Long-context modeling is crucial for next-generation language models, yet the high computational cost of standard attention mechanisms poses significant computational challenges. Sparse attention offers a promising direction for improving efficiency while maintaining model capabilities. We present NSA, a Natively trained Sparse Attention mechanism that integrates algorithmic innovations with hardware-aligned optimizations to achieve efficient long-context modeling. NSA employs a dynamic hierarchical sparse strategy, combining coarse-grained token compression with fine-grained token selection to preserve both global context awareness and local precision. Our approach advances sparse attention design with two key innovations: (1) We achieve substantial speedups through arithmetic intensity-balanced algorithm design, with implementation optimizations for modern hardware. (2) We enable end-to-end training, reducing pretraining computation without sacrificing model performance. As shown in Figure 1, experiments show the model pretrained with NSA maintains or exceeds Full Attention models across general benchmarks, long-context tasks, and instruction-based reasoning. Meanwhile, NSA achieves substantial speedups over Full Attention on 64k-length sequences across decoding, forward propagation, and backward propagation, validating its efficiency throughout the model lifecycle.
论文检索
输入标题、作者或关键词，从 3,086 篇学术成果中精准定位
Knowledge distillation (KD) compresses large language models (LLMs), known as teacher models, into lightweight versions called student models, enabling efficient inference and downstream applications. However, prevailing approaches accomplish this by predominantly focusing on matching the final output distributions of student/teacher models. Drawing on the perspective that transformers can be viewed as discretizing ordinary differential equation (ODEs) on integer time steps (corresponding to layer indices), where intermediate features evolve across layers, we argue that effective KD requires aligning the entire feature dynamics between teacher and student models, which we call feature dynamics distillation (FDD). This alignment involves matching both the feature trajectory and its first-order derivative, rather than just the final states. Our approach extends the original KD objective with two additional loss terms: layer-wise feature KD, which matches discretized feature trajectory, and layer feature delta KD, which matches first-order changes in features across adjacent layers. Extensive experiments on various tasks validate the effectiveness of our distillation method.
Large Language Models (LLMs) have advanced rapidly but face significant memory demands. While quantization has shown promise for LLMs, current methods typically require lengthy training to alleviate the performance degradation from quantization loss. However, deploying LLMs across diverse scenarios with different resource constraints, e.g., servers and personal computers, requires repeated training per application, which amplifies the lengthy training problem. Given that, it is advantageous to train a once-for-all (OFA) supernet capable of yielding diverse optimal subnets for downstream applications through one-shot training. Nonetheless, the scale of current language models impedes efficiency and amplifies interference from weight sharing between subnets. We make an initial attempt to extend the once-for-all framework to large language models. Specifically, we decouple shared weights to eliminate the interference and incorporate Low-Rank adapters for training efficiency. Furthermore, we observe the imbalance allocation of training resources from the traditional uniform sampling. A non-parametric scheduler is introduced to adjust the sampling rate for each quantization configuration, achieving a more balanced allocation among subnets with varying demands. We validate the approach on LLaMA2 families and Mistral on downstream evaluation, demonstrating high performance while significantly reducing deployment time faced with multiple scenarios.
HoPE: A Novel Positional Encoding Without Long-Term Decay for Enhanced Context Awareness and Extrapolation
PDF ↗Many positional encodings (PEs) are designed to exhibit long-term decay, based on an entrenched and long-standing inductive opinion: tokens farther away from the current position carry less relevant information. We argue that long-term decay is outdated in the era of LLMs, as LLMs are now applied to tasks demanding precise retrieval of in-context information from arbitrary positions. Firstly, we present empirical analyses on various PEs, demonstrating that models inherently learn attention with only a local-decay pattern while forming a U-shape pattern globally, contradicting the principle of long-term decay. Furthermore, we conduct a detailed analysis of rotary position encoding (RoPE, a prevalent relative positional encoding in LLMs), and found that the U-shape attention is caused by some learned components, which are also the key factor limiting RoPE’s expressiveness and extrapolation. Inspired by these insights, we propose High-frequency rotary Position Encoding (HoPE). HoPE replaces the specific components in RoPE with position-independent ones, retaining only high-frequency signals, which also breaks the principle of long-term decay in theory. HoPE achieves two major advantages: (1) Without constraints imposed by long-term decay, contradictory factors that limit attention optimization are removed. Thus, the model’s context awareness is enhanced. (2) HoPE exhibits greater robustness to the out-of-distribution behavior in attention patterns during extrapolation. The effectiveness of HoPE is validated through extensive experiments and with a large language model of up to 3 billion parameters.
WET: Overcoming Paraphrasing Vulnerabilities in Embeddings-as-a-Service with Linear Transformation Watermarks
PDF ↗Embeddings-as-a-Service (EaaS) is a service offered by large language model (LLM) developers to supply embeddings generated by LLMs. Previous research suggests that EaaS is prone to imitation attacks—attacks that clone the underlying EaaS model by training another model on the queried embeddings. As a result, EaaS watermarks are introduced to protect the intellectual property of EaaS providers. In this paper, we first show that existing EaaS watermarks can be removed by paraphrasing when attackers clone the model. Subsequently, we propose a novel watermarking technique that involves linearly transforming the embeddings, and show that it is empirically and theoretically robust against paraphrasing.
From English to Second Language Mastery: Enhancing LLMs with Cross-Lingual Continued Instruction Tuning
PDF ↗Supervised Fine-Tuning (SFT) with translated instruction data effectively adapts Large Language Models (LLMs) from English to non-English languages. We introduce Cross-Lingual Continued Instruction Tuning (X-CIT), which fully leverages translation-based parallel instruction data to enhance cross-lingual adaptability. X-CIT emulates the human process of second language acquisition and is guided by Chomsky’s Principles and Parameters Theory. It first fine-tunes the LLM on English instruction data to establish foundational capabilities (i.e. Principles), then continues with target language translation and customized chat-instruction data to adjust “parameters” specific to the target language. This chat-instruction data captures alignment information in translated parallel data, guiding the model to initially think and respond in its native language before transitioning to the target language. To further mimic human learning progression, we incorporate Self-Paced Learning (SPL) during continued training, allowing the model to advance from simple to complex tasks. Implemented on Llama-2-7B across five languages, X-CIT was evaluated against three objective benchmarks and an LLM-as-a-judge benchmark, improving the strongest baseline by an average of 1.97% and 8.2% in these two benchmarks, respectively.
Are We in the AI-Generated Text World Already? Quantifying and Monitoring AIGT on Social Media
PDF ↗Social media platforms are experiencing a growing presence of AI-Generated Texts (AIGTs). However, the misuse of AIGTs could have profound implications for public opinion, such as spreading misinformation and manipulating narratives. Despite its importance, it remains unclear how prevalent AIGTs are on social media. To address this gap, this paper aims to quantify and monitor the AIGTs on online social media platforms. We first collect a dataset (SM-D) with around 2.4M posts from 3 major social media platforms: Medium, Quora, and Reddit. Then, we construct a diverse dataset (AIGTBench) to train and evaluate AIGT detectors. AIGTBench combines popular open-source datasets and our AIGT datasets generated from social media texts by 12 LLMs, serving as a benchmark for evaluating mainstream detectors. With this setup, we identify the best-performing detector (OSM-Det). We then apply OSM-Det to SM-D to track AIGTs across social media platforms from January 2022 to October 2024, using the AI Attribution Rate (AAR) as the metric. Specifically, Medium and Quora exhibit marked increases in AAR, rising from 1.77% to 37.03% and 2.06% to 38.95%, respectively. In contrast, Reddit shows slower growth, with AAR increasing from 1.31% to 2.45% over the same period. Our further analysis indicates that AIGTs on social media differ from human-written texts across several dimensions, including linguistic patterns, topic distributions, engagement levels, and the follower distribution of authors. We envision our analysis and findings on AIGTs in social media can shed light on future research in this domain.
Alignment techniques have become central to ensuring that Large Language Models (LLMs) generate outputs consistent with human values. However, existing alignment paradigms often model an averaged or monolithic preference, failing to account for the diversity of perspectives across cultures, demographics, and communities. This limitation is particularly critical in health-related scenarios, where plurality is essential due to the influence of culture, religion, personal values, and conflicting opinions. Despite progress in pluralistic alignment, no prior work has focused on health, likely due to the unavailability of publicly available datasets. To address this gap, we introduce VITAL, a new benchmark dataset comprising 13.1K value-laden situations and 5.4K multiple-choice questions focused on health, designed to assess and benchmark pluralistic alignment methodologies. Through extensive evaluation of eight LLMs of varying sizes, we demonstrate that existing pluralistic alignment techniques fall short in effectively accommodating diverse healthcare beliefs, underscoring the need for tailored AI alignment in specific domains. This work highlights the limitations of current approaches and lays the groundwork for developing health-specific alignment solutions.
To facilitate robust and trustworthy deployment of large language models (LLMs), it is essential to quantify the reliability of their generations through uncertainty estimation. While recent efforts have made significant advancements by leveraging the internal logic and linguistic features of LLMs to estimate uncertainty scores, our empirical analysis highlights the pitfalls of these methods to strike a harmonized estimation between indication, balance, and calibration, which hinders their broader capability for accurate uncertainty estimation. To address this challenge, we propose CUE (Corrector for Uncertainty Estimation): A straightforward yet effective method that employs a lightweight model trained on data aligned with the target LLM’s performance to adjust uncertainty scores. Comprehensive experiments across diverse models and tasks demonstrate its effectiveness, which achieves consistent improvements of up to 60% over existing methods.
“What do you call a dog that is incontrovertibly true? Dogma”: Testing LLM Generalization through Humor
PDF ↗Humor, requiring creativity and contextual understanding, is a hallmark of human intelligence, showcasing adaptability across linguistic scenarios. While recent advances in large language models (LLMs) demonstrate strong reasoning on various benchmarks, it remains unclear whether they truly adapt to new tasks like humans (i.e., generalize) or merely replicate memorized content. To explore this, we introduce Phunny, a new humor-based question-answering benchmark designed to assess LLMs’ reasoning through carefully crafted puns. Our dataset is manually curated to ensure novelty and minimize data contamination, providing a robust evaluation of LLMs’ linguistic comprehension. Experiments on pun comprehension, resolution, and generation reveal that most LLMs struggle with generalization, even on simple tasks, consistently underperforming the human baseline. Additionally, our detailed error analysis provides valuable insights to guide future research.
Mining the uncertainty patterns of humans and models in the annotation of moral foundations and human values
PDF ↗The NLP community has converged on considering disagreement in annotation (or human label variation, HLV) as a constitutive feature of subjective tasks. This paper makes a further step by investigating the relationship between HLV and model uncertainty, and the impact of linguistic features of the items on both. We focus on the identification of moral foundations (e.g., care, fairness, loyalty) and human values (e.g., be polite, be honest) in text. We select three standard datasets and proceed into two steps. First, we focus on HLV and analyze the linguistic features (complexity, polarity, pragmatic phenomena, lexical choices) that correlate with HLV. Next, we proceed to uncertainty and its relationship to HLV. We experiment with RoBERTa and Flan-T5 in a number of training setups and evaluation metrics that test the calibration of uncertainty to HLV and its relationship to performance beyond majority vote; next, we analyze the impact of linguistic features on uncertainty. We find that RoBERTa with soft loss is better calibrated to HLV, and we find alignment between calibrated models and humans in the features (textual complexity and polarity) triggering variation.
Reasoning over time and space is essential for understanding our world. However, the abilities of language models in this area are largely unexplored as previous work has tested their abilities for logical reasoning in terms of time and space in isolation or only in simple or artificial environments. In this paper, we present the first evaluation of the ability of language models to jointly reason over time and space. To enable our analysis, we create GeoTemp, a dataset of 320k prompts covering 289 cities in 217 countries and 37 time zones. Using GeoTemp, we evaluate eight open chat models of three different model families for different combinations of temporal and geographic knowledge. We find that most models perform well on reasoning tasks involving only temporal knowledge and that overall performance improves with scale. However, performance remains constrained in tasks that require connecting temporal and geographical information. We do not find clear correlations of performance with specific geographic regions. Instead, we find a significant performance increase for location names with low model perplexity, suggesting their repeated occurrence during model training. We further demonstrate that their performance is heavily influenced by prompt formulation - a direct injection of geographical knowledge leads to performance gains, whereas, surprisingly, techniques like chain-of-thought prompting decrease performance on simpler tasks.
A Statistical and Multi-Perspective Revisiting of the Membership Inference Attack in Large Language Models
PDF ↗The lack of data transparency in Large Language Models (LLMs) has highlighted the importance of Membership Inference Attack (MIA), which differentiates trained (member) and untrained (non-member) data. Though it shows success in previous studies, recent research reported a near-random performance in different settings, highlighting a significant performance inconsistency. We assume that a single setting doesn’t represent the distribution of the vast corpora, causing members and non-members with different distributions to be sampled and causing inconsistency. In this study, instead of a single setting, we statistically revisit MIA methods from various settings with thousands of experiments for each MIA method, along with study in text feature, embedding, threshold decision, and decoding dynamics of members and non-members. We found that (1) MIA performance improves with model size and varies with domains, while most methods do not statistically outperform baselines, (2) Though MIA performance is generally low, a notable amount of differentiable member and non-member outliers exists and vary across MIA methods, (3) Deciding a threshold to separate members and non-members is an overlooked challenge, (4) Text dissimilarity and long text benefit MIA performance, (5) Differentiable or not is reflected in the LLM embedding, (6) Member and non-members show different decoding dynamics.
Recent advancements have demonstrated the advantage of converting pretrained large language models into powerful text encoders by enabling bidirectional attention in transformer layers. However, existing methods often require extensive training on large-scale datasets, posing challenges in low-resource, domain-specific scenarios. In this work, we show that a pretrained large language model can be converted into a strong text encoder without additional training. We first conduct a comprehensive empirical study to investigate different conversion strategies and identify the impact of the attention sink phenomenon on the performance of converted encoder models. Based on our findings, we propose a novel approach that enables bidirectional attention and suppresses the attention sink phenomenon, resulting in superior performance. Extensive experiments on multiple domains demonstrate the effectiveness of our approach. Our work provides new insights into the training-free conversion of text encoders in low-resource scenarios and contributes to the advancement of domain-specific text representation generation. Our code is available at https://github.com/bigai-nlco/Look-Both-Ways-and-No-Sink.
Being able to thoroughly assess massive multi-task language understanding (MMLU) capabilities is essential for advancing the applicability of multilingual language models. However, preparing such benchmarks in high quality native language is often costly and therefore limits the representativeness of evaluation datasets. While recent efforts focused on building more inclusive MMLU benchmarks, these are conventionally built using machine translation from high-resource languages, which may introduce errors and fail to account for the linguistic and cultural intricacies of the target languages. In this paper, we address the lack of native language MMLU benchmark especially in the under-represented Turkic language family with distinct morphosyntactic and cultural characteristics. We propose two benchmarks for Turkic language MMLU: TUMLU is a comprehensive, multilingual, and natively developed language understanding benchmark specifically designed for Turkic languages. It consists of middle- and high-school level questions spanning 11 academic subjects in Azerbaijani, Crimean Tatar, Karakalpak, Kazakh, Kyrgyz, Tatar, Turkish, Uyghur, and Uzbek. We also present TUMLU-mini, a more concise, balanced, and manually verified subset of the dataset. Using this dataset, we systematically evaluate a diverse range of open and proprietary multilingual large language models (LLMs), including Claude, Gemini, GPT, and LLaMA, offering an in-depth analysis of their performance across different languages, subjects, and alphabets. To promote further research and development in multilingual language understanding, we release TUMLU-mini and all corresponding evaluation scripts.
We propose EdiText, a controllable text editing method that modifies the reference text to desired attributes at various scales. We integrate an SDEdit-based editing technique that allows for broad adjustments in the degree of text editing. Additionally, we introduce a novel fine-level editing method based on self-conditioning, which allows subtle control of reference text. While being capable of editing on its own, this fine-grained method, integrated with the SDEdit approach, enables EdiText to make precise adjustments within the desired range. EdiText demonstrates its controllability to robustly adjust reference text at a broad range of levels across various tasks, including toxicity control and sentiment control.
To meet the demands of content moderation, online platforms have resorted to automated systems. Newer forms of real-time engagement (\textit{e.g.}, users commenting on live streams) on platforms like Twitch exert additional pressures on the latency expected of such moderation systems. Despite their prevalence, relatively little is known about the effectiveness of these systems. In this paper, we conduct an audit of Twitch’s automated moderation tool (\texttt{AutoMod}) to investigate its effectiveness in flagging hateful content. For our audit, we create streaming accounts to act as siloed test beds, and interface with the live chat using Twitch’s APIs to send over 107,000 comments collated from 4 datasets. We measure \texttt{AutoMod}‘s accuracy in flagging blatantly hateful content containing misogyny, racism, ableism and homophobia. Our experiments reveal that a large fraction of hateful messages, up to 94% on some datasets, \text{\textit{bypass moderation}}. Contextual addition of slurs to these messages results in 100% removal, revealing \texttt{AutoMod}‘s reliance on slurs as a hate signal. We also find that contrary to Twitch’s community guidelines, \texttt{AutoMod} blocks up to 89.5% of benign examples that use sensitive words in pedagogical or empowering contexts. Overall, our audit points to large gaps in \texttt{AutoMod}‘s capabilities and underscores the importance for such systems to understand context effectively.
CENTAUR: Bridging the Impossible Trinity of Privacy, Efficiency, and Performance in Privacy-Preserving Transformer Inference
PDF ↗With the growing deployment of pre-trained models like Transformers on cloud platforms, privacy concerns about model parameters and inference data are intensifying. Existing Privacy-Preserving Transformer Inference (PPTI) frameworks face the “impossible trinity” of balancing privacy, efficiency, and performance: Secure Multi-Party Computation (SMPC)-based approaches ensure strong privacy but suffer from high computational overhead and performance losses; Conversely, permutation-based methods achieve near-plaintext efficiency and accuracy but compromise privacy by exposing sensitive model parameters and intermediate results. Bridging this gap with a single approach presents substantial challenges, motivating the introduction of CENTAUR, a groundbreaking PPTI framework that seamlessly integrates random permutations and SMPC to address the “impossible trinity”. By designing efficient PPTI algorithms tailored to the structural properties of Transformer models, CENTAUR achieves an unprecedented balance among privacy, efficiency, and performance. Our experiments demonstrate CENTAUR’s ability to resist diverse data reconstruction attacks, achieve plaintext-level inference accuracy, and boost inference speed by 5.0~30.4 times, unlocking new possibilities for secure and efficient AI deployment.
Large language models (LLMs) generate human-like text, raising concerns about their misuse in creating deceptive content. Detecting LLM-generated comments (LGC) in online news is essential for preserving online discourse integrity and preventing opinion manipulation. However, effective detection faces two key challenges; the brevity and informality of news comments limit traditional methods, and the absence of a publicly available LGC dataset hinders model training, especially for languages other than English. To address these challenges, we propose a twofold approach. First, we develop an LGC generation framework to construct a high-quality dataset with diverse and complex examples. Second, we introduce XDAC (\textbf{X}AI-Driven \textbf{D}etection and \textbf{A}ttribution of LLM-Generated \textbf{C}omments), a framework utilizing explainable AI, designed for the detection and attribution of short-form LGC in Korean news articles. XDAC leverages XAI to uncover distinguishing linguistic patterns at both token and character levels. We present the first large-scale benchmark dataset, comprising 1.3M human-written comments from Korean news platforms and 1M LLM-generated comments from 14 distinct models. XDAC outperforms existing methods, achieving a 98.5% F1 score in LGC detection with a relative improvement of 68.1%, and an 84.3% F1 score in attribution. To validate real-world applicability, we analyze 5.24M news comments from Naver, South Korea’s leading online news platform, identifying 27,029 potential LLM-generated comments.
Critical text assessment is at the core of many expert activities, such as fact-checking, peer review, and essay grading. Yet, existing work treats critical text assessment as a black box problem, limiting interpretability and human-AI collaboration. To close this gap, we introduce Structured Reasoning in Critical Text Assessment (STRICTA), a novel specification framework to model text assessment as an explicit, step-wise reasoning process. STRICTA breaks down the assessment into a graph of interconnected reasoning steps drawing on causality theory (Pearl, 1995). This graph is populated based on expert interaction data and used to study the assessment process and facilitate human-AI collaboration. We formally define STRICTA and apply it in a study on biomedical paper assessment, resulting in a dataset of over 4000 reasoning steps from roughly 40 biomedical experts on more than 20 papers. We use this dataset to empirically study expert reasoning in critical text assessment, and investigate if LLMs are able to imitate and support experts within these workflows. The resulting tools and datasets pave the way for studying collaborative expert-AI reasoning in text assessment, in peer review and beyond.