Current evaluations of tool-integrated LLM agents typically focus on end-to-end tool-usage evaluation while neglecting their stability. This limits their real-world applicability, as various internal or external factors can cause agents to crash or behave abnormally. Our research addresses this by investigating whether agents are vulnerable to errors throughout the entire tool invocation process, including reading tool documentation, selecting tools and generating parameters, and processing the tool's response. Through extensive experiments, we observe that agents are highly susceptible to errors at each stage and agents based on open-source models are more vulnerable than those based on proprietary models. We also find that increasing the model size does not significantly improve tool invocation reasoning and may make agents more vulnerable to attacks resembling normal user instructions. This highlights the importance of evaluating agent stability and offers valuable insights for future LLM development and evaluation.

Today's text-to-image generative models are trained on millions of images sourced from the Internet, each paired with a detailed caption produced by Vision-Language Models (VLMs). This part of the training pipeline is critical for supplying the models with large volumes of high-quality image-caption pairs during training. However, recent work suggests that VLMs are vulnerable to stealthy adversarial attacks, where adversarial perturbations are added to images to mislead the VLMs into producing incorrect captions. In this paper, we explore the feasibility of adversarial mislabeling attacks on VLMs as a mechanism to poisoning training pipelines for text-to-image models. Our experiments demonstrate that VLMs are highly vulnerable to adversarial perturbations, allowing attackers to produce benign-looking images that are consistently miscaptioned by the VLM models. This has the effect of injecting strong "dirty-label" poison samples into the training pipeline for text-to-image models, successfully altering their behavior with a small number of poisoned samples. We find that while potential defenses can be effective, they can be targeted and circumvented by adaptive attackers. This suggests a cat-and-mouse game that is likely to reduce the quality of training data and increase the cost of text-to-image model development. Finally, we demonstrate the real-world effectiveness of these attacks, achieving high attack success (over 73%) even in black-box scenarios against commercial VLMs (Google Vertex AI and Microsoft Azure).

Internal world models (WMs) enable agents to understand the world's state and predict transitions, serving as the basis for advanced deliberative reasoning. Recent large Vision-Language Models (VLMs), such as OpenAI o3, GPT-4o and Gemini, exhibit potential as general-purpose WMs. While the latest studies have evaluated and shown limitations in specific capabilities such as visual understanding, a systematic evaluation of VLMs' fundamental WM abilities remains absent. Drawing on comparative psychology and cognitive science, we propose a two-stage framework that assesses Perception (visual, spatial, temporal, quantitative, and motion) and Prediction (mechanistic simulation, transitive inference, compositional inference) to provide an atomic evaluation of VLMs as WMs. Guided by this framework, we introduce WM-ABench, a large-scale benchmark comprising 23 fine-grained evaluation dimensions across 6 diverse simulated environments with controlled counterfactual simulations. Through 660 experiments on 15 latest commercial and open-source VLMs, we find that these models exhibit striking limitations in basic world modeling abilities. For instance, almost all models perform at near-random accuracy when distinguishing motion trajectories. Additionally, they lack disentangled understanding -- e.g., some models tend to believe blue objects move faster than green ones. More rich results and analyses reveal significant gaps between VLMs and human-level world modeling.

This study explores Machine Translationese (MTese) -- the linguistic peculiarities of machine translation outputs -- focusing on the under-researched English-to-Chinese language pair in news texts. We construct a large dataset consisting of 4 sub-corpora and employ a comprehensive five-layer feature set. Then, a chi-square ranking algorithm is applied for feature selection in both classification and clustering tasks. Our findings confirm the presence of MTese in both Neural Machine Translation systems (NMTs) and Large Language Models (LLMs). Original Chinese texts are nearly perfectly distinguishable from both LLM and NMT outputs. Notable linguistic patterns in MT outputs are shorter sentence lengths and increased use of adversative conjunctions. Comparing LLMs and NMTs, we achieve approximately 70% classification accuracy, with LLMs exhibiting greater lexical diversity and NMTs using more brackets. Additionally, translation-specific LLMs show lower lexical diversity but higher usage of causal conjunctions compared to generic LLMs. Lastly, we find no significant differences between LLMs developed by Chinese firms and their foreign counterparts.

Hybrid-based retrieval methods, which unify dense-vector and lexicon-based retrieval, have garnered considerable attention in the industry due to performance enhancement. However, despite their promising results, the application of these hybrid paradigms in Chinese retrieval contexts has remained largely underexplored. In this paper, we introduce HyReC, an innovative end-to-end optimization method tailored specifically for hybrid-based retrieval in Chinese. HyReC enhances performance by integrating the semantic union of terms into the representation model. Additionally, it features the Global-Local-Aware Encoder (GLAE) to promote consistent semantic sharing between lexicon-based and dense retrieval while minimizing the interference between them. To further refine alignment, we incorporate a Normalization Module (NM) that fosters mutual benefits between the retrieval approaches. Finally, we evaluate HyReC on the C-MTEB retrieval benchmark to demonstrate its effectiveness.

Evaluating the performance and biases of large language models (LLMs) through role-playing scenarios is becoming increasingly common, as LLMs often exhibit biased behaviors in these contexts. Building on this line of research, we introduce PapersPlease, a benchmark consisting of 3,700 moral dilemmas designed to investigate LLMs' decision-making in prioritizing various levels of human needs. In our setup, LLMs act as immigration inspectors deciding whether to approve or deny entry based on the short narratives of people. These narratives are constructed using the Existence, Relatedness, and Growth (ERG) theory, which categorizes human needs into three hierarchical levels. Our analysis of six LLMs reveals statistically significant patterns in decision-making, suggesting that LLMs encode implicit preferences. Additionally, our evaluation of the impact of incorporating social identities into the narratives shows varying responsiveness based on both motivational needs and identity cues, with some models exhibiting higher denial rates for marginalized identities. All data is publicly available at https://github.com/yeonsuuuu28/papers-please.

Zero-shot skeleton-based action recognition aims to develop models capable of identifying actions beyond the categories encountered during training. Previous approaches have primarily focused on aligning visual and semantic representations but often overlooked the importance of fine-grained action patterns in the semantic space (e.g., the hand movements in drinking water and brushing teeth). To address these limitations, we propose a Frequency-Semantic Enhanced Variational Autoencoder (FS-VAE) to explore the skeleton semantic representation learning with frequency decomposition. FS-VAE consists of three key components: 1) a frequency-based enhancement module with high- and low-frequency adjustments to enrich the skeletal semantics learning and improve the robustness of zero-shot action recognition; 2) a semantic-based action description with multilevel alignment to capture both local details and global correspondence, effectively bridging the semantic gap and compensating for the inherent loss of information in skeleton sequences; 3) a calibrated cross-alignment loss that enables valid skeleton-text pairs to counterbalance ambiguous ones, mitigating discrepancies and ambiguities in skeleton and text features, thereby ensuring robust alignment. Evaluations on the benchmarks demonstrate the effectiveness of our approach, validating that frequency-enhanced semantic features enable robust differentiation of visually and semantically similar action clusters, improving zero-shot action recognition.

The field of artificial intelligence is currently abuzz with discussions surrounding "agentic AI" or "AI agents." However, despite the widespread excitement, the term agent itself often lacks a precise, universally agreed-upon definition within these conversations. Recently, significant focus has shifted towards agents built upon large language models (LLMs), leveraging some reasoning and language understanding capabilities to execute complex tasks, interact with external tools, and learn from feedback [53, 56, 63, 66, 67]. This move towards more autonomous, goal-directed LLM systems represents a promising yet challenging frontier in AI development. During this time, AI algorithms have also reached superhuman performance in numerous tasks such as game playing [9,57,62,65] and text and image processing [2, 15, 51]. On the other hand, there are still significant obstacles that modern AI has yet to overcome. Grosz [21] proposed a revised Turing Test to create: "A computer team member that can behave, over the long term and in uncertain, dynamic environments, in such a way that people on the team will not notice that it is not human."

Automated scientific idea generation systems have made remarkable progress, yet the automatic evaluation of idea novelty remains a critical and underexplored challenge. Manual evaluation of novelty through literature review is labor-intensive, prone to error due to subjectivity, and impractical at scale. To address these issues, we propose the Idea Novelty Checker, an LLM-based retrieval-augmented generation (RAG) framework that leverages a two-stage retrieve-then-rerank approach. The Idea Novelty Checker first collects a broad set of relevant papers using keyword and snippet-based retrieval, then refines this collection through embedding-based filtering followed by facet-based LLM re-ranking. It incorporates expert-labeled examples to guide the system in comparing papers for novelty evaluation and in generating literature-grounded reasoning. Our extensive experiments demonstrate that our novelty checker achieves approximately 13% higher agreement than existing approaches. Ablation studies further showcases the importance of the facet-based re-ranker in identifying the most relevant literature for novelty evaluation.

Large Language Models (LLMs) have demonstrated immense advances in a wide range of natural language tasks. However, these models are susceptible to hallucinations and errors on particularly temporal understanding tasks involving multiple entities in answers. In such tasks, they fail to associate entities with accurate time intervals, generate a complete list of entities in answers or reason about events associated with specific temporal bounds. Existing works do not extensively evaluate the abilities of the model to perform implicit and explicit temporal understanding in a list answer construction setup. To bridge this gap, we propose the Time referenced List based Question Answering or TLQA benchmark that requires structured answers in list format aligned with corresponding time periods. Our TLQA benchmark, requires both list construction and temporal understanding simultaneously, which to the best of our knowledge has not been explored in prior benchmarks. We investigate the temporal understanding and list construction capabilities of state-of-the-art generative models on TLQA in closed-book and open-domain settings. Our findings reveal significant shortcomings in current models, particularly their inability to provide complete answers and temporally align facts in a closed-book setup and the need to improve retrieval in open-domain setup, providing clear future directions for research on TLQA. The benchmark and code at https://github.com/elixir-research-group/TLQA.