Current Multimodal Large Language Models (MLLMs) excel in general visual reasoning but remain underexplored in Abstract Visual Reasoning (AVR), which demands higher-order reasoning to identify abstract rules beyond simple perception. Existing AVR benchmarks focus on single-step reasoning, emphasizing the end result but neglecting the multi-stage nature of reasoning process. Past studies found MLLMs struggle with these benchmarks, but it doesn't explain how they fail. To address this gap, we introduce MultiStAR, a Multi-Stage AVR benchmark, based on RAVEN, designed to assess reasoning across varying levels of complexity. Additionally, existing metrics like accuracy only focus on the final outcomes while do not account for the correctness of intermediate steps. Therefore, we propose a novel metric, MSEval, which considers the correctness of intermediate steps in addition to the final outcomes. We conduct comprehensive experiments on MultiStAR using 17 representative close-source and open-source MLLMs. The results reveal that while existing MLLMs perform adequately on basic perception tasks, they continue to face challenges in more complex rule detection stages.

The reliability of large language models (LLMs) is greatly compromised by their tendency to hallucinate, underscoring the need for precise identification of knowledge gaps within LLMs. Various methods for probing such gaps exist, ranging from calibration-based to prompting-based methods. To evaluate these probing methods, in this paper, we propose a new process based on using input variations and quantitative metrics. Through this, we expose two dimensions of inconsistency in knowledge gap probing. (1) Intra-method inconsistency: Minimal non-semantic perturbations in prompts lead to considerable variance in detected knowledge gaps within the same probing method; e.g., the simple variation of shuffling answer options can decrease agreement to around 40%. (2) Cross-method inconsistency: Probing methods contradict each other on whether a model knows the answer. Methods are highly inconsistent -- with decision consistency across methods being as low as 7% -- even though the model, dataset, and prompt are all the same. These findings challenge existing probing methods and highlight the urgent need for perturbation-robust probing frameworks.

V oice cloning (VC)-resistant watermarking is an emerging technique for tracing and preventing unauthorized cloning. Existing methods effectively trace traditional VC models by training them on watermarked audio but fail in zero-shot VC scenarios, where models synthesize audio from an audio prompt without training. To address this, we propose V oiceMark, the first zero-shot VC-resistant watermarking method that leverages speaker-specific latents as the watermark carrier, allowing the watermark to transfer through the zero-shot VC process into the synthesized audio. Additionally, we introduce VC-simulated augmentations and V AD-based loss to enhance robustness against distortions. Experiments on multiple zero-shot VC models demonstrate that V oiceMark achieves over 95% accuracy in watermark detection after zero-shot VC synthesis, significantly outperforming existing methods, which only reach around 50%. See our code and demos at: https://huggingface.co/spaces/haiyunli/V oiceMark.

Large language model (LLM) research has grown rapidly, along with increasing concern about their limitations such as failures in reasoning, hallucinations, and limited multilingual capability. While prior reviews have addressed these issues, they often focus on individual limitations or consider them within the broader context of evaluating overall model performance. This survey addresses the gap by presenting a data-driven, semi-automated review of research on limitations of LLMs (LLLMs) from 2022 to 2025, using a bottom-up approach. From a corpus of 250,000 ACL and arXiv papers, we extract 14,648 relevant limitation papers using keyword filtering and LLM-based classification, validated against expert labels. Using topic clustering (via two approaches, HDBSCAN+BERTopic and LlooM), we identify between 7 and 15 prominent types of limitations discussed in recent LLM research across the ACL and arXiv datasets. We find that LLM-related research increases nearly sixfold in ACL and nearly fifteenfold in arXiv between 2022 and 2025, while LLLMs research grows even faster, by a factor of over 12 in ACL and nearly 28 in arXiv. Reasoning remains the most studied limitation, followed by generalization, hallucination, bias, and security. The distribution of topics in the ACL dataset stays relatively stable over time, while arXiv shifts toward safety and controllability (with topics like security risks, alignment, hallucinations, knowledge editing), and multimodality between 2022 and 2025. We offer a quantitative view of trends in LLM limitations research and release a dataset of annotated abstracts and a validated methodology, available at: https://github.com/a-kostikova/LLLMs-Survey.

Conventional AI evaluation approaches concentrated within the AI stack exhibit systemic limitations for exploring, navigating and resolving the human and societal factors that play out in real world deployment such as in education, finance, healthcare, and employment sectors. AI capability evaluations can capture detail about first-order effects, such as whether immediate system outputs are accurate, or contain toxic, biased or stereotypical content, but AI's second-order effects, i.e. any long-term outcomes and consequences that may result from AI use in the real world, have become a significant area of interest as the technology becomes embedded in our daily lives. These secondary effects can include shifts in user behavior, societal, cultural and economic ramifications, workforce transformations, and long-term downstream impacts that may result from a broad and growing set of risks. This position paper argues that measuring the indirect and secondary effects of AI will require expansion beyond static, single-turn approaches conducted in silico to include testing paradigms that can capture what actually materializes when people use AI technology in context. Specifically, we describe the need for data and methods that can facilitate contextual awareness and enable downstream interpretation and decision making about AI's secondary effects, and recommend requirements for a new ecosystem.

Traditional single-modal sensing systems-based solely on either radio frequency (RF) or visual data-struggle to cope with the demands of complex and dynamic environments. Furthermore, single-device systems are constrained by limited perspectives and insufficient spatial coverage, which impairs their effectiveness in urban or non-line-of-sight scenarios. To overcome these challenges, we propose a novel large language model (LLM)-driven distributed integrated multimodal sensing and semantic communication (LLM-DiSAC) framework. Specifically, our system consists of multiple collaborative sensing devices equipped with RF and camera modules, working together with an aggregation center to enhance sensing accuracy. First, on sensing devices, LLM-DiSAC develops an RF-vision fusion network (RVFN), which employs specialized feature extractors for RF and visual data, followed by a cross-attention module for effective multimodal integration. Second, a LLM-based semantic transmission network (LSTN) is proposed to enhance communication efficiency, where the LLM-based decoder leverages known channel parameters, such as transceiver distance and signal-to-noise ratio (SNR), to mitigate semantic distortion. Third, at the aggregation center, a transformer-based aggregation model (TRAM) with an adaptive aggregation attention mechanism is developed to fuse distributed features and enhance sensing accuracy. To preserve data privacy, a two-stage distributed learning strategy is introduced, allowing local model training at the device level and centralized aggregation model training using intermediate features. Finally, evaluations on a synthetic multi-view RF-visual dataset generated by the Genesis simulation engine show that LLM-DiSAC achieves a good performance.

Vision-language models (VLMs) have achieved strong results on coding and math benchmarks that are challenging for humans, yet their ability to perform tasks that come naturally to humans--such as perception, spatial navigation, and memory management--remains understudied. Real video games are crafted to be intuitive for humans to learn and master by leveraging innate inductive biases, making them an ideal testbed for evaluating such capabilities in VLMs. To this end, we introduce VideoGameBench, a benchmark consisting of 10 popular video games from the 1990s that VLMs directly interact with in real-time. VideoGameBench challenges models to complete entire games with access to only raw visual inputs and a high-level description of objectives and controls, a significant departure from existing setups that rely on game-specific scaffolding and auxiliary information. We keep three of the games secret to encourage solutions that generalize to unseen environments. Our experiments show that frontier vision-language models struggle to progress beyond the beginning of each game. We find inference latency to be a major limitation of frontier models in the real-time setting; therefore, we introduce VideoGameBench Lite, a setting where the game pauses while waiting for the LM's next action. The best performing model, Gemini 2.5 Pro, completes only 0.48% of VideoGameBench and 1.6% of VideoGameBench Lite. We hope that the formalization of the human skills mentioned above into this benchmark motivates progress in these research directions.

De novo peptide sequencing is a critical task in proteomics. However, the performance of current deep learning-based methods is limited by the inherent complexity of mass spectrometry data and the heterogeneous distribution of noise signals, leading to data-specific biases. We present RankNovo, the first deep reranking framework that enhances de novo peptide sequencing by leveraging the complementary strengths of multiple sequencing models. RankNovo employs a list-wise reranking approach, modeling candidate peptides as multiple sequence alignments and utilizing axial attention to extract informative features across candidates. Additionally, we introduce two new metrics, PMD (Peptide Mass Deviation) and RMD (residual Mass Deviation), which offer delicate supervision by quantifying mass differences between peptides at both the sequence and residue levels. Extensive experiments demonstrate that RankNovo not only surpasses its base models used to generate training candidates for reranking pre-training, but also sets a new state-of-the-art benchmark. Moreover, RankNovo exhibits strong zero-shot generalization to unseen models whose generations were not exposed during training, highlighting its robustness and potential as a universal reranking framework for peptide sequencing. Our work presents a novel reranking strategy that fundamentally challenges existing single-model paradigms and advances the frontier of accurate de novo sequencing. Our source code is provided on GitHub.

Recent studies on the safety alignment of large language models (LLMs) have revealed that existing approaches often operate superficially, leaving models vulnerable to various adversarial attacks. Despite their significance, these studies generally fail to offer actionable solutions beyond data augmentation for achieving more robust safety mechanisms. This paper identifies a fundamental cause of this superficiality: existing alignment approaches often presume that models can implicitly learn a safety-related reasoning task during the alignment process, enabling them to refuse harmful requests. However, the learned safety signals are often diluted by other competing objectives, leading models to struggle with drawing a firm safety-conscious decision boundary when confronted with adversarial attacks. Based on this observation, by explicitly introducing a safety-related binary classification task and integrating its signals with our attention and decoding strategies, we eliminate this ambiguity and allow models to respond more responsibly to malicious queries. We emphasize that, with less than 0.2x overhead cost, our approach enables LLMs to assess the safety of both the query and the previously generated tokens at each necessary generating step. Extensive experiments demonstrate that our method significantly improves the resilience of LLMs against various adversarial attacks, offering a promising pathway toward more robust generative AI systems.

Large Language Models (LLMs) have rapidly become central to NLP, demonstrating their ability to adapt to various tasks through prompting techniques, including sentiment analysis. However, we still have a limited understanding of how these models capture sentiment-related information. This study probes the hidden layers of Llama models to pinpoint where sentiment features are most represented and to assess how this affects sentiment analysis. Using probe classifiers, we analyze sentiment encoding across layers and scales, identifying the layers and pooling methods that best capture sentiment signals. Our results show that sentiment information is most concentrated in mid-layers for binary polarity tasks, with detection accuracy increasing up to 14% over prompting techniques. Additionally, we find that in decoder-only models, the last token is not consistently the most informative for sentiment encoding. Finally, this approach enables sentiment tasks to be performed with memory requirements reduced by an average of 57%. These insights contribute to a broader understanding of sentiment in LLMs, suggesting layer-specific probing as an effective approach for sentiment tasks beyond prompting, with potential to enhance model utility and reduce memory requirements.