Video anomaly understanding (VAU) is critical for real-world scenarios. Recent advances in Video Large Language Models (Video-LLMs) enhance the ability of VAU models to describe and interpret anomalies. However, progress in anomaly localization is still limited by two key issues. First, most existing video anomaly datasets only annotate segments that are clearly inconsistent with the context, often omitting subsequent segments that are semantically part of the same abnormal event. Second, the field lacks systematic evaluation protocols. To bridge these gaps, we introduce VALU, a new benchmark that explicitly defines anomalies across five semantic levels and provides comprehensive temporal boundaries and detailed textual descriptions for each. Based on these annotations, we design three evaluation tasks that comprehensively assess models’ capabilities across different dimensions, including temporal grounding, anomaly localization, and anomaly detail discrimination. Evaluation results reveal persistent challenges in current models’ capabilities on VAU. We further analyze and discuss these findings, and hope that both VALU and insights will advance research in VAU and the development of Video-LLMs. Our benchmark will be publicly available.

The massive size of Large Language Models (LLMs) imposes substantial computational and storage burdens, particularly on devices with limited hardware resources. Compared to foundation models, smaller and more specialized models are often more suitable for practical deployment. Existing customization approaches, such as the conventional “prune-then-finetune” paradigm or task-agnostic deployment strategies, either incur excessive computational costs or lead to suboptimal task performance. The recently popular Mixture-of-Experts (MoE) architecture exhibits a strong ability to mitigate inter-task interference, offering a new perspective on model deployment. In this paper, we introduce ModularMoE, a training framework that converts pre-trained LLMs into parameter-sharing MoE models for lightweight deployment. Exploiting the emergent modularity within LLMs, we split the feed-forward layers into multiple disjoint modules. Each expert is then constructed as a combination of such modules, enabling knowledge sharing across experts and thereby improving parameter efficiency within MoEs. Extensive experiments across multiple downstream tasks demonstrate that ModularMoE outperforms other state-of-the-art baselines at the same sparsity level, achieving an average performance improvement of 4.10% to 28.75% while delivering up to 2.71× inference speedup.

In this paper, we investigate knowledge forgetting in large language models with a focus on its generalisation—ensuring that models forget not only specific training samples but also related implicit knowledge. To this end, we begin by identifying a broader unlearning scope that includes both target data and logically associated samples, including rephrased, subject-replaced, relation-reversed, and one-hop reasoned data. We then conduct a rigorous evaluation of 15 state-of-the-art methods across three datasets, revealing that unlearned models still recall paraphrased answers and retain target facts in their intermediate layers. This motivates us to take a preliminary step toward more generalised implicit knowledge forgetting by proposing PERMU—a novel probability perturbation-based unlearning paradigm. PERMU simulates adversarial unlearning samples to eliminate fact-related tokens from the logit distribution, collectively reducing the probabilities of all answer-associated tokens. Experiments are conducted on a diverse range of datasets, including TOFU, Harry Potter, ZsRE, WMDP, and MUSE, using models ranging from 1.3B to 13B in scale. The results demonstrate that PERMU delivers up to a 50.40% improvement in unlearning vanilla target data while maintaining a 40.73% boost in forgetting implicit knowledge. Our code can be found in the supplementary material.

Recent advances in large language models (LLMs) have significantly improved code-generation capabilities, particularly through retrieval-augmented generation (RAG) for private libraries. While RAG leverages API documentation to address the scarcity of private code corpora, its performance critically depends on the quality of retrieved examples. Existing approaches often overlook the intrinsic characteristics of these examples, particularly how factors such as complexity, readability, and correctness impact their effectiveness. In this study, we systematically investigate these three critical aspects—complexity, readability, and correctness—and find that optimal examples should exhibit moderate complexity, semantic correctness, and step-by-step execution patterns. Based on these findings, we propose ComboPrompt, a novel example enhancement method that strategically combines existing API examples to improve complexity, refines code structure for readability, and incorporates automated validation ensuring correctness. Extensive evaluations across five private library benchmarks and different LLMs demonstrate that ComboPrompt achieves up to 22% accuracy improvement over baseline approaches. Code is available at [Anonymous Github](https://github.com/FireAndWin/ComboPrompt_ExampleQualityMatters).

Prompts, especially high-quality ones, play an invaluable role in assisting large language models (LLMs) to accomplish various natural language processing tasks. However, carefully crafted prompts can also manipulate model behavior. Therefore, the security risks that “prompts themselves face” and those “arising from harmful prompts” cannot be overlooked and we define the Prompt Threat (PT) issues. In this paper, we review the latest attack methods related to prompt threats, focusing on prompt leakage attacks and prompt jailbreak attacks. Additionally, we summarize the experimental setups of these methods and explore the relationship between prompt threats and prompt injection attacks.

Large Vision-Language Models (LVLMs) have a significant issue with object hallucinations, where researchers have noted that LVLMs often mistakenly determine objects as present in images where they do not actually exist. Some recent studies evaluate the occurrence of object hallucinations by asking LVLMs whether they see objects that do not exist in input images. However, we observe that these evaluation methods have some limitations, such as the objects being questioned potentially having little relevance to the image. In this paper, we introduce a more challenging benchmark for evaluating object hallucinations by removing objects from images and then asking the model whether it can still see the removed objects. Our evaluation result reveals that LVLMs suffer from severe hallucinations, as they often still claim to see the removed objects. Through our analysis, we find that biases in training result in LVLMs lacking guidance on learning about the absence of objects, which in turn leads to a lack of ability to determine that objects do not exist in images. To address this issue, we further propose oDPO, a direct preference optimization objective based on visual objects. By guiding LVLMs to learn to determine the existence of objects, oDPO effectively alleviates object hallucinations. It achieves more competitive results than other hallucination mitigation approaches across multiple object hallucination benchmarks and enhances the performance of LVLMs in various vision-language tasks.

Sparse attention can effectively alleviate the significant demands on memory when large language models (LLMs) process long contexts. Existing methods typically apply the same sparse pattern across different attention heads and inputs. However, this uniform approach fails to capture the inherent diversity of attention patterns within LLMs — the intrinsic attention clustering. To address this, we propose ClusterAttn, a training-free sparse attention method that provides an efficient prompt cache compression scheme under intrinsic attention clustering for efficient LLM inference.Our findings show that attention heads consistently focus on specific clusters of the prompt during decoding, a pattern detectable from an observation window at the prompt’s end. ClusterAttn adaptively fits these clusters utilizing a density-based attention clustering algorithm, thus compressing the KV cache of the prompt. Evaluations on different models across various benchmarks demonstrate ClusterAttn’s superior compression rates and efficiency. By utilizing only 1024 tokens, it can reduce memory usage by 10%–65%, resulting in a latency reduction of 12%–23% and a throughput increase of 2.6–4.8 times, all with nearly no accuracy loss. Additionally, ClusterAttn can handle up to 128k context on a single A100-80GB GPU, outperforming existing methods.

Multimodal large language models (MLLMs) have recently achieved state-of-the-art performance on tasks ranging from visual question answering to video understanding. However, existing studies have concentrated mainly on visual–textual misalignment, leaving largely unexplored the MLLMs’ ability to preserve an originally correct answer when confronted with misleading information. We reveal a response uncertainty phenomenon: across nine standard datasets, twelve state-of-the-art open-source MLLMs overturn a previously correct answer in 65% of cases after receiving a single deceptive cue. To systematically quantify this vulnerability, we propose a two-stage evaluation pipeline: (1) elicit each model’s original response on unperturbed inputs; (2) inject explicit (false-answer hints) and implicit (contextual contradictions) misleading instructions, and compute the misleading rate—the fraction of correct-to-incorrect flips. Leveraging the most susceptible examples, we curate the Multimodal Uncertainty Benchmark (MUB), a collection of image–question pairs stratified into low, medium, and high difficulty based on how many of twelve state-of-the-art MLLMs they mislead. Extensive evaluation on twelve open-source and five closed-source models reveals a high uncertainty: average misleading rates exceed 86%, with explicit cues over 67.19% and implicit cues over 80.67%. To reduce the misleading rate, we then fine-tune all open-source MLLMs on a compact 2,000-sample mixed-instruction dataset, reducing misleading rates to 6.97% (explicit) and 32.77% (implicit), boosting consistency by nearly 29.37% on highly deceptive inputs, and slightly improving accuracy on standard benchmarks.

Recommendation Models (RMs) are crucial for predicting user preferences and enhancing personalized experiences on large-scale platforms. As the application of recommendation models grows, optimizing their online serving performance has become a significant challenge. However, current serving systems perform poorly under highly concurrent scenarios. To address this, we introduce RecStream, a system designed to optimize stream configurations based on model characteristics for handling high concurrency requests. We employ a hybrid Graph Neural Network architecture to determine the best configurations for various RMs. Experimental results demonstrate that RecStream achieves significant performance improvements, reducing latency by up to 74%.

Entity alignment (EA) is crucial for integrating multi-source knowledge graphs (KGs), aiming to identify equivalent entities across different graphs. However, most existing EA decoding methods rely on both entity and relation embeddings, limiting their generalizability and efficiency, especially in GNN-based models. To address these challenges, we propose Triple Feature Propagation (TFP), an adaptable and fast EA decoding framework that only utilizes entity embeddings. TFP reconstructs KG representation by maximizing the smoothness of entity embeddings. The discretized smoothness-maximization process yields the explicit Euler solution of TFP. We also generalize multi-view matrices: entity-to-entity, entity-to-relation, relation-to-entity, and relation-to-triple, to capture structural diversity. Extensive experiments on public datasets demonstrate that TFP is fast and adaptable to various encoders, achieving comparable results to state-of-the-art methods in under 6 seconds, and surpassing them in many cases.

Existing research in multi-hop questions has identified two reasoning modes: latent reasoning and factual shortcuts, but has not deeply investigated how these modes differ during inference. This impacts both model generalization ability and downstream reasoning tasks. In this work, we systematically examine these distinctions and propose a simple and efficient classification metric, Attribute Rate Ratio (ARR). First, we construct specialized datasets corresponding to the two reasoning modes based on our proposed criteria. Then, using reverse engineering methods, including attention knockout and logit lens techniques, we reveal that subject representations differ significantly across modes: latent reasoning encodes bridge-related information for final answer extraction, while factual shortcuts bypass intermediate reasoning and resemble single-hop factual queries. Finally, our proposed ARR achieves around 90% accuracy on our datasets and demonstrates effectiveness in RAG conflict scenarios, showing that model behavior under conflicting prompts is closely tied to its underlying reasoning mode. Our findings and proposed metric have significant potential for advancing LLM development and applications.

Micro-enterprises and individual developers emerge analysis demands for long sequence with powerful Large Language Models (LLMs). They try to deploy the LLMs at local, but only possess various commodity devices and the unreliable interconnection between devices. Existing parallel techniques do not lead to the same effectiveness in limited environment. The heterogeneity of devices, coupled with their limited capacity and expensive communication, brings challenges to private deployment for maximized utilization of available devices while masking latency. Hence, we introduce HPipe, a pipeline inference framework that successfully mitigates LLMs from high-performance clusters to heterogeneous commodity devices. By ensuring a balanced distribution of workloads, HPipe facilitates the parallel execution of LLMs through pipelining the sequences on the token dimension. The evaluation conducted on LLaMA-7B and GPT3-2B demonstrates that HPipe holds the potential for context analysis on LLM with heterogeneity devices, achieving an impressive speedup in latency and throughput up to 2.28 times.

Code summarization provides a natural language description for a given piece of code. In this work, we focus on scripting code—programming languages that interact with specific devices through commands. The low-resource nature of scripting languages makes traditional code summarization methods challenging to apply. To address this, we introduce a novel framework: distantly supervised contrastive learning for low-resource scripting language summarization. This framework leverages limited atomic commands and category constraints to enhance code representations. Extensive experiments demonstrate our method’s superiority over competitive baselines.

Recently, retrieval-based in-context learning (ICL) methods for selecting demonstrations have been widely investigated. Existing methods train a dense retriever to retrieve the most appropriate demonstrations for a given test query, which improves ICL performance. However, we find that distinct LLMs exhibit different biases for “what is a good demonstration” since they possess differences in training data, model architectures and training methods. As a result, a demonstration suitable for one LLM may not be appropriate for others.Previous approaches ignore the model bias and fail to retrieve the most appropriate demonstrations for different inference LLMs, resulting in a degradation of ICL performance.To address this problem, we propose a simple yet effective metric to evaluate the appropriateness of demonstrations for a specific inference LLM. Furthermore, we introduce a Model-specific Demonstration Retrieval (MDR) method for ICL at inference time, which considers the biases of different LLMs. We test MDR on seen and unseen tasks with multi-scale inference LLMs, such as GPT-Neo-2.7B, LLaMA-7B and Vicuna-13B. Experiments on 23 datasets across 11 data domains highlight the remarkable effectiveness of MDR, showcasing improvements of up to 41.2% in comparison to methods that neglect model biases.

Language Models (LMs) acquire factual knowledge during pre-training and store it in the parameters, which can be valuable for downstream tasks. As world evolves, some facts may be incorrectly induced or become obsolete over time. Various model editing methods have been proposed to modify specific examples in LMs. However, existing training-based methods still suffer from sub-optimal locality, where irrelevant neighborhood examples can be adversely influenced. Model’s gradients are still struggling to identify the appropriate direction when updating the parameters. To address this issue, we find that directing the hidden state of the edit example towards spaces where semantics are sparse tends to help preserve the semantics of irrelevant neighborhood examples. Based on this hypothesis, we propose a novel metric, named SSS, to evaluate the degree of sparsity around a sentence embedding in the semantic space without any human or machine annotation. Subsequently, we incorporate SSS into the original loss function of the existing training-based methods to enhance locality. Experiments conducted on two datasets across various models demonstrate that SSS is effective in improving both locality and reasoning capability.

Aspect-based sentiment analysis (ABSA) aims to distinguish sentiment polarity of every specific aspect in a given sentence. Previous researches have realized the importance of interactive learning with context and aspects. However, these methods are ill-studied to learn complex sentence with multiple aspects due to overlapped polarity feature. And they do not consider the correlation between aspects to distinguish overlapped feature. In order to solve this problem, we propose a new method called Recurrent Inverse Learning Guided Network (RILGNet). Our RILGNet has two points to improve the modeling of aspect correlation and the selecting of aspect feature. First, we use Recurrent Mechanism to improve the joint representation of aspects, which enhances the aspect correlation modeling iteratively. Second, we propose Inverse Learning Guidance to improve the selection of aspect feature by considering aspect correlation, which provides more useful information to determine polarity. Experimental results on SemEval 2014 Datasets demonstrate the effectiveness of RILGNet, and we further prove that RILGNet is state-of-the-art method in multiaspect scenarios.

Cross-domain sentiment classification aims to address the lack of massive amounts of labeled data. It demands to predict sentiment polarity on a target domain utilizing a classifier learned from a source domain. In this paper, we investigate how to efficiently apply the pre-training language model BERT on the unsupervised domain adaptation. Due to the pre-training task and corpus, BERT is task-agnostic, which lacks domain awareness and can not distinguish the characteristic of source and target domain when transferring knowledge. To tackle these problems, we design a post-training procedure, which contains the target domain masked language model task and a novel domain-distinguish pre-training task. The post-training procedure will encourage BERT to be domain-aware and distill the domain-specific features in a self-supervised way. Based on this, we could then conduct the adversarial training to derive the enhanced domain-invariant features. Extensive experiments on Amazon dataset show that our model outperforms state-of-the-art methods by a large margin. The ablation study demonstrates that the remarkable improvement is not only from BERT but also from our method.

Aspect-level sentiment classification is a crucial task for sentiment analysis, which aims to identify the sentiment polarities of specific targets in their context. The main challenge comes from multi-aspect sentences, which express multiple sentiment polarities towards different targets, resulting in overlapped feature representation. However, most existing neural models tend to utilize static pooling operation or attention mechanism to identify sentimental words, which therefore insufficient for dealing with overlapped features. To solve this problem, we propose to utilize capsule network to construct vector-based feature representation and cluster features by an EM routing algorithm. Furthermore, interactive attention mechanism is introduced in the capsule routing procedure to model the semantic relationship between aspect terms and context. The iterative routing also enables encoding sentence from a global perspective. Experimental results on three datasets show that our proposed model achieves state-of-the-art performance.

As an essential component of natural language processing, text classification relies on deep learning in recent years. Various neural networks are designed for text classification on the basis of word embedding. However, polysemy is a fundamental feature of the natural language, which brings challenges to text classification. One polysemic word contains more than one sense, while the word embedding procedure conflates different senses of a polysemic word into a single vector. Extracting the distinct representation for the specific sense could thus lead to fine-grained models with strong generalization ability. It has been demonstrated that multiple senses of a word actually reside in linear superposition within the word embedding so that specific senses can be extracted from the original word embedding. Therefore, we propose to use capsule networks to construct the vectorized representation of semantics and utilize hyperplanes to decompose each capsule to acquire the specific senses. A novel dynamic routing mechanism named ‘routing-on-hyperplane’ will select the proper sense for the downstream classification task. Our model is evaluated on 6 different datasets, and the experimental results show that our model is capable of extracting more discriminative semantic features and yields a significant performance gain compared to other baseline methods.