Current Video Large Language Models (VideoLLMs) suffer from quadratic computational complexity and key-value cache scaling, due to their reliance on processing excessive redundant visual tokens. To address this problem, we propose SharpV, a minimalist and efficient method for adaptive pruning of visual tokens and KV cache. Different from most uniform compression approaches, SharpV dynamically adjusts pruning ratios based on spatial-temporal information. Remarkably, this adaptive mechanism occasionally achieves performance gains over dense models, offering a novel paradigm for adaptive pruning. During the KV cache pruning stage, based on observations of visual information degradation, SharpV prunes degraded visual features via a self-calibration manner, guided by similarity to original visual features. In this way, SharpV achieves hierarchical cache pruning from the perspective of information bottleneck, offering a new insight into VideoLLMs' information flow. Experiments on multiple public benchmarks demonstrate the superiority of SharpV. Moreover, to the best of our knowledge, SharpV is notably the first two-stage pruning framework that operates without requiring access to exposed attention scores, ensuring full compatibility with hardware acceleration techniques like Flash Attention.

Accurately recognizing distracted driving activities in real-world scenarios is essential for improving road and pedestrian safety. However, existing approaches are prone to attending to irrelevant scene context and are susceptible to interference from redundant frames, compromising their robustness in complex driving environments. To overcome these limitations, we propose DualScope, a novel framework that captures behaviorally critical information from both spatial and temporal perspectives. In the spatial domain, we introduce a Synergistic Behavior-Centric Distillation mechanism that leverages two key information sources: (1) position-aware knowledge derived from the SAM model, which enhances the perception of critical regions and their semantic interaction structures; and (2) fine-grained visual details obtained from cropped key regions, which improve the model's ability to capture detailed patterns within behavior-relevant areas. In the temporal domain, we present the Saliency-Aware Fine-to-Coarse Temporal Modeling module, comprising three components: a Fine-Grained Motion Encoder for capturing local inter-frame dependencies; a Dynamic Difference Extractor for generating salient motion dynamics; and a Saliency-Aware Temporal Pyramid Mamba for integrating these representations to enable multi-scale temporal modeling. This design effectively captures both short-term motions and long-term behavioral patterns. Furthermore, incorporating salient dynamics enhances the model's focus on significant behavioral variations. Extensive experiments on seven publicly available DDAR datasets demonstrate that DualScope consistently outperforms state-of-the-art methods, validating its effectiveness in capturing behavioral cues across spatial and temporal dimensions.

Mixture-of-Experts (MoE) architectures decouple model capacity from per-token computation, enabling scaling beyond the computational limits imposed by dense scaling laws. Yet how MoE architectures shape knowledge acquisition during pre-training—and how this process differs from dense architectures—remains unknown. To address this issue, we introduce Gated-LPI (Log-Probability Increase), a neuron-level attribution metric that decomposes log-probability increase across neurons. We present a time-resolved comparison of knowledge acquisition dynamics in MoE and dense architectures, tracking checkpoints over 1.2M (~ 5.0T tokens) and 600K (~ 2.5T tokens) training steps, respectively. Our experiments uncover three patterns: (1) Low-entropy backbone. The top approximately 1% of MoE neurons capture over 45% of positive updates, forming a high-utility core, which is absent in the dense baseline. (2) Early consolidation. The MoE model locks into a stable importance profile within < 100K steps, whereas the dense model remains volatile throughout training. (3) Functional robustness. Masking the ten most important MoE attention heads reduces relational HIT@10 by < 10%, compared with > 50% for the dense model, showing that sparsity fosters distributed—rather than brittle—knowledge storage. These patterns collectively demonstrate that sparsity fosters an intrinsically stable and distributed computational backbone from early in training, helping bridge the gap between sparse architectures and training-time interpretability.

Chenfei Liao, Wensong Wang, Zichen Wen, Xu Zheng, Yiyu Wang, Haocong He, Yuanhuiyi Lyu, Lutao Jiang, Xin Zou, Yuqian Fu, Bin Ren, Linfeng Zhang, Xuming Hu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Shuliang Liu, Songbo Yang, Dong Fang, Sihang Jia, Yuqi Tang, Lingfeng Su, Ruoshui Peng, Yibo Yan, Xin Zou, Xuming Hu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Nowadays, regulatory compliance has become a cornerstone of corporate governance, ensuring adherence to systematic legal frameworks. At its core, financial regulations often comprise highly intricate provisions, layered logical structures, and numerous exceptions, which inevitably result in labor-intensive or comprehension challenges. To mitigate this, recent Regulatory Technology (RegTech) and Large Language Models (LLMs) have gained significant attention in automating the conversion of regulatory text into executable compliance logic. However, their performance remains suboptimal particularly when applied to Chinese-language financial regulations, due to three key limitations: (1) incomplete domain-specific knowledge representation, (2) insufficient hierarchical reasoning capabilities, and (3) failure to maintain temporal and logical coherence. One promising solution is to develop a domain specific and code-oriented dataset for model training. Existing datasets such as LexGLUE, LegalBench, and CODE-ACCORD are often English-focused, domain-mismatched, or lack fine-grained granularity for compliance code generation. To fill these gaps, we present Compliance-to-Code, the first large-scale Chinese dataset dedicated to financial regulatory compliance. Covering 1,159 annotated clauses from 361 regulations across ten categories, each clause is modularly structured with four logical elements-subject, condition, constraint, and contextual information-along with regulation relations. We provide deterministic Python code mappings, detailed code reasoning, and code explanations to facilitate automated auditing. To demonstrate utility, we present FinCheck: a pipeline for regulation structuring, code generation, and report generation. Compliance-to-Code establishes a new benchmark for LLM-based compliance automation. Experimental evaluation shows that GLM-4-9B-0414 achieves the best performance on key task in regulation structuring and DeepSeek-R1-0528 performs the best in compliance code generation tasks.

Automated waterway environment perception is crucial for enabling unmanned surface vessels (USVs) to understand their surroundings and make informed decisions. Most existing waterway perception models primarily focus on instance-level object perception paradigms (e.g., detection, segmentation). However, due to the complexity of waterway environments, current perception datasets and models fail to achieve global semantic understanding of waterways, limiting large-scale monitoring and structured log generation. With the advancement of vision-language models (VLMs), we leverage image captioning to introduce WaterCaption, the first captioning dataset specifically designed for waterway environments. WaterCaption focuses on fine-grained, multi-region long-text descriptions, providing a new research direction for visual geo-understanding and spatial scene cognition. Exactly, it includes 20.2k image-text pair data with 1.8 million vocabulary size. Additionally, we propose Da Yu, an edge-deployable multi-modal large language model for USVs, where we propose a novel vision-to-language projector called Nano Transformer Adaptor (NTA). NTA effectively balances computational efficiency with the capacity for both global and fine-grained local modeling of visual features, thereby significantly enhancing the model’s ability to generate long-form textual outputs. Da Yu achieves an optimal balance between performance and efficiency, surpassing state-of-the-art models on WaterCaption and several other captioning benchmarks. The project is available at https://github.com/GuanRunwei/WaterCaption.

Diffusion Transformers (DiT) have become the dominant methods in image and video generation yet still suffer substantial computational costs. As an effective approach for DiT acceleration, feature caching methods are designed to cache the features of DiT in previous timesteps and reuse them in the next timesteps, allowing us to skip the computation in the next timesteps. Among them, token-wise feature caching has been introduced to perform different caching ratios for different tokens in DiTs, aiming to skip the computation for unimportant tokens while still computing the important ones. In this paper, we propose to carefully check the effectiveness in token-wise feature caching with the following two questions: 1) Is it really necessary to compute the so-called "important" tokens in each step? 2) Are so-called important tokens really important? Surprisingly, this paper gives some counter-intuition answers, demonstrating that consistently computing the selected "important tokens" in all steps is not necessary. The selection of the so-called "important tokens" is often ineffective, and even sometimes shows inferior performance than random selection. Based on these observations, this paper introduces dual feature caching referred to as DuCa, which performs aggressive caching strategy and conservative caching strategy iteratively and selects the tokens for computing randomly. Extensive experimental results demonstrate the effectiveness of our method in DiT, PixArt, FLUX, and OpenSora, demonstrating significant improvements than the previous token-wise feature caching.

Fine-tuning large language models (LLMs) on task-specific data is essential for their effective deployment. As dataset sizes grow, efficiently selecting optimal subsets for training becomes crucial to balancing performance and computational costs. Traditional data selection methods often require fine-tuning a scoring model on the target dataset, which is time-consuming and resource-intensive, or rely on heuristics that fail to fully leverage the model’s predictive capabilities. To address these challenges, we propose Data Whisperer, an efficient, training-free, attention-based method that leverages few-shot in-context learning with the model to be fine-tuned. Comprehensive evaluations were conducted on both raw and synthetic datasets across diverse tasks and models. Notably, Data Whisperer achieves superior performance compared to the full GSM8K dataset on the Llama-3-8B-Instruct model, using just 10% of the data, and outperforms existing methods with a 3.1-point improvement and a 7.4× speedup.

The interpretability of Mixture-of-Experts (MoE) models, especially those with heterogeneous designs, remains underexplored.Existing attribution methods for dense models fail to capture dynamic routing-expert interactions in sparse MoE architectures.To address this issue, we propose a cross-level attribution algorithm to analyze sparse MoE architectures (Qwen 1.5-MoE, OLMoE, Mixtral-8x7B) against dense models (Qwen 1.5-7B, Llama-7B, Mistral-7B).Results show MoE models achieve 31% higher per-layer efficiency via a "mid-activation, lateamplification" pattern: early layers screen experts, while late layers refine knowledge collaboratively.Ablation studies reveal a "basicrefinement" framework-shared experts handle general tasks (entity recognition), while routed experts specialize in domain-specific processing (geographic attributes).Semanticdriven routing is evidenced by strong correlations between attention heads and experts (r = 0.68), enabling task-aware coordination.Notably, architectural depth dictates robustness: deep Qwen 1.5-MoE mitigates expert failures (e.g., 43% MRR drop in geographic tasks when blocking top-10 experts) through shared expert redundancy, whereas shallow OLMoE suffers severe degradation (76% drop).Task sensitivity further guides design: core-sensitive tasks (geography) require concentrated expertise, while distributed-tolerant tasks (object attributes) leverage broader participation.These insights advance MoE interpretability, offering principles to balance efficiency, specialization, and robustness.

Leyi Pan, Aiwei Liu, Shiyu Huang, Yijian Lu, Xuming Hu, Lijie Wen, Irwin King, Philip S. Yu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Boyang Xue, Fei Mi, Qi Zhu, Hongru Wang, Rui Wang, Sheng Wang, Erxin Yu, Xuming Hu, Kam-Fai Wong. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Despite the recent breakthroughs achieved by Large Vision Language Models (LVLMs) in understanding and responding to complex visual-textual contexts, their inherent hallucination tendencies limit their practical application in real-world scenarios that demand high levels of precision. Existing methods typically either fine-tune the LVLMs using additional data, which incurs extra costs in manual annotation and computational resources or perform comparisons at the decoding stage, which may eliminate useful language priors for reasoning while introducing inference time overhead. Therefore, we propose ICT, a lightweight, training-free method that calculates an intervention direction to shift the model’s focus towards different levels of visual information, enhancing its attention to high-level and fine-grained visual details. During the forward pass stage, the intervention is applied to the attention heads that encode the overall image information and the fine-grained object details, effectively mitigating the phenomenon of overly language priors, and thereby alleviating hallucinations. Extensive experiments demonstrate that ICT achieves strong performance with a small amount of data and generalizes well across different datasets and models. Our codes are publicly available at:https://github.com/THU-BPM/ICT/.

Dataset distillation has emerged as a powerful approach for reducing data requirements in deep learning. Among various methods, distribution matching-based approaches stand out for their balance of computational efficiency and strong performance. However, existing distance metrics used in distribution matching often fail to accurately capture distributional differences, leading to unreliable measures of discrepancy. In this paper, we reformulate dataset distillation as a minmax optimization problem and introduce Neural Characteristic Function Discrepancy (NCFD), a comprehensive and theoretically grounded metric for measuring distributional differences. NCFD leverages the Characteristic Function (CF) to encapsulate full distributional information, employing a neural network to optimize the sampling strategy for the CF’s frequency arguments, thereby maximizing the discrepancy to enhance distance estimation. Simultaneously, we minimize the difference between real and synthetic data under this optimized NCFD measure. Our approach, termed Neural Characteristic Function Matching (NCFM), inherently aligns the phase and amplitude of neural features in the complex plane for both real and synthetic data, achieving a balance between realism and diversity in synthetic samples. Experiments demonstrate that our method achieves significant performance gains over state-of-the-art methods on both low- and high-resolution datasets. Notably, we achieve a 22.9% accuracy boost on ImageSquawk. Our method also reduces GPU memory usage by over 300× and achieves 20× faster processing speeds compared to state-of-the-art methods. To the best of our knowledge, this is the first work to achieve lossless compression of CIFAR-100 on a single NVIDIA 2080 Ti GPU using only 2.3 GB of memory. The code for this work is publicly available at: https://github.com/gszfwsb/NCFM.

Yunkai Dang, Mengxi Gao, Yibo Yan, Xin Zou, Yanggan Gu, Jungang Li, Jingyu Wang, Peijie Jiang, Aiwei Liu, Jia Liu, Xuming Hu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Large Language Models (LLMs), despite their remarkable capabilities, are hampered by hallucinations.A particularly challenging variant, knowledge overshadowing, occurs when one piece of activated knowledge inadvertently masks another relevant piece, leading to erroneous outputs even with high-quality training data.Current understanding of overshadowing is largely confined to inference-time observations, lacking deep insights into its origins and internal mechanisms during model training.Therefore, we introduce PHANTOMCIRCUIT, a novel framework designed to comprehensively analyze and detect knowledge overshadowing.By innovatively employing knowledge circuit analysis, PHANTOMCIRCUIT dissects the function of key components in the circuit and how the attention pattern dynamics contribute to the overshadowing phenomenon and its evolution throughout the training process.Extensive experiments demonstrate PHANTOMCIRCUIT 's effectiveness in identifying such instances, offering novel insights into this elusive hallucination and providing the research community with a new methodological lens for its potential mitigation.Our code can be found in https://github.com/halfmorepiece/PhantomCircuit.

Mixture-of-Experts (MoE) models offer immense capacity via sparsely gated expert subnetworks, yet adapting them to multiple domains without catastrophic forgetting remains an open challenge.Existing approaches either incur prohibitive computation, suffer crossdomain interference, or require separate runs per domain.We propose DES-MoE, a dynamic expert specialization framework for multi-domain adaptation of Mixture-of-Experts models.DES-MoE addresses catastrophic forgetting through three innovations: (1) an adaptive router balancing pre-trained knowledge retention and task-specific updates via distillation, (2) real-time expert-domain correlation mapping to isolate domain-specific gradients, and (3) a three-phase adaptive fine-tuning schedule that progressively freezes non-specialized parameters.Evaluated on six domains (math, code, law, etc.), DES-MoE matches singledomain ESFT performance while training one unified model, reduces forgetting by 89% compared to full fine-tuning as domains scale from 2 to 6, and achieves 68% faster convergence than conventional methods.Our work establishes dynamic expert isolation as a scalable paradigm for multi-task MoE adaptation.

Shuliang Liu, Zheng Qi, Jesse Jiaxi Xu, Yibo Yan, Junyan Zhang, He Geng, Aiwei Liu, Peijie Jiang, Jia Liu, Yik-Cheung Tam, Xuming Hu. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

We show that large language models (LLMs) exhibit an internal chain-of-thought: they sequentially decompose and execute composite tasks layer-by-layer.Two claims ground our study: (i) distinct subtasks are learned at different network depths, and (ii) these subtasks are executed sequentially across layers.On a benchmark of 15 two-step composite tasks, we employ layer-from context-masking and propose a novel cross-task patching method, confirming (i).To examine claim (ii), we apply LogitLens to decode hidden states, revealing a consistent layerwise execution pattern.We further replicate our analysis on the real-world TRACE benchmark, observing the same stepwise dynamics.Together, our results enhance LLMs transparency by showing their capacity to internally plan and execute subtasks (or instructions), opening avenues for fine-grained, instruction-level activation steering.

With the widespread use of mobile devices and the rapid growth of micro-video platforms such as TikTok and Kwai, the demand for personalized micro-video recommendation systems has significantly increased. Micro-videos typically contain diverse information, such as textual metadata, visual cues (e.g., cover images), and dynamic video content, significantly affecting user interaction and engagement patterns. However, most existing approaches often suffer from the problem of over-smoothing, which limits their ability to capture comprehensive interaction information effectively. Additionally, cold-start scenarios present ongoing challenges due to sparse interaction data and the underutilization of available interaction signals.To address these issues, we propose a Multi-view Hypergraph-based Contrastive learning model for cold-start micro-video Recommendation (MHCR). MHCR introduces a multi-view multi-modal feature extraction layer to capture interaction signals from various perspectives and incorporates multi-view self-supervised learning tasks to provide additional supervisory signals. Through extensive experiments on two real-world datasets, we show that MHCR significantly outperforms existing video recommendation models and effectively mitigates cold-start challenges. Our code is available at https://github.com/sisuolv/MHCR.

Fusing and balancing multi-modal inputs from novel sensors for dense prediction tasks, particularly semantic segmentation, is critically important yet remains a significant challenge. One major limitation is the tendency of multi-modal frameworks to over-rely on easily learnable modalities, a phenomenon referred to as unimodal dominance or bias. This issue becomes especially problematic in real-world scenarios where the dominant modality may be unavailable, resulting in severe performance degradation. To this end, we apply a simple but effective plug-and-play regularization term based on functional entropy, which introduces no additional parameters or modules. This term is designed to intuitively balance the contribution of each visual modality to the segmentation results. Specifically, we leverage the log-Sobolev inequality to bound functional entropy using functional-Fisher-information. By maximizing the information contributed by each visual modality, our approach mitigates unimodal dominance and establishes a more balanced and robust segmentation framework. A multi-scale regularization module is proposed to apply our proposed plug-and-play term on high-level features and also segmentation predictions for more balanced multi-modal learning. Extensive experiments on three datasets demonstrate that our proposed method achieves superior performance, i.e., +13.94%, +3.25%, and +3.64%, without introducing any additional parameters.

Entity Set Expansion (ESE) aims to identify new entities belonging to the same semantic class as the given set of seed entities. Traditional methods solely relied on positive seed entities to represent the target fine-grained semantic class, rendering them tough to represent ultra-fine-grained semantic classes. Specifically, merely relying on positive seed entities leads to two inherent shortcomings: (i) Ambiguity among ultra-fine-grained semantic classes. (ii) Inability to define “unwanted” semantics. Hence, previous ESE methods struggle to address the ultra-fine-grained ESE (Ultra-ESE) task. To solve this issue, we first introduce negative seed entities in the inputs, which jointly describe the ultra-fine-grained semantic class with positive seed entities. Negative seed entities eliminate the semantic ambiguity by providing a contrast between positive and negative attributes. Meanwhile, it provides a straightforward way to express “unwanted”. To assess model performance in Ultra-ESE and facilitate further research, we also constructed UltraWiki, the first large-scale dataset tailored for Ultra-ESE. UltraWiki encompasses 50,973 entities and 394,097 sentences, alongside 236 ultra-fine-grained semantic classes, where each class is represented with 3–5 positive and negative seed entities. Moreover, a retrieval-based framework RetExpan and a generation-based framework GenExpan are proposed to provide powerful baselines for Ultra-ESE. Additionally, we devised two strategies to enhance models' comprehension of ultra-fine-grained entities' semantics: contrastive learning and chain-of-thought reasoning. Extensive experiments confirm the effectiveness of our proposed strategies and also reveal that there remains a large space for improvement in Ultra-ESE. All the codes, dataset, and supplementary notes are available at https://github.com/THUKElab/UltraWiki.

The debate between self-interpretable models and post-hoc explanations for black-box models is central to Explainable AI (XAI). Self-interpretable models, such as concept-based networks, offer insights by connecting decisions to human-understandable concepts but often struggle with performance and scalability. Conversely, post-hoc methods like Shapley values, while theoretically robust, are computationally expensive and resource-intensive. To bridge the gap between these two lines of research, we propose a novel method that combines their strengths, providing theoretically guaranteed self-interpretability for black-box models without compromising prediction accuracy. Specifically, we introduce a parameter-efficient pipeline, AutoGnothi, which integrates a small side network into the black-box model, allowing it to generate Shapley value explanations without changing the original network parameters. This side-tuning approach significantly reduces memory, training, and inference costs, outperforming traditional parameter-efficient methods, where full fine-tuning serves as the optimal baseline. AutoGnothi enables the black-box model to predict and explain its predictions with minimal overhead. Extensive experiments show that AutoGnothi offers accurate explanations for both vision and language tasks, delivering superior computational efficiency with comparable interpretability.

Text watermarking for Large Language Models (LLMs) has made significant progress in detecting LLM outputs and preventing misuse. Current watermarking techniques offer high detectability, minimal impact on text quality, and robustness to text editing. However, current researches lack investigation into the imperceptibility of watermarking techniques in LLM services. This is crucial as LLM providers may not want to disclose the presence of watermarks in real-world scenarios, as it could reduce user willingness to use the service and make watermarks more vulnerable to attacks. This work is the first to investigate the imperceptibility of watermarked LLMs. We design an identification algorithm called Water-Probe that detects watermarks through well-designed prompts to the LLM. Our key motivation is that current watermarked LLMs expose consistent biases under the same watermark key, resulting in similar differences across prompts under different watermark keys. Experiments show that almost all mainstream watermarking algorithms are easily identified with our well-designed prompts, while Water-Probe demonstrates a minimal false positive rate for non-watermarked LLMs. Finally, we propose that the key to enhancing the imperceptibility of watermarked LLMs is to increase the randomness of watermark key selection. Based on this, we introduce the Water-Bag strategy, which significantly improves watermark imperceptibility by merging multiple watermark keys.

Multimodal Large Language Models (MLLMs) have emerged as a central focus in both industry and academia, but often suffer from biases introduced by visual and language priors, which can lead to multimodal hallucination. These biases arise from the visual encoder and the Large Language Model (LLM) backbone, affecting the attention mechanism responsible for aligning multimodal inputs. Existing decoding-based mitigation methods focus on statistical correlations and overlook the causal relationships between attention mechanisms and model output, limiting their effectiveness in addressing these biases. To tackle this issue, we propose a causal inference framework termed CausalMM that applies structural causal modeling to MLLMs, treating modality priors as a confounder between attention mechanisms and output. Specifically, by employing backdoor adjustment and counterfactual reasoning at both the visual and language attention levels, our method mitigates the negative effects of modality priors and enhances the alignment of MLLM's inputs and outputs, with a maximum score improvement of 65.3% on 6 VLind-Bench indicators and 164 points on MME Benchmark compared to conventional methods. Extensive experiments validate the effectiveness of our approach while being a plug-and-play solution. Our code is available at: https://github.com/The-Martyr/CausalMM.

Despite their impressive capabilities, Multimodal Large Language Models (MLLMs) are prone to hallucinations, i.e., the generated content that is nonsensical or unfaithful to input sources. Unlike in LLMs, hallucinations in MLLMs often stem from the sensitivity of text decoder to visual tokens, leading to a phenomenon akin to "amnesia" about visual information. To address this issue, we propose MemVR, a novel decoding paradigm inspired by common cognition: when the memory of an image seen the moment before is forgotten, people will look at it again for factual answers. Following this principle, we treat visual tokens as supplementary evidence, re-injecting them into the MLLM through Feed Forward Network (FFN) as “key-value memory” at the middle trigger layer. This look-twice mechanism occurs when the model exhibits high uncertainty during inference, effectively enhancing factual alignment. Comprehensive experimental evaluations demonstrate that MemVR significantly mitigates hallucination across various MLLMs and excels in general benchmarks without incurring additional time overhead.

Accurate weather forecasts are important for disaster prevention, agricultural planning, etc. Traditional numerical weather prediction (NWP) methods offer physically interpretable high-accuracy predictions but are computationally expensive and fail to fully leverage rapidly growing historical data. In recent years, deep learning models have made significant progress in weather forecasting, but challenges remain, such as balancing global and regional high-resolution forecasts, excessive smoothing in extreme event predictions, and insufficient dynamic system modeling. To address these issues, this paper proposes a global-regional nested weather forecasting framework (OneForecast) based on graph neural networks. By combining a dynamic system perspective with multi-grid theory, we construct a multi-scale graph structure and densify the target region to capture local high-frequency features. We introduce an adaptive messaging mechanism, using dynamic gating units to deeply integrate node and edge features for more accurate extreme event forecasting. For high-resolution regional forecasts, we propose a neural nested grid method to mitigate boundary information loss. Experimental results show that OneForecast performs excellently across global to regional scales and short-term to long-term forecasts, especially in extreme event predictions. Codes link: \url{https://github.com/YuanGao-YG/OneForecast}.

Recent text-to-image generative models, e.g., Stable Diffusion V3 and Flux, have achieved notable progress. However, these models are strongly restricted to their limited knowledge, a.k.a., their own fixed parameters, that are trained with closed datasets. This leads to significant hallucinations or distortions when facing fine-grained and unseen novel real-world objects, e.g., the appearance of the Tesla Cybertruck. To this end, we present the first real-object-based retrieval-augmented generation framework (RealRAG), which augments fine-grained and unseen novel object generation by learning and retrieving real-world images to overcome the knowledge gaps of generative models. Specifically, to integrate missing memory for unseen novel object generation, we train a reflective retriever by self-reflective contrastive learning, which injects the generator’s knowledge into the sef-reflective negatives, ensuring that the retrieved augmented images compensate for the model’s missing knowledge. Furthermore, the real-object-based framework integrates fine-grained visual knowledge for the generative models, tackling the distortion problem and improving the realism for fine-grained object generation. Our Real-RAG is superior in its modular application to all types of state-of-the-art text-to-image generative models and also delivers remarkable performance boosts with all of them, such as a gain of 16.18% FID score with the auto-regressive model on the Stanford Car benchmark.

The domain of time series forecasting has gained significant attention due to its critical applications in multimedia-rich web traffic (including video streaming workloads and dynamic content delivery) and cross-platform advertisement click predictions, which are essential for web operations planning. While models like TimeSieve have demonstrated strong capabilities in predicting web visitation metrics, they suffer from critical unfaithfulness issues, including sensitivity to random seeds, input noise, layer noise, and parametric perturbations. To address these limitations, we propose Faithful TimeSieve (FTS), an enhanced framework designed to improve prediction reliability and robustness. Our approach systematically detects and mitigates unfaithfulness in TimeSieve, significantly enhancing its stability and consistency. Experimental results demonstrate that FTS substantially improves the model's faithfulness, setting a new standard for temporal forecasting methods. This advancement not only increases TimeSieve's reliability but also contributes to more robust temporal modeling, particularly crucial for web traffic forecasting where prediction accuracy directly impacts operational decisions. Our work thus represents a significant step toward more dependable time series predictions in web-related applications.

Quantization and fine-tuning are crucial for deploying large language models (LLMs) on resource-constrained edge devices. However, fine-tuning quantized models presents significant challenges, primarily stemming from: First, the mismatch in data types between the low-precision quantized weights (e.g., 4-bit) and the high-precision adaptation weights (e.g., 16-bit). This mismatch limits the computational efficiency advantage offered by quantized weights during inference. Second, potential accuracy degradation when merging these high-precision adaptation weights into the low-precision quantized weights, as the adaptation weights often necessitate approximation or truncation. Third, as far as we know, no existing methods support the lossless merging of adaptation while adjusting all quantized weights. To address these challenges, we introduce lossless ternary adaptation for quantization-aware fine-tuning (LoTA-QAF). This is a novel fine-tuning method specifically designed for quantized LLMs, enabling the lossless merging of ternary adaptation weights into quantized weights and the adjustment of all quantized weights. LoTA-QAF operates through a combination of: i) A custom-designed ternary adaptation (TA) that aligns ternary weights with the quantization grid and uses these ternary weights to adjust quantized weights. ii) A TA-based mechanism that enables the lossless merging of adaptation weights. iii) Ternary signed gradient descent (t-SignSGD) for updating the TA weights. We apply LoTA-QAF to Llama-3.1/3.3 and Qwen-2.5 model families and validate its effectiveness on several downstream tasks. On the MMLU benchmark, our method effectively recovers performance for quantized models, surpassing 16-bit LoRA by up to 5.14\%. For task-specific fine-tuning, 16-bit LoRA achieves superior results, but LoTA-QAF still outperforms other methods. Code is available in github.com/KingdalfGoodman/LoTA-QAF.

Large Language Models (LLMs) require significant GPU memory when processing long texts, with the key value (KV) cache consuming up to 70\% of total memory during inference. Although existing compression methods reduce memory by evaluating the importance of individual tokens, they overlook critical semantic relationships between tokens, resulting in fragmented context and degraded performance. We introduce \method{}, which fundamentally reimagines KV cache compression by treating semantic chunks - rather than isolated tokens - as basic compression units. This approach preserves complete linguistic structures and contextual integrity, ensuring that essential meaning is retained even under aggressive compression. Our innovation includes a novel layer-wise index reuse technique that exploits the higher cross-layer similarity of preserved indices in \method{}, reducing computational overhead and improving throughput by 26.5\%. Comprehensive evaluations on challenging benchmarks: LongBench, Needle-In-A-HayStack, GSM8K, and JailbreakV demonstrate that \method{} outperforms state-of-the-art methods by up to 8.7\% in precision while maintaining the same compression ratio. These results confirm that semantic-aware compression significantly enhances both efficiency and performance for long-context LLM inference, providing a simple yet effective solution to the memory bottleneck problem. \emph{The code is available at \href{https://github.com/NVIDIA/kvpress}{link}.}

Multimodal Large Language Models (MLLMs) increasingly excel at perception,understanding, and reasoning. However, current benchmarks inadequately evaluate their ability to perform these tasks continuously in dynamic, real-world environments. To bridge this gap, we introduce RT V-Bench, a fine-grained benchmark for MLLM real-time video analysis. RTV-Bench includes three key principles: (1) Multi-Timestamp Question Answering (MTQA), where answers evolve with scene changes; (2) Hierarchical Question Structure, combining basic and advanced queries; and (3) Multi-dimensional Evaluation, assessing the ability of continuous perception, understanding, and reasoning. RTV-Bench contains 552 diverse videos (167.2 hours) and 4,631 high-quality QA pairs. We evaluated leading MLLMs, including proprietary (GPT-4o, Gemini 2.0), open-source offline (Qwen2.5-VL, VideoLLaMA3), and open-source real-time (VITA-1.5, InternLM-XComposer2.5-OmniLive) models. Experiment results show open-source real-time models largely outperform offline ones but still trail top proprietary models. Our analysis also reveals that larger model size or higher frame sampling rates do not significantly boost RTV-Bench performance, sometimes causing slight decreases.This underscores the need for better model architectures optimized for video stream processing and long sequences to advance real-time video analysis with MLLMs.

Despite their powerful capabilities, multimodal large language models (MLLMs) suffer from considerable computational overhead due to their reliance on massive visual tokens. Recent studies have explored token pruning to alleviate this problem, which typically uses text-vision cross-attention or [CLS] attention to assess and discard redundant visual tokens. In this work, we identify a critical limitation of such attention-first pruning approaches, i.e., they tend to preserve semantically similar tokens, resulting in pronounced performance drops under high pruning rates. To this end, we propose HoloV, a simple yet effective, plug-and-play visual token pruning framework for efficient inference. Distinct from previous attention-first schemes, HoloV rethinks token retention from a holistic perspective. By adaptively distributing the pruning budget across different spatial crops, HoloV ensures that the retained tokens capture the global visual context rather than isolated salient features. This strategy minimizes representational collapse and maintains task-relevant information even under aggressive pruning. Experimental results demonstrate that our HoloV achieves superior performance across various tasks, MLLM architectures, and pruning ratios compared to SOTA methods. For instance, LLaVA1.5 equipped with HoloV preserves 95.8% of the original performance after pruning 88.9% of visual tokens, achieving superior efficiency-accuracy trade-offs.

Given that substantial amounts of domain-specific knowledge are stored in structured formats, such as web data organized through HTML, Large Language Models (LLMs) are expected to fully comprehend this structured information to broaden their applications in various real-world downstream tasks. Current approaches for applying LLMs to structured data fall into two main categories: serialization-based and operation-based methods. Both approaches, whether relying on serialization or using SQL-like operations as an intermediary, encounter difficulties in fully capturing structural relationships and effectively handling sparse data. To address these unique characteristics of structured data, we propose HyperG, a hypergraph-based generation framework aimed at enhancing LLMs' ability to process structured knowledge. Specifically, HyperG first augment sparse data with contextual information, leveraging the generative power of LLMs, and incorporate a prompt-attentive hypergraph learning (PHL) network to encode both the augmented information and the intricate structural relationships within the data. To validate the effectiveness and generalization of HyperG, we conduct extensive experiments across two different downstream tasks requiring structured knowledge. Our code is publicly available at: https://github.com/s1ruihuang/HyperG.

Waterway perception is critical for the special operations and autonomous navigation of Unmanned Surface Vessels (USVs), but current perception schemes are sensor-based, neglecting the interaction between humans and USVs for embodied perception in various operations. Therefore, inspired by visual grounding, we present WaterVG, the inaugural visual grounding dataset tailored for USV-based waterway perception guided by human prompts. WaterVG contains a wealth of prompts describing multiple targets, with instance-level annotations, including bounding boxes and masks. Specifically, WaterVG comprises 11,568 samples and 34,987 referred targets, integrating both visual and radar characteristics. The text-guided two-sensor pattern provides a fine granularity of text prompts aligned with the visual and radar features of the referent targets, containing both qualitative and numeric descriptions. To enhance the endurance and maintain the normal operations of USVs in open waterways, we propose Potamoi, a low-power visual grounding model. Potamoi is a multi-task model employing a sophisticated Phased Heterogeneous Modality Fusion (PHMF) mechanism, which includes Adaptive Radar Weighting (ARW) and Multi-Head Slim Cross Attention (MHSCA). The ARW module utilizes a gating mechanism to adaptively extract essential radar features for fusion with visual inputs, ensuring prompt alignment. MHSCA, characterized by its low parameter count and computational efficiency (FLOPs), effectively integrates contextual information from both sensors with linguistic features, delivering outstanding performance in visual grounding tasks. Comprehensive experiments and evaluations on WaterVG demonstrate that Potamoi achieves state-of-the-art results compared to existing methods. The project is available athttps://github.com/GuanRunwei/WaterVG.

Cross-domain named entity recognition (NER) tasks encourage NER models to transfer knowledge from data-rich source domains to sparsely labeled target domains. Previous works adopt the paradigms of pre-training on the source domain followed by fine-tuning on the target domain. However, these works ignore that general labeled NER source domain data can be easily retrieved in the real world, and soliciting more source domains could bring more benefits. Unfortunately, previous paradigms cannot efficiently transfer knowledge from multiple source domains. In this work, to transfer multiple source domains' knowledge, we decouple the NER task into the pipeline tasks of mention detection and entity typing, where the mention detection unifies the training object across domains, thus providing the entity typing with higher-quality entity mentions. Additionally, we request multiple general source domain models to suggest the potential named entities for sentences in the target domain explicitly, and transfer their knowledge to the target domain models through the knowledge progressive networks implicitly. Furthermore, we propose two methods to analyze in which source domain knowledge transfer occurs, thus helping us judge which source domain brings the greatest benefit. In our experiment, we develop a Chinese cross-domain NER dataset. Our model improved the F1 score by an average of 12.50% across 8 Chinese and English datasets compared to models without source domain data.

Junzhe Chen, Xuming Hu, Shuodi Liu, Shiyu Huang, Wei-Wei Tu, Zhaofeng He, Lijie Wen. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Projecting visual features into word embedding space has become a significant fusion strategy adopted by Multimodal Large Language Models (MLLMs).However, its internal mechanisms have yet to be explored.Inspired by multilingual research, we identify domain-specific neurons in multimodal large language models.Specifically, we investigate the distribution of domain-specific neurons and the mechanism of how MLLMs process features from diverse domains.Furthermore, we propose a threestage mechanism for language model modules in MLLMs when handling projected image features, and verify this hypothesis using logit lens.Extensive experiments indicate that while current MLLMs exhibit Visual Question Answering (VQA) capability, they may not fully utilize domain-specific information.Manipulating domain-specific neurons properly will result in a 10% change of accuracy at most, shedding light on the development of cross-domain, all-encompassing MLLMs in the future.The source code is available at this URL.

Recently, large language models (LLMs) have achieved tremendous breakthroughs in the field of NLP, but still lack understanding of their internal neuron activities when processing different languages. We designed a method to convert dense LLMs into fine-grained MoE architectures, and then visually studied the multilingual activation patterns of LLMs through expert activation frequency heatmaps. Through comprehensive experiments on different model families, different model sizes, and different variants, we analyzed the similarities and differences in the internal neuron activation patterns of LLMs when processing different languages. Specifically, we investigated the distribution of high-frequency activated experts, multilingual shared experts, whether multilingual activation patterns are related to language families, and the impact of instruction tuning on activation patterns. We further explored leveraging the discovered differences in expert activation frequencies to guide sparse activation and pruning. Experimental results demonstrated that our method significantly outperformed random expert pruning and even exceeded the performance of unpruned models in some languages. Additionally, we found that configuring different pruning rates for different layers based on activation level differences could achieve better results. Our findings reveal the multilingual processing mechanisms within LLMs and utilize these insights to offer new perspectives for applications such as sparse activation and model pruning.

Large language models (LLMs) have recently driven striking performance improvements across a range of natural language processing tasks. The factual knowledge acquired during pretraining and instruction tuning can be useful in various downstream tasks, such as question answering, and language generation. Unlike conventional Knowledge Bases (KBs) that explicitly store factual knowledge, LLMs implicitly store facts in their parameters. Content generated by the LLMs can often exhibit inaccuracies or deviations from the truth, due to facts that can be incorrectly induced or become obsolete over time. To this end, we aim to explore the extent and scope of factual knowledge within LLMs by designing the benchmark Pinocchio. Pinocchio contains 20K diverse factual questions that span different sources, timelines, domains, regions, and languages. Furthermore, we investigate whether LLMs can compose multiple facts, update factual knowledge temporally, reason over multiple pieces of facts, identify subtle factual differences, and resist adversarial examples. Extensive experiments on different sizes and types of LLMs show that existing LLMs still lack factual knowledge and suffer from various spurious correlations. We believe this is a critical bottleneck for realizing trustworthy artificial intelligence. The dataset Pinocchio and our codes are publicly available at: https://github.com/THU-BPM/Pinocchio.

Recently, text watermarking algorithms for large language models (LLMs) have been proposed to mitigate the potential harms of text generated by LLMs, including fake news and copyright issues. However, current watermark detection algorithms require the secret key used in the watermark generation process, making them susceptible to security breaches and counterfeiting during public detection. To address this limitation, we propose an unforgeable publicly verifiable watermark algorithm named UPV that uses two different neural networks for watermark generation and detection, instead of using the same key at both stages. Meanwhile, the token embedding parameters are shared between the generation and detection networks, which makes the detection network achieve a high accuracy very efficiently. Experiments demonstrate that our algorithm attains high detection accuracy and computational efficiency through neural networks. Subsequent analysis confirms the high complexity involved in forging the watermark from the detection network. Our code is available at https://github.com/THU-BPM/unforgeable_watermark

Watermark algorithms for large language models (LLMs) have achieved extremely high accuracy in detecting text generated by LLMs. Such algorithms typically involve adding extra watermark logits to the LLM's logits at each generation step. However, prior algorithms face a trade-off between attack robustness and security robustness. This is because the watermark logits for a token are determined by a certain number of preceding tokens; a small number leads to low security robustness, while a large number results in insufficient attack robustness. In this work, we propose a semantic invariant watermarking method for LLMs that provides both attack robustness and security robustness. The watermark logits in our work are determined by the semantics of all preceding tokens. Specifically, we utilize another embedding LLM to generate semantic embeddings for all preceding tokens, and then these semantic embeddings are transformed into the watermark logits through our trained watermark model. Subsequent analyses and experiments demonstrated the attack robustness of our method in semantically invariant settings: synonym substitution and text paraphrasing settings. Finally, we also show that our watermark possesses adequate security robustness.

Weize Liu, Guocong Li, Kai Zhang, Bang Du, Qiyuan Chen, Xuming Hu, Hongxia Xu, Jintai Chen, Jian Wu. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Recently, Large Language Models (LLMs) make remarkable evolutions in language understanding and generation. Following this, various benchmarks for measuring all kinds of capabilities of LLMs have sprung up. In this paper, we challenge the reasoning and understanding abilities of LLMs by proposing a FaLlacy Understanding Benchmark (FLUB) containing cunning texts that are easy for humans to understand but difficult for models to grasp. Specifically, the cunning texts that FLUB focuses on mainly consist of the tricky, humorous, and misleading texts collected from the real internet environment. And we design three tasks with increasing difficulty in the FLUB benchmark to evaluate the fallacy understanding ability of LLMs. Based on FLUB, we investigate the performance of multiple representative and advanced LLMs, reflecting our FLUB is challenging and worthy of more future study. Interesting discoveries and valuable insights are achieved in our extensive experiments and detailed analyses. We hope that our benchmark can encourage the community to improve LLMs' ability to understand fallacies. Our data and codes are available at https://github.com/THUKElab/FLUB.

As large language models (LLMs) become increasingly capable and widely deployed, the risk of them generating misinformation poses a critical challenge. Misinformation from LLMs can take various forms, from factual errors due to hallucination to intentionally deceptive content, and can have severe consequences in high-stakes domains.This tutorial covers comprehensive strategies to prevent and detect misinformation generated by LLMs. We first introduce the types of misinformation LLMs can produce and their root causes. We then explore two broad categories: Preventing misinformation generation: a) AI alignment training techniques to reduce LLMs' propensity for misinformation and refuse malicious instructions during model training. b) Training-free mitigation methods like prompt guardrails, retrieval-augmented generation (RAG), and decoding strategies to curb misinformation at inference time. Detecting misinformation after generation, including a) using LLMs themselves to detect misinformation through embedded knowledge or retrieval-enhanced judgments, and b) distinguishing LLM-generated text from human-written text through black-box approaches (e.g., classifiers, probability analysis) and white-box approaches (e.g., watermarking). We also discuss the challenges and limitations of detecting LLM-generated misinformation.

Open relation extraction is the task of extracting open-domain relation facts from natural language sentences. Existing works either utilize distant-supervised annotations to train a supervised classifier over pre-defined relations, or adopt unsupervised methods with additional dependency on external assumptions. However, these works can only obtain information signals from limited existing knowledge bases or datasets. In this work, we propose a self-supervised framework namedWeb-SelfORE, which exploits self-supervised signals by requiring a large pretrained language model to extensively read real-world relevant documents from the web, and obtain contextualized relational features by mixing contextualized representations of entities from different documents. We perform adaptive clustering on contextualized relational features and bootstrap the self-supervised signals by improving contextualized features in relation classification. We additionally compare the effectiveness of self-supervisions brought by different document sources, and introduce relevance and redundancy evaluation metrics to obtain higher-quality self-supervisions. Experimental results on four public datasets show the effectiveness and robustness ofWeb-SelfOREon open-domain relation extraction task when comparing with competitive baselines.

Concept relatedness estimation (CRE) aims to determine whether two given concepts are related. Existing methods only consider the pairwise relationship between concepts, while overlooking the higher-order relationship that could be encoded in a concept-level graph structure. We discover that this underlying graph satisfies a set of intrinsic properties of CRE, including reflexivity, commutativity, and transitivity. In this paper, we formalize the CRE properties and introduce a graph structure named ConcreteGraph. To address the data scarcity issue in CRE, we introduce a novel data augmentation approach to sample new concept pairs from the graph. As it is intractable for data augmentation to fully capture the structural information of the ConcreteGraph due to a large amount of potential concept pairs, we further introduce a novel Graph Component Contrastive Learning framework to implicitly learn the complete structure of the ConcreteGraph. Empirical results on three datasets show significant improvement over the state-of-the-art model. Detailed ablation studies demonstrate that our proposed approach can effectively capture the high-order relationship among concepts.

Aspect-based sentiment analysis (ABSA) is a fine-grained sentiment classification task.Many recent works have used dependency trees to extract the relation between aspects and contexts and have achieved significant improvements.However, further improvement is limited due to the potential mismatch between the dependency tree as a syntactic structure and the sentiment classification as a semantic task.To alleviate this gap, we replace the syntactic dependency tree with the semantic structure named Abstract Meaning Representation (AMR) and propose a model called AMR-based Path Aggregation Relational Network (APARN) to take full advantage of semantic structures.In particular, we design the path aggregator and the relation-enhanced selfattention mechanism that complement each other.The path aggregator extracts semantic features from AMRs under the guidance of sentence information, while the relationenhanced self-attention mechanism in turn improves sentence features with refined semantic information.Experimental results on four public datasets demonstrate 1.13% average F1 improvement of APARN in ABSA when compared with state-of-the-art baselines.1

How to identify semantic relations among entities in a document when only a few labeled documents are available?Few-shot documentlevel relation extraction (FSDLRE) is crucial for addressing the pervasive data scarcity problem in real-world scenarios.Metric-based meta-learning is an effective framework widely adopted for FSDLRE, which constructs class prototypes for classification.However, existing works often struggle to obtain class prototypes with accurate relational semantics: 1) To build prototype for a target relation type, they aggregate the representations of all entity pairs holding that relation, while these entity pairs may also hold other relations, thus disturbing the prototype.2) They use a set of generic NOTA (none-of-the-above) prototypes across all tasks, neglecting that the NOTA semantics differs in tasks with different target relation types.In this paper, we propose a relation-aware prototype learning method for FSDLRE to strengthen the relational semantics of prototype representations.By judiciously leveraging the relation descriptions and realistic NOTA instances as guidance, our method effectively refines the relation prototypes and generates task-specific NOTA prototypes.Extensive experiments demonstrate that our method outperforms state-of-the-art approaches by average 2.61% F 1 across various settings of two FSDLRE benchmarks. 1

Named Entity Recognition (NER) is a well and widely studied task in natural language processing. Recently, the nested NER has attracted more attention since its practicality and difficulty. Existing works for nested NER ignore the recognition order and boundary position relation of nested entities. To address these issues, we propose a novel seq2seq model named GPRL, which formulates the nested NER task as an entity triplet sequence generation process. GPRL adopts the reinforcement learning method to generate entity triplets de-coupling the entity order in gold labels and expects to learn a reasonable recognition order of entities via trial and error. Based on statistics of boundary distance for nested entities, GPRL designs a Gaussian prior to represent the boundary distance distribution between nested entities and adjust the out-put probability distribution of nested boundary tokens. Experiments on three nested NER datasets demonstrate that GPRL outperforms previous nested NER models.

How can we better extract entities and relations from text? Using multimodal extraction with images and text obtains more signals for entities and relations, and aligns them through graphs or hierarchical fusion, aiding in extraction. Despite attempts at various fusions, previous works have overlooked many unlabeled image-caption pairs, such as NewsCLIPing. This paper proposes innovative pre-training objectives for entity-object and relation-image alignment, extracting objects from images and aligning them with entity and relation prompts for soft pseudo-labels. These labels are used as self-supervised signals for pre-training, enhancing the ability to extract entities and relations. Experiments on three datasets show an average 3.41% F1 improvement over prior SOTA. Additionally, our method is orthogonal to previous multimodal fusions, and using it on prior SOTA fusions further improves 5.47% F1.

Multimodal Named Entity Recognition (MNER) and Multimodal Relation Extraction (MRE) are tasks in information retrieval that aim to recognize entities and extract relations among them using information from multiple modalities, such as text and images. Although current methods have attempted a variety of modality fusion approaches to enhance the information in text, a large amount of readily available internet retrieval data has not been considered. Therefore, we attempt to retrieve real-world text related to images, objects, and entire sentences from the internet and use this retrieved text as input for cross-modal fusion to improve the performance of entity and relation extraction tasks in the text.

Low-resource relation extraction (LRE) aims to extract potential relations from limited labeled corpus to handle the problem of scarcity of human annotations. Previous works mainly consist of two categories of methods: (1) Self-training methods, which improve themselves through the models' predictions, thus suffering from confirmation bias when the predictions are wrong. (2) Self-ensembling methods, which learn task-agnostic representations, therefore, generally do not work well for specific tasks. In our work, we propose a novel LRE architecture named SelfLRE, which leverages two complementary modules, one module uses self-training to obtain pseudo-labels for unlabeled data, and the other module uses self-ensembling learning to obtain the task-agnostic representations, and leverages the existing pseudo-labels to refine the better task-specific representations on unlabeled data. The two models are jointly trained through multi-task learning to iteratively improve the effect of LRE task. Experiments on three public datasets show that SelfLRE achieves 1.81% performance gain over the SOTA baseline. Source code is available at: https://github.com/THU-BPM/SelfLRE.

As social media platforms are evolving from text-based forums into multi-modal environments, the nature of misinformation in social media is also transforming accordingly. Misinformation spreaders have recently targeted contextual connections between the modalities e.g., text and image. However, existing datasets for rumor detection mainly focus on a single modality i.e., text. To bridge this gap, we construct MR2, a multimodal multilingual retrieval-augmented dataset for rumor detection. The dataset covers rumors with images and texts, and provides evidence from both modalities that are retrieved from the Internet. Further, we develop established baselines and conduct a detailed analysis of the systems evaluated on the dataset. Extensive experiments show that MR2 will provide a challenging testbed for developing rumor detection systems designed to retrieve and reason over social media posts. Source code and data are available at: https://github.com/THU-BPM/MR2.

Real-world fact verification task aims to verify the factuality of a claim by retrieving evidence from the source document. The quality of the retrieved evidence plays an important role in claim verification. Ideally, the retrieved evidence should be faithful (reflecting the model's decision-making process in claim verification) and plausible (convincing to humans), and can improve the accuracy of verification task. Although existing approaches leverage the similarity measure of semantic or surface form between claims and documents to retrieve evidence, they all rely on certain heuristics that prevent them from satisfying all three requirements. In light of this, we propose a fact verification model named ReRead to retrieve evidence and verify claim that: (1) Train the evidence retriever to obtain interpretable evidence (i.e., faithfulness and plausibility criteria); (2) Train the claim verifier to revisit the evidence retrieved by the optimized evidence retriever to improve the accuracy. The proposed system is able to achieve significant improvements upon best-reported models under different settings.

Relation extraction (RE) aims to extract potential relations according to the context of two entities, thus, deriving rational contexts from sentences plays an important role. Previous works either focus on how to leverage the entity information (e.g., entity types, entity verbalization) to inference relations, but ignore context-focused content, or use counterfactual thinking to remove the model's bias of potential relations in entities, but the relation reasoning process will still be hindered by irrelevant content. Therefore, how to preserve relevant content and remove noisy segments from sentences is a crucial task. In addition, retained content needs to be fluent enough to maintain semantic coherence and interpretability. In this work, we propose a novel rationale extraction framework named RE2, which leverages two continuity and sparsity factors to obtain relevant and coherent rationales from sentences. To solve the problem that the gold rationales are not labeled, RE2 applies an optimizable binary mask to each token in the sentence, and adjust the rationales that need to be selected according to the relation label. Experiments on four datasets show that RE2 surpasses baselines.

Natural Language Inference (NLI) is a growingly essential task in natural language understanding, which requires inferring the relationship between the sentence pairs (premiseandhypothesis). Recently, low-resource natural language inference has gained increasing attention, due to significant savings in manual annotation costs and a better fit with real-world scenarios. Existing works fail to characterize discriminative representations between different classes with limited training data, which may cause faults in label prediction. Here we propose a multi-level supervised contrastive learning framework named MultiSCL for low-resource natural language inference. MultiSCL leverages a sentence-level and pair-level contrastive learning objective to discriminate between different classes of sentence pairs by bringing those in one class together and pushing away those in different classes. MultiSCL adopts a data augmentation module that generates different views for input samples to better learn the latent representation. The pair-level representation is obtained from a cross attention module. We conduct extensive experiments on two public NLI datasets in low-resource settings, and the accuracy of MultiSCL exceeds other models by 1.8%, 3.1% and 4.1% on SNLI, MNLI and Sick with 5 instances per label respectively. Moreover, our method outperforms the previous state-of-the-art method on cross-domain tasks of text classification.

Utilizing Visualization-oriented Natural Language Interfaces (V-NLI) as a complementary input modality to direct manipulation for visual analytics can provide an engaging user experience. It enables users to focus on their tasks rather than having to worry about how to operate visualization tools on the interface. In the past two decades, leveraging advanced natural language processing technologies, numerous V-NLI systems have been developed in academic research and commercial software, especially in recent years. In this article, we conduct a comprehensive review of the existing V-NLIs. In order to classify each article, we develop categorical dimensions based on a classic information visualization pipeline with the extension of a V-NLI layer. The following seven stages are used: query interpretation, data transformation, visual mapping, view transformation, human interaction, dialogue management, and presentation. Finally, we also shed light on several promising directions for future work in the V-NLI community.

A scene graph is a semantic representation that expresses the objects, attributes, and relationships between objects in a scene. Scene graphs play an important role in many cross modality tasks, as they are able to capture the interactions between images and texts. In this paper, we focus on scene graph modification (SGM), where the system is required to learn how to update an existing scene graph based on a natural language query. Unlike previous approaches that rebuilt the entire scene graph, we frame SGM as a graph expansion task by introducing the incremental structure expanding (ISE). ISE constructs the target graph by incrementally expanding the source graph without changing the unmodified structure. Based on ISE, we further propose a model that iterates between nodes prediction and edges prediction, inferring more accurate and harmonious expansion decisions progressively. In addition, we construct a challenging dataset that contains more complicated queries and larger scene graphs than existing datasets. Experiments on four benchmarks demonstrate the effectiveness of our approach, which surpasses the previous state-of-the-art model by large margins.

Successful Machine Learning based Named Entity Recognition models could fail on texts from some special domains, for instance, Chinese addresses and e-commerce titles, where requires adequate background knowledge. Such texts are also difficult for human annotators. In fact, we can obtain some potentially helpful information from correlated texts, which have some common entities, to help the text understanding. Then, one can easily reason out the correct answer by referencing correlated samples. In this paper, we suggest enhancing NER models with correlated samples. We draw correlated samples by the sparse BM25 retriever from large-scale in-domain unlabeled data. To explicitly simulate the human reasoning process, we perform a training-free entity type calibrating by majority voting. To capture correlation features in the training stage, we suggest to model correlated samples by the transformer-based multi-instance cross-encoder. Empirical results on datasets of the above two domains show the efficacy of our methods.

We propose the first character-level white-box adversarial attack method against transformer models.The intuition of our method comes from the observation that words are split into subtokens before being fed into the transformer models and the substitution between two close subtokens has a similar effect to the character modification.Our method mainly contains three steps.First, a gradient-based method is adopted to find the most vulnerable words in the sentence.Then we split the selected words into subtokens to replace the origin tokenization result from the transformer tokenizer.Finally, we utilize an adversarial loss to guide the substitution of attachable subtokens in which the Gumbel-softmax trick is introduced to ensure gradient propagation.Meanwhile, we introduce the visual and length constraint in the optimization process to achieve minimum character modifications.Extensive experiments on both sentence-level and token-level tasks demonstrate that our method could outperform the previous attack methods in terms of success rate and edit distance.Furthermore, human evaluation verifies our adversarial examples could preserve their origin labels.

Joint entity and relation extraction has been a core task in the field of information extraction.Recent approaches usually consider the extraction of relational triples from a stereoscopic perspective, either learning a relation-specific tagger or separate classifiers for each relation type.However, they still suffer from error propagation, relation redundancy and lack of highlevel connections between triples.To address these issues, we propose a novel query-based approach to construct instance-level representations for relational triples.By metric-based comparison between query embeddings and token embeddings, we can extract all types of triples in one step, thus eliminating the error propagation problem.In addition, we learn the instance-level representation of relational triples via contrastive learning.In this way, relational triples can not only enclose rich classlevel semantics but also access to high-order global connections.Experimental results show that our proposed method achieves the state of the art on five widely used benchmarks.

Natural language inference (NLI) is an increasingly important task for natural language understanding, which requires one to infer the relationship between the sentence pair (premise and hypothesis). Many recent works have used contrastive learning by incorporating the relationship of the sentence pair from NLI datasets to learn sentence representation. However, these methods only focus on comparisons with sentence-level representations. In this paper, we propose a Pair-level Supervised Contrastive Learning approach (PairSCL). We adopt a cross attention module to learn the joint representations of the sentence pairs. A contrastive learning objective is designed to distinguish the varied classes of sentence pairs by pulling those in one class together and pushing apart the pairs in other classes. We evaluate PairSCL on two public datasets of NLI where the accuracy of PairSCL outperforms other methods by 2.1% on average. Furthermore, our method outperforms the previous state-of-the-art method on seven transfer tasks of text classification.1

The generalizability to new databases is of vital importance to Text-to-SQL systems which aim to parse human utterances into SQL statements. Existing works achieve this goal by leveraging the exact matching method to identify the lexical matching between the question words and the schema items. However, these methods fail in other challenging scenarios, such as the synonym substitution in which the surface form differs between the corresponding question words and schema items. In this paper, we propose a framework named ISESL-SQL to iteratively build a semantic enhanced schema-linking graph between question tokens and database schemas. First, we extract a schema linking graph from PLMs through a probing procedure in an unsupervised manner. Then the schema linking graph is further optimized during the training process through a deep graph learning method. Meanwhile, we also design an auxiliary task called graph regularization to improve the schema information mentioned in the schema-linking graph. Extensive experiments on three benchmarks demonstrate that ISESL-SQL could consistently outperform the baselines and further investigations show its generalizability and robustness.

Xuming Hu, Zhijiang Guo, GuanYu Wu, Aiwei Liu, Lijie Wen, Philip Yu. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Shuliang Liu, Xuming Hu, Chenwei Zhang, Shu’ang Li, Lijie Wen, Philip Yu. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

Prediction over event sequences is critical for many real-world applications in Information Retrieval and Natural Language Processing. Future Event Generation (FEG) is a challenging task in event sequence prediction because it requires not only fluent text generation but also commonsense reasoning to maintain the logical coherence of the entire event story. In this paper, we propose a novel explainable FEG framework, Coep. It highlights and integrates two types of event knowledge, sequential knowledge of direct event-event relations and inferential knowledge that reflects the intermediate character psychology between events, such as intents, causes, reactions, which intrinsically pushes the story forward. To alleviate the knowledge forgetting issue, we design two modules, IM and GM, for each type of knowledge, which are combined via prompt tuning. First, IM focuses on understanding inferential knowledge to generate commonsense explanations and provide a soft prompt vector for GM. We also design a contrastive discriminator for better generalization ability. Second, GM generates future events by modeling direct sequential knowledge with the guidance of IM. Automatic and human evaluation demonstrate that our approach can generate more coherent, specific, and logical future events.

Distributed in-memory computing frameworks usually have lots of parameters (e.g., the buffer size of shuffle) to form a configuration for each execution. A well-tuned configuration can bring large improvements of performance. However, to improve resource utilization, jobs are often share the same cluster, which causes dynamic cluster load conditions. According to our observation, the variation of cluster load reduces effectiveness of configuration tuning. Besides, as a common problem of cluster computing jobs, overestimation of resources also occurs during configuration tuning. It is challenging to efficiently find the optimal configuration in a shared cluster with the consideration of memory-sparing. In this article, we introduce MespaConfig, a job-level configuration optimizer for distributed in-memory computing jobs. Advancements of MespaConfig over previous work are features including memory-sparing and load-sensitive. We evaluate MespaConfig by 6 typical Spark programs under different load conditions. The evaluation results show that MespaConfig improves the performance of six typical programs by up to 12× compared with default configurations. MespaConfig also achieves at most 41 percent reduction of configuration memory usage and reduces the optimization time overhead by 10.8× compared with the state-of-the-art approach.

Low-resource Relation Extraction (LRE) aims to extract relation facts from limited labeled corpora when human annotation is scarce.Existing works either utilize self-training scheme to generate pseudo labels that will cause the gradual drift problem, or leverage metalearning scheme which does not solicit feedback explicitly.To alleviate selection bias due to the lack of feedback loops in existing LRE learning paradigms, we developed a Gradient Imitation Reinforcement Learning method to encourage pseudo label data to imitate the gradient descent direction on labeled data and bootstrap its optimization capability through trial and error.We also propose a framework called GradLRE, which handles two major scenarios in low-resource relation extraction.Besides the scenario where unlabeled data is sufficient, GradLRE handles the situation where no unlabeled data is available, by exploiting a contextualized augmentation method to generate data.Experimental results on two public datasets demonstrate the effectiveness of GradLRE on low resource relation extraction when comparing with baselines.Source code is available 1 .

Open relation extraction is the task of extracting open-domain relation facts from natural language sentences.Existing works either utilize heuristics or distant-supervised annotations to train a supervised classifier over pre-defined relations, or adopt unsupervised methods with additional assumptions that have less discriminative power.In this work, we propose a self-supervised framework named SelfORE, which exploits weak, self-supervised signals by leveraging large pretrained language model for adaptive clustering on contextualized relational features, and bootstraps the self-supervised signals by improving contextualized features in relation classification.Experimental results on three datasets show the effectiveness and robustness of SelfORE on open-domain Relation Extraction when comparing with competitive baselines.Source code is available 1 .