论文检索

Haoyuan Shi, Yunxin li, Nanhao Deng, Zhenran Xu, Xinyu Chen, Longyue Wang, Baotian Hu, Min Zhang

The evolution of video generation toward complex, multi-shot narratives has exposed a critical deficit in current evaluation methods. Existing benchmarks remain anchored to single-shot paradigms, lacking the comprehensive story assets and cross-shot metrics required to assess long-form coherence and appeal. To bridge this gap, we introduce \textbf{MSVBench}, the first comprehensive benchmark featuring hierarchical scripts and reference images tailored for \textbf{M}ulti-\textbf{S}hot \textbf{V}ideo generation. We propose a hybrid evaluation framework that synergizes the high-level semantic reasoning of Large Multimodal Models (LMMs) with the fine-grained perceptual rigor of domain-specific expert models. Evaluating 20 video generation methods across diverse paradigms, we find that current models—despite strong visual fidelity—primarily behave as visual interpolators rather than true world models. We further validate the reliability of our benchmark by demonstrating a state-of-the-art Spearman’s rank correlation of \textbf{0.944} with human judgments. Finally, MSVBench extends beyond evaluation by providing a scalable supervisory signal. Fine-tuning a lightweight model on its pipeline-refined reasoning traces yields human-aligned performance comparable to commercial models like Gemini-2.5-Flash.

Zaitang LI, Pin-Yu Chen, Tsung-Yi Ho

Large Language Models (LLMs) have revolutionized generative tasks, but concerns about their trustworthiness and vulnerability to adversarial attacks persist. This paper introduces the Generative Robustness Evaluation (GRE) Score, a novel metric designed to assess LLMs’ resilience against adversarial red teaming attempts that may compromise model compliance and elicit undesired responses. Our approach utilizes conditional generation for synthetic text creation, offering an attack-independent evaluation of LLM robustness. By calculating the margin in refusal scores, we quantify the robustness of LLMs in an attack-agnostic manner. We evaluate our method on five dimensions with specified datasets, encompassing ethical considerations, safety protocols, and potential misuse scenarios. We present four contributions: (1) The GRE Score framework, which establishes a textual robustness certificate for LLMs against adversarial red teaming attempts, providing a theoretical foundation for quantifying model resilience. (2) Comprehensive evaluations across five dimensions using eight prominent LLMs, validating GRE Scores with adversarial red teaming attacks. Our method demonstrates a consistent ranking of LLM robustness when compared to the attack-based model ranking on TrustLLM (CITATION) while achieving a significant 5-8x speedup compared to traditional evaluation techniques. (3) Insights into the non-linear relationship between model scaling and performance, revealing that larger models do not always perform better, and an analysis of how instruction-tuning impacts robustness across LLMs. (4) The discovery that all evaluated LLMs exhibit lower performance in robustness and privacy tasks compared to other areas, highlighting a critical gap in capabilities.

Rose Sathyanathan, Kinshuk Vasisht, Danish Pruthi

Millions of users turn to AI models for their information needs. It is conceivable that a large number of user queries contain assumptions that may be factually inaccurate. Prior work notes that large language models (LLMs) often fail to challenge such erroneous assumptions, and can reinforce users’ misinformed opinions. However, given the recent advances, especially in model’s reasoning capabilities, we revisit whether large reasoning models (LRMs) can reason about the underlying assumptions and respond to user queries appropriately. We construct queries with varying degrees of presuppositions spanning health, science, and general knowledge, and use it to evaluate several widely-deployed models When compared to non-reasoning models, we find that reasoning models achieve a slightly higher accuracy (2-11%), but they still fail to challenge a large fraction (26-42%) of false presuppositions. Further, reasoning models remain susceptible to how strongly the presupposition is expressed.

Xinyu Zhong, Peng Lan, Zhifang Liao

Large Language Models (LLMs) have demonstrated a remarkable ability in code generation, yet ensuring the security and functionality of the produced code remains a critical challenge. Existing security code generation methods often rely solely on abstract security knowledge, typically resulting in a suboptimal trade-off: they either produce code with lingering vulnerabilities due to insufficient guidance or sacrifice functionality for the sake of absolute security. To address this limitation, we propose SAFENOTE, a novel framework that integrates a Security Error Notebook and a Function Error Notebook to transform failure experiences into concrete, actionable guidance. This method facilitates a form of contrastive guidance during inference, effectively steering the LLMs away from identified vulnerabilities while preserving functional correctness. Extensive experiment results across five LLMs on CodeGuard+ and LiveCodeBench benchmarks demonstrate the effectiveness of our method. Specifically, SAFENOTE achieves a substantial leap in SP@1 metric, with GPT-4o-mini performance improving from 60.21% to 66.7% on CodeGuard+. Furthermore, SAFENOTE provides security and functional guidance that generalizes effectively to “unseen” CWE scenarios, significantly outperforming existing baselines.

Shu Zhou, Rui Ling, Junan Chen, Xin Wang, Tao Fan, Hao Wang

Scaling test-time compute through extended chains of thought has become a dominant paradigm for improving large language model reasoning. However, existing research implicitly assumes that longer thinking always yields better results. This assumption remains largely unexamined. We systematically investigate how the marginal utility of additional reasoning tokens changes as compute budgets increase. We find that marginal returns diminish substantially at higher budgets and that models exhibit overthinking, where extended reasoning is associated with abandoning previously correct answers. Furthermore, we show that optimal thinking length varies across problem difficulty, suggesting that uniform compute allocation is suboptimal. Our cost-aware evaluation framework reveals that stopping at moderate budgets can reduce computation significantly while maintaining comparable accuracy.

Jianing Hao, Yuhe Wu, Yuanjian Xu, Shichang Meng, Shuai Yuan, Wei Zeng, Zixuan Wang, Guang Zhang

Large language models (LLMs) hold great promise for business applications, yet business analysis remains inherently complex, demanding rigorous reasoning and the integration of diverse knowledge sources. Existing benchmarks typically target narrow tasks and thus leave a fundamental question unanswered: how can LLMs be reliably applied in business, and how are these applications grounded in underlying theoretical capabilities? To address this gap, we introduce BizCompass, a benchmark explicitly designed to connect theoretical foundations with practical business knowledge and applications. At the knowledge level, BizCompass covers four core domains—finance, economics, statistics, and operations management. At the application level, it structures tasks around three representative roles: the analyst, the trader, and the consultant. This dual-axis design not only exposes performance differences across realistic scenarios but also diagnoses which foundational capabilities enable or constrain success. We systematically evaluate both open-source and commercial LLMs, revealing how theoretical knowledge translates into practical performance in business. The results provide actionable insights for model selection and training optimization in real-world business contexts. All datasets and evaluation code are publicly released to support reproducibility and future research: https://bizcompass.dev.ypemc.com.

Chau Minh Pham, Zichao Wang, Puneet Mathur, Alexa Siu, Akriti Jain, Aparna Garimella, Ananya B. Sai, Nedim Lipka, Mohit Iyyer, Varun Manjunatha

Large language models are increasingly used to draft long-form multimodal documents, but their end-to-end performance on professional report generation remains systematically understudied. We introduce AnalystBench, a continually extensible benchmark of 20 real-world report generation tasks grounded in multimodal document collections, where models must process millions of input tokens to produce long-form professional reports. Using expert-validated quality checklists and groundedness evaluation, we evaluate LLMs and coding agents and find that the best model, GPT-5.1, scores highly on executive summarization tasks (exceeding 90% on quality checklists) but degrades substantially on tasks requiring long-horizon synthesis over large inputs (dropping to 25-40%). Agent-based generation substantially benefits strong closed-source models such as GPT-5.1, with checklist scores improving by 20.24 percentage points and visual coverage by 37.41 points over vanilla generation, but offers little or negative gains for open-source models like DeepSeek-R1 (-3.02 points). Expert reviewers note that while generated reports are grounded and clearly separate factual description from interpretation, they often fall short in actionability, clarity, and quantitative precision, which highlights the gap between system performance and real-world professional needs.

Yifan Wang, Yun Fu

Recent large multimodal models (LMMs) have demonstrated impressive capabilities in image understanding, yet they still struggle to perform complex reasoning on challenging multimodal problems. In this paper, we present UnAC (Understanding, Abstracting, and Checking), a multimodal prompting method that strengthens reasoning for complex multimodal tasks in LMMs (e.g., GPT-4o, Gemini 1.5, and GPT-4V). To improve image understanding and capture fine details, we propose an adaptive visual prompting strategy that enables LMMs to focus on salient regions. We further design an image-abstraction prompt to effectively extract key information from images. In addition, we introduce a gradual self-checking scheme that improves reasoning by verifying each decomposed subquestion and its answer. Extensive experiments on three public benchmarks—MathVista, MM-Vet, and MMMU—demonstrate the effectiveness of our method.

Pujun Zheng, Jiacheng Yao, Jinquan Zheng, Chenyang Gu, Guoxiu He, Jiawei Liu, Yong Huang, Tianrui Guo, Wei Lu

Large language models (LLMs) are currently applied to scientific paper evaluation by assigning an absolute score to each paper independently. However, since score scales vary across conferences, time periods, and evaluation criteria, models trained on absolute scores are prone to fitting narrow, context-specific rules rather than developing robust scholarly judgment. To overcome this limitation, we propose shifting paper evaluation from isolated scoring to collaborative ranking. In particular, we design a \textbf{C}omparison-\textbf{N}ative framework for \textbf{P}aper \textbf{E}valuation (\textbf{CNPE}), integrating comparison into both data construction and model learning. We first propose a graph-based similarity ranking algorithm to facilitate the sampling of more informative and discriminative paper pairs from a collection. We then enhance relative quality judgment through supervised fine-tuning and reinforcement learning with comparison-based rewards. At inference, the model performs pairwise comparisons over sampled paper pairs and aggregates these preference signals into a global relative quality ranking. Experimental results demonstrate that our framework achieves an average relative improvement of \textbf{21.8}% over the strong baseline DeepReview-14B, while exhibiting robust generalization to five previously unseen datasets. Our code is available at https://github.com/ECNU-Text-Computing/ComparisonReview.

Xingyuan Li, Mengyue Wu

Detecting medical conditions from speech acoustics is fundamentally a weakly-supervised learning problem: a single, often noisy, session-level label must be linked to nuanced patterns within a long, complex audio recording. This task is further hampered by severe data scarcity and the subjective nature of clinical annotations. While semi-supervised learning (SSL) offers a viable path to leverage unlabeled data, existingaudio methods often fail to address the core challenge that pathological traits are not uniformly expressed in a patient’s speech. We propose a novel, audio-only SSL framework that explicitly models this hierarchy by jointly learning from frame-level, segment-level, and session-level representations within unsegmented clinical dialogues. Our end-to-end approach dynamically aggregates these multi-granularity features and generates high-quality pseudo-labels to efficiently utilize unlabeled data. Extensive experiments show the framework is model-agnostic, robust across languages and conditions, and highly data-efficient—achieving, for instance, 90% of fully-supervised performance using only 11 labeled samples. This work provides a principled approach to learning from weak, far-end supervision in medical speech analysis.The code is available at https://github.com/fispresent/semi_pathological.

Yiming Lei, Yize Fan, Zeming Liu, Jiaji Dong, Hui Qiu, Haitao Leng, Qingjie Liu, Kehai Chen, Tingting Gao, Yunhong Wang

Recent advancements in Multimodal Large Language Models (MLLMs) have achieved significant success in understanding static pre-recorded video scenarios (e.g., event-centric or narrative-driven content). However, existing MLLMs are largely trained on datasets restricted to static content due to the scarcity of high-quality interleaved data, causing them to struggle with dynamic interactions. Distinct from pre-recorded videos, live streaming is characterized by high-density, interleaved multimodal turns, where viewer comments (danmaku) are tightly coupled with real-time audio-visual evidence and evolving dialogue context. In such settings, purely textual annotations fail to capture fine-grained visual and temporal dependencies. To bridge this gap, we introduce **Live-Aid**, the first large-scale interleaved live interaction Chinese dataset with **human-annotated**, temporally aligned video responses, spanning over **1,100 hours** and 80,037 dialogue turns across 8,053 video sessions. Building on this, we leverage these high-quality annotations within a novel multi-agent pipeline to construct evaluation tasks targeting core capabilities of live interactions. Extensive evaluations of strong Video-LLMs and Omni-LLMs reveal critical limitations in interleaved multi-turn interactions requiring temporal reasoning, highlighting the value of **Live-Aid** in advancing interleaved multimodal reasoning and dynamic audio-visual dependencies.

Xing Yang, Chengxiang Tan

In Document-level Relation Extraction (DocRE), relational facts are typically organized around a few salient entities. Explicitly capturing this topological structure is pivotal to addressing the two critical bottlenecks of the task: the extreme class imbalance and the complexity of multi-hop reasoning. Based on this insight, we first introduce the concept of the salient entity and propose a novel approach that decouples the extraction space into dense and sparse scenarios. Specifically, our approach restricts the search space for dense pairs to mitigate the dominance of the negative samples, and innovatively injects the rich semantic knowledge of salient entities to explicitly reconstruct the document for bridging disjoint evidence in multi-hop reasoning. Extensive experiments demonstrate that our approach yields consistent improvements over various backbone models and achieves advanced performance compared to existing enhancement methods.

Chi Cui, Yixin Wu, Michael Backes, Yang Zhang

Prompt injection attacks are recognized as one of the primary risks faced by LLM-integrated applications in recent years. However, common evaluation frameworks remain insufficient, lacking comprehensiveness and real-world relevance. To bridge this gap, we revisit the common evaluation framework and conduct an extensive evaluation across eight different evaluation settings, including 37 real-world applications, 185 injected tasks, 21 attack instructions, and a total of 143,745 queries. The evaluation highlights several findings. For example, real-world applications are more vulnerable to prompt injection attacks compared to those used in research settings. While complex attack instructions are more sophisticated, they are less effective than simple attack instructions. We further conduct an assessment of both prompt-level and model-level defense mechanisms and highlight their limitations in real-world applications. By exploring more diverse scenarios across different dimensions, our framework provides a solid foundation for assessing vulnerabilities in LLM-integrated applications and evaluating the efficacy of defensive strategies.

Shanshan Xu, Santosh T.Y.S.S, Barbara Plank

Human Label Variation (HLV) refers to legitimate disagreement in annotation that reflects the diversity of human perspectives rather than mere error. Long treated in NLP as \textit{noise} to be eliminated, HLV has only recently been reframed as a \textit{signal} for improving model robustness. With the rise of large language models (LLMs) and post-training methods such as human feedback-based alignment, the role of HLV has become increasingly consequential. Yet current preference-learning datasets routinely collapse multiple annotations into a single label, flattening diverse perspectives into artificial consensus. Preserving HLV is necessary not only for pluralistic alignment but also for sociotechnical safety evaluation, where model behavior must be assessed in relation to human interaction and societal context.This position paper argues that preserving HLV as an embodiment of human pluralism must be treated as a \textit{Selbstzweck}, an intrinsic value in itself. We analyze the limitations of existing preference datasets and propose actionable strategies for incorporating HLV into dataset construction to better preserve pluralistic human values.

Qingbao Huang, Cheng Yang, Jiawei Yao, Zhiyue Liu, Yi Cai, Xingmao Zhang

Metaphorical text expresses meaning through cross-domain mappings rather than literal surface content, which makes it difficult for text-to-image systems to generate semantically faithful images. We propose CMIG, a structured prompting framework inspired by Conceptual Metaphor Theory (CMT). CMIG identifies source–target mappings, filters projectable source attributes, and selects a visual realization strategy in a reproducible reasoning workflow. Experiments on DALL\cdotE 3, Imagen 2, and FLUX-1 show that CMIG consistently improves semantic alignment and yields a better overall balance of human-rated metaphor quality, visual coherence, and controllability on metaphorical prompts. To support systematic evaluation, we also construct a 3,500-instance visual metaphor benchmark.

Zhaoyang Han, Yihe Liu, Kai Zhang, Ping Li

The rapid discovery of jailbreak prompts has revealed the alarming fragility of safety alignment in frontier large language models (LLMs). While jailbreak techniques play a critical role in red-teaming and safety evaluation, existing methods exhibit three key limitations: (i) poor transferability across model families, requiring model-specific manual tuning; (ii) heavy reliance on large-scale prompt enumeration or exhaustive search, causing prohibitive query costs and poor scalability; and (iii) high sensitivity to input preprocessing or refusal-oriented fine-tuning, leading to attack failures once the underlying model is updated. To address these, we propose Experience-driven Multi-agent Jailbreak Optimization (EMJO), which couples three collaborating agents (Attacker, Analyzer, and Judge) into a closed-loop “probe–evaluate–revise” process, together with a dynamic experience bank accumulating high-quality successful prompts and reusable strategy patterns across iterations and tasks. This design enables query-efficient and transferable jailbreak optimization under black-box access. Extensive experiments on diverse LLMs demonstrate that EMJO consistently outperforms existing black-box jailbreak baselines, achieving up to 11% absolute improvement in attack success rate while reducing the average query cost by up to 7.9\times across two benchmark datasets. These results indicate that EMJO offers an effective and scalable paradigm for systematic jailbreak discovery.

Yifei Yao, Hanrong Zhang, Mengnan Du

Understanding the internal representations of large language models (LLMs) remains a central challenge for interpretability research. Sparse autoencoders (SAEs) offer a promising solution by decomposing activations into interpretable features, but existing approaches rely on fixed sparsity constraints that fail to account for input complexity. We propose AdaptiveK SAE (Adaptive Top K Sparse Autoencoders), a novel framework that dynamically adjusts sparsity levels based on the semantic complexity of each input. Leveraging linear probes, we demonstrate that context complexity is linearly encoded in LLM representations, and we use this signal to guide feature allocation during training. Experiments across ten language models demonstrate that this complexity-driven adaptation outperforms fixed-sparsity approaches on reconstruction fidelity, explained variance, cosine similarity and interpretability metrics while eliminating the burden of extensive hyperparameter tuning. Our code is available at: https://github.com/hiyukie/adaptiveK.

Kewei Chen, Yayu Long, Shuai Li, Mingsheng Shang

Diffusion policies have demonstrated exceptional performance in embodied AI. However, their iterative denoising process results in high latency, and existing acceleration methods often sacrifice physical consistency. To address this, we propose ElasticFlow, a distillation-free, physics-consistent one-step policy framework. We reconstruct the Mean Field Theory by directly modeling the average velocity field, enabling a direct single-step mapping from noise to action. Addressing the Temporal Heterogeneity of robotic tasks, we introduce the Elastic Time Horizons mechanism. This mechanism effectively overcomes Spectral Bias by explicitly encoding control granularity, achieving efficient alignment between semantic instructions and physical execution horizons. Experiments on benchmarks such as LIBERO, CALVIN, and RoboTwin demonstrate that ElasticFlow achieves efficient 1-NFE inference (approximately 71Hz). Furthermore, it outperforms state-of-the-art methods, including OpenVLA and \pi_0, on long-horizon tasks, highlighting its potential for efficient, robust, and semantically aligned control.

Xinhao Zhuang, Qiongyu Tian, Yalin Chen, Tianle Xin, Yongyong Fu, Yunchao Ling, Guoqing Zhang

High-quality annotated data is crucial for NLP, yet manual annotation is costly and difficult to scale in low-resource settings. Large Language Models (LLMs) have demonstrated strong zero-shot and few-shot generalization in NLP tasks, but existing annotation tools either lack LLM support or use LLMs only as one-off pre-annotation engines, without incorporating collaboration or quality control, compromising data reliability. We present BNLP, a text annotation platform that embeds LLM-assisted labeling into a quality-aware collaborative workflow. BNLP treats LLM outputs as intermediate, revisable states and integrates multi-role collaboration, iterative review cycles, and consistency analysis to enable continuous quality monitoring while preserving efficiency gains. BNLP also natively supports AI-ready formats such as Excel and JSON, ensuring seamless data flow from manual annotation to model training. Experiments show that BNLP reduces annotation time by 74.3% and improves annotation quality by 11.6% over purely manual annotation in LLM-assisted settings.

Anas Mohamed, Azal Ahmad Khan, Xinran Wang, Ahmad Faraz Khan, Shuwen Ge, Saman Bahzad Khan, Ayaan Ahmad, Ali Anwar

Generative AI can now synthesize strikingly realistic images from text, yet output quality remains highly sensitive to how prompts are phrased. Direct Preference Optimization (DPO) offers a lightweight, off-policy alternative to RL for automatic prompt engineering, but its token-level regularization leaves semantic inconsistency unchecked as prompts that win higher preference scores can still drift away from the user’s intended meaning. We introduce Sem-DPO, a variant of DPO that preserves semantic consistency yet retains its simplicity and efficiency. Sem-DPO adjusts the DPO loss using a weight based on how different the winning prompt is from the original, reducing the impact of training examples that are semantically misaligned. We provide the first analytical bound on semantic drift for preference-tuned prompt generators, showing that Sem-DPO keeps learned prompts within a provably bounded neighborhood of the original text. On three standard text-to-image prompt-optimization benchmarks and three language models, Sem-DPO achieves 8–12% higher CLIP similarity and 5–9% higher human-preference scores (HPSv2.1, PickScore) than DPO, while also outperforming state-of-the-art prompt optimization baselines as well as several DPO variants. These findings suggest that strong flat baselines augmented with semantic weighting should become the new standard for prompt-optimization studies and lay the groundwork for broader, semantics-aware preference optimization in language models.