an attempt to reveal or override system instructions, leading the malicious request to be detected and blocked before it reaches the target model. ProB therefore requires no access to model weights or prompts, and is deployable entirely at the API level. Experiments across multiple model families demonstrate that ProB achieves state-of-the-art resilience against prompt leakage and jailbreak attacks. Notably, our approach outperforms baselines and achieves up to 98.8% defense effectiveness, and shows robust protection across both open and closed-source LLMs when suitably paired with proxy models, while also keeping response quality intact.
Large language models are prone to generating hallucination that deviates from factual information. Existing studies mainly focus on detecting the presence of hallucinations but lack a systematic classification approach, which hinders deeper exploration of their characteristics. To address this, we introduce the concept of belief state, which quantifies the model’s confidence in its own responses. We define the belief state of the model based on self-consistency, leveraging answer repetition rates to label confident and uncertain states. Based on this, we categorize factuality hallucination into two types: Overconfident Hallucination and Unaware Hallucination. Furthermore, we propose BAFH, a factuality hallucination type detection method. By training a classifier on model’s hidden states, we establish a link between hidden states and belief states, enabling efficient and automatic hallucination type detection. Experimental results demonstrate the effectiveness of BAFH and the differences between hallucination types.
Large language models (LLMs) are increasingly used for data generation. However, creating evaluation benchmarks raises the bar for this emerging paradigm. Benchmarks must target specific phenomena, penalize exploiting shortcuts, and be challenging. Through two case studies, we ask whether LLMs are ready to meet these demands—by generating reasoning-over-text benchmarks and comparing them to those that were created through careful crowdsourcing. Specifically, we evaluate both the *validity* and *difficulty* of LLM-generated versions of two high-quality reading comprehension datasets: CondaQA, which evaluates reasoning about negation, and DROP, which targets reasoning about quantities. We find that prompting LLMs can produce variants of these datasets that are often valid according to the annotation guidelines, at a fraction of the cost of the original crowdsourcing effort. However, we show that they are *less challenging for LLMs* than their human-authored counterparts. This finding sheds light on what may have been lost by generating evaluation data with LLMs, and calls for critically reassessing the immediate use of this increasingly prevalent approach to benchmark creation.