Effective query expansion for web search benefits from promoting both exploration and result diversity to capture multiple interpretations and facets of a query. While recent LLM-based methods have improved retrieval performance and demonstrate strong domain generalization without additional training, they often generate narrowly focused expansions that overlook these desiderata. We propose ThinkQE, a test-time query expansion framework addressing this limitation through two key components: a thinking-based expansion process that encourages deeper and comprehensive semantic exploration, and a corpus-interaction strategy that iteratively refines expansions using retrieval feedback from the corpus. Experiments on diverse web search benchmarks (DL19, DL20, and BRIGHT) show ThinkQE consistently outperforms prior approaches, including training-intensive dense retrievers and rerankers.
论文检索
输入标题、作者或关键词，从 7,591 篇学术成果中精准定位
Semantic Component Analysis: Introducing Multi-Topic Distributions to Clustering-Based Topic Modeling
PDF ↗Topic modeling is a key method in text analysis, but existing approaches fail to efficiently scale to large datasets or are limited by assuming one topic per document. Overcoming these limitations, we introduce Semantic Component Analysis (SCA), a topic modeling technique that discovers multiple topics per sample by introducing a decomposition step to the clustering-based topic modeling framework. We evaluate SCA on Twitter datasets in English, Hausa and Chinese. There, it achieves competetive coherence and diversity compared to BERTopic, while uncovering at least double the topics and maintaining a noise rate close to zero. We also find that SCA outperforms the LLM-based TopicGPT in scenarios with similar compute budgets. SCA thus provides an effective and efficient approach for topic modeling of large datasets.
The influence of personas on Large Language Models (LLMs) has been widely studied, yet their direct impact on performance remains uncertain. This work explores a novel approach to guiding LLM behaviour through role vectors, an alternative to persona-based prompting. We construct 29 role vectors derived from model activations and evaluate their impact on benchmark performance across multiple domains. Our analysis investigates whether these vectors can effectively steer models toward domain-specific expertise. We measure two key interventions: (i) activation addition, which reinforces role-specific directions, and (ii) directional ablation, which removes them. Results on well-established benchmarks indicate that role vectors do, in fact, influence model behaviour, improving in-domain task performance while also yielding unexpected cross-domain gains.This, in turn, suggests that manipulating internal model representations has a greater impact on outcomes than persona-based prompting.
Position IDs Matter: An Enhanced Position Layout for Efficient Context Compression in Large Language Models
PDF ↗Using special tokens (e.g., gist, memory, or compressed tokens) to compress context information is a common practice for large language models (LLMs). However, existing approaches often neglect that position encodings inherently induce local inductive biases in models, causing the compression process to ignore holistic contextual dependencies. We propose **Enhanced Position Layout (EPL)**, a simple yet effective method that improves the context compression capability of LLMs by only adjusting position IDs, the numerical identifiers that specify token positions. EPL minimizes the distance between context tokens and their corresponding special tokens and at the same time maintains the sequence order in position IDs between context tokens, special tokens, and the subsequent tokens. Integrating EPL into our best performing context compression model results in 1.9 ROUGE-1 F1 improvement on out-of-domain question answering datasets in average. When extended to multimodal scenarios, EPL brings an average accuracy gain of 2.6 to vision compression LLMs.
ET-MIER: Entity Type-guided Key Mention Identification and Evidence Retrieval for Document-level Relation Extraction
PDF ↗Document-level relation extraction (DocRE) task aims to identify relations between entities in a document. In DocRE, an entity may appear in multiple sentences of a document in the form of mentions. In addition, relation inference requires the use of evidence sentences that can provide key clues to entity pairs. These make DocRE more challenging than sentencelevel relation extraction. Existing work does not fully distinguish the contribution of different mentions to entity representation and the importance of mentions in evidence sentences. To address these issues, we observe that entity types can provide consistent semantic constraints for entities of the same type and implicitly preclude impossible relations between entities, which may help the model better understand both intra- and inter-entity mentions. Therefore, we propose a novel model ET-MIER, which for the first time leverages **E**ntity **T**ypes to guide key **M**ention **I**dentification and **E**vidence **R**etrieval. In this way, entity types not only help learn better entity representation but also enhance evidence retrieval, both of which are crucial for DocRE. We conduct experiments on widely-adopted datasets and show that our model achieves state-of-the-art performance. Our code is available at: https://github.com/NEU-IDKE/ET-MIER
Beneath the Facade: Probing Safety Vulnerabilities in LLMs via Auto-Generated Jailbreak Prompts
PDF ↗The rapid proliferation of large language models and multimodal generative models has raised concerns about their potential vulnerabilities to a wide range of real-world safety risks. However, a critical gap persists in systematic assessment, alongside the lack of evaluation frameworks to keep pace with the breadth and variability of real-world risk factors. In this paper, we introduce TroGEN, an automated jailbreak prompt generation framework that assesses these vulnerabilities by deriving scenario-driven jailbreak prompts using an adversarial agent. Moving beyond labor-intensive dataset construction, TroGEN features an extensible design that covers broad range of risks, supports plug-and-play jailbreak strategies, and adapts seamlessly to multimodal settings. Experimental results demonstrate that TroGEN effectively uncovers safety weaknesses, revealing susceptibilities to adversarial attacks that conceal malicious intent beneath an apparently benign facade, like a Trojan horse. Furthermore, such stealthy attacks exhibit resilience even against existing jailbreak defense methods.
The Transformer architecture has long dominated the development of large language models, but its quadratic complexity in sequence length presents scalability challenges. Recent advances in State Space Models, particularly Mamba series, offer a promising alternative with linear-time inference and competitive performance. While scaling model capacity via sparsification, exemplified by Mixture-of-Experts, has proven effective in reducing computation while expanding knowledge capacity, the integration of sparsification with Mamba remains largely unexplored. Existing attempts typically apply naive block-level stacking, failing to leverage Mamba’s internal structure for fine-grained sparsification. In this work, we mainly explore how to sparsify the parameters inside Mamba. We found that the effects of using sparsification strategies on parameters related to various mechanisms inside mamba are significantly different. Our proposed Mamba-MoZ framework introduces a flexible and effective sparsification mechanism inside Mamba, which can independently achieve parameter scalability and has stronger performance.
SafeToolBench: Pioneering a Prospective Benchmark to Evaluating Tool Utilization Safety in LLMs
PDF ↗Large language models (LLMs) have exhibited great performance in autonomously calling various tools in external environments, leading to better problems solving and task automation capabilities. However, these external tools also amplify potential risks such as financial loss or privacy leaking with ambiguous or malicious user instructions. Compared to previous studies, which mainly assess the safety awareness of LLMs after obtaining the tool execution results (i.e., retrospective evaluation), this paper focuses on prospective ways to assess the safety of LLM tool utilization, aiming to avoid irreversible harm caused by directly executing tools. To this end, we propose SafeToolBench, the first benchmark to comprehensively assess tool utilization security in a prospective manner, covering malicious user instructions and diverse practical toolsets. Additionally, we propose a novel framework, SafeInstructTool, which aims to enhance LLMs’ awareness of tool utilization security through three perspectives (i.e., User Instruction, Tool Itself, and Joint Instruction-Tool), leading to nine detailed dimensions in total. We experiment with four LLMs using different methods, revealing that existing approaches fail to fully capture all risks in tool utilization. In contrast, our framework significantly enhances LLMs’ self-awareness, enabling a more safer and trustworthy tool utilization.
Large language models are reshaping internet services. Serving these models is often costly, as it requires multiple high-end GPUs. Consumer-grade GPUs offer cheaper computational power, providing an opportunity for more cost-efficient LLM serving.Prior efforts have explored distributed serving at scale, primarily focusing on model deployment strategies. However, communication efficiency has emerged as a challenge due to the imbalance in data transfer volumes between the two phases of inference: prefill and decode. Prefill requests can involve transmitting up to 1000 times more data than decode requests, leading to decode requests being delayed. Consequently, servers are underutilized while waiting for decode requests. In this paper, we present MoLink, an efficient distributed LLM serving system. It splits the prolonged transmission volume of prefill requests into smaller chunks and carefully scheduling their transmission. It consists of two parts: (i) a transmission scheduling algorithm that fairly determines whether to transmit prefill or decode requests, and (ii) a chunking determination algorithm that determines the transmit volume for prefill requests just-in-time. Our evaluation demonstrates that MoLink reduces TTFT, TPOT, and latency compared to the state-of-the-art distributed LLM serving system, with a maximum reduction of up to 46%.
Beyond Surface Alignment: Rebuilding LLMs Safety Mechanism via Probabilistically Ablating Refusal Direction
PDF ↗Jailbreak attacks pose persistent threats to large language models (LLMs). Current safety alignment methods have attempted to address these issues, but they experience two significant limitations: insufficient safety alignment depth and unrobust internal defense mechanisms. These limitations make them vulnerable to adversarial attacks such as prefilling and refusal direction manipulation. We introduce DeepRefusal, a robust safety alignment framework that overcomes these issues. DeepRefusal forces the model to dynamically rebuild its refusal mechanisms from jailbreak states. This is achieved by probabilistically ablating the refusal direction across layers and token depths during fine-tuning. Our method not only defends against prefilling and refusal direction attacks but also demonstrates strong resilience against other unseen jailbreak strategies. Extensive evaluations on four open-source LLM families and six representative attacks show that DeepRefusal reduces attack success rates by approximately 95%, while maintaining model capabilities with minimal performance degradation.
This work fosters research on the interaction between natural language use and gambling disorders. We have built a new Spanish corpus for screening standardized gambling symptoms. We employ search methods to find on-topic sentences, top-k pooling to form the assessment pools of sentences, and thorough annotation guidelines. The labeling task is challenging, given the need to identify topic relevance and explicit evidence about the symptoms. Additionally, we explore using state-of-the-art LLMs for annotation and compare different sentence search models.
Commonsense reasoning in multimodal contexts remains a foundational challenge in artificial intelligence. We introduce Multimodal UNcommonsense (MUN), a benchmark designed to evaluate models’ ability to handle scenarios that deviate from typical visual or contextual expectations. MUN pairs visual scenes with surprising or unlikely outcomes described in natural language, prompting models to either rationalize seemingly odd images using everyday logic or uncover unexpected interpretations in ordinary scenes. To support this task, we propose a retrieval-based in-context learning (R-ICL) framework that transfers reasoning capabilities from larger models to smaller ones without additional training. Leveraging a novel Multimodal Ensemble Retriever (MER), our method identifies semantically relevant exemplars even when image and text pairs are deliberately discordant. Experiments show an average improvement of 8.3% over baseline ICL methods, highlighting the effectiveness of R-ICL in low-frequency, atypical settings. MUN opens new directions for evaluating and improving visual-language models’ robustness and adaptability in real-world, culturally diverse, and non-prototypical scenarios.
Vision-language models like CLIP demonstrate exceptional generalization capabilities but face significant adaptation challenges due to parameter scale, prompt sensitivity, and cross-modal alignment difficulties. Existing approaches primarily focus on single-modality adjustments, leading to suboptimal alignment and limited generalization. We introduce MAFMO, a plug-and-play framework comprising: (1) a Harmonic Cross-Modal Adapter enabling efficient cross-modal knowledge transfer; (2) a Meta-Template Optimization module dynamically generating input-dependent templates; and (3) a Cross-Modal Knowledge Synthesis mechanism preserving critical structural relationships during adaptation. Extensive experiments across multiple fine-grained visual recognition benchmarks demonstrate MAFMO consistently improves existing methods’ performance on both novel classes and harmonic mean, while maintaining robustness under various challenging conditions with minimal computational overhead.
We critically assess the efficacy of the current SOTA in hallucination detection and find that its performance on the RAGTruth dataset is largely driven by a spurious correlation with data. Controlling for this effect, state-of-the-art performs no better than supervised linear probes, while requiring extensive hyperparameter tuning across datasets. Out-of-distribution generalization is currently out of reach, with all of the analyzed methods performing close to random. We propose a set of guidelines for hallucination detection and its evaluation.
Beyond Binary Preferences: Semi-Online Label-Free GRACE-KTO with Group-Wise Adaptive Calibration for High-Quality Long-Text Generation
PDF ↗Generating high-quality long-text remains challenging for Large Language Models (LLMs), as conventional supervised fine-tuning fails to ensure overall quality due to its teacher-forcing nature. Kahneman-Tversky Optimization (KTO), as a model alignment method that can holistically optimize generation quality, overcomes the need for paired preference data required by previous methods. However, it still suffers from binary supervision that inadequately reflects varying quality degrees. To address this, we propose GRACE-KTO, a semi-online framework that transforms KTO’s binary signals into dynamically calibrated intra-group rewards. Specifically, GRACE-KTO aggregates responses to identical queries into groups, computes rank-sum scores across multiple linguistic quality dimensions, and applies group-wise and global normalization to adaptively redistribute sample importance. We adopt a semi-online training strategy to reduce costly online sampling while outperforming offline variants. By leveraging query generation with seed data, we minimize labeled data dependency, using the model’s own knowledge to enhance its long-text generation capabilities. Additionally, we extend the context window to 32k tokens using YaRN during inference, enabling the model to generate longer texts while maintaining perplexities. Experiments demonstrate GRACE-KTO’s superiority over vanilla KTO on both automatic metrics and LLM-as-a-Judge evaluations, advancing long-text generation through group-wise adaptive calibration.
In this paper we introduce a Hindi verb alternations benchmark to investigate whether pretrained large language models (LLMs) can infer the frame-selectional properties of Hindi verbs. Our benchmark consists of minimal pairs such as ‘Tina cut the wood’/*‘Tina disappeared the wood’. We create four variants of these alternations for Hindi to test knowledge of verbal morphology and argument case-marking. Our results show that a masked monolingual model performs the best, while causal models fare poorly. We further test the quality of the predictions using a cloze-style sentence completion task. While the models appear to infer the right mapping between verbal morphology and valency in the acceptability task, they do not generate the right verbal morphology in the cloze task. The model completions also lack pragmatic and world knowledge, crucial for making generalizations about verbal alternations. Our work points towards the need for more cross-linguistic research of verbal alternations.
Recent advancements in large language models (LLMs) have enabled their successful application to a broad range of tasks. However, in information-intensive tasks, the prompt length can grow fast, leading to increased computational requirements, performance degradation, and induced biases from irrelevant or redundant information. Recently, various prompt compression techniques have been introduced to optimize the trade-off between reducing input length and retaining performance. We propose a holistic evaluation framework that allows for in-depth analysis of prompt compression methods. We focus on three key aspects, besides compression ratio: (i) downstream task performance, (ii) grounding in the input context, and (iii) information preservation. Using our framework, we analyze state-of-the-art soft and hard compression methods and show that some fail to preserve key details from the original prompt, limiting performance on complex tasks. By identifying these limitations, we are able to improve one soft prompting method by controlling compression granularity, achieving up to +23% in downstream performance, +8 BERTScore points in grounding, and 2.7× more entities preserved in compression. Ultimately, we find that the best effectiveness/compression rate trade-off is achieved with soft prompting combined with sequence-level training.
Domain adaptation is widely adopted in text retrieval scenarios where large labeled data is unavailable. To improve model adaptability, existing methods try to expand more source datasets. However, we found from experiments that indiscriminately using a large amount of source data from various text tasks does not guarantee improved adaptability, but may negatively impact model performance. To tackle this issue, we propose Trait, a framework that can effectively improve model adaptability by selecting beneficial data without evaluating all source data. Specifically, we first divide multiple source datasets into data chunks of the same size as the minimum selection unit to form the whole selection space. Then we devise an iterative process that includes Bayesian optimization-based selection and transfer-aware chunk evaluation to incrementally select beneficial chunks. To reduce unnecessary evaluation costs, we also design backtracking and pruning actions to adjust the selection subspace. Extensive experimental results show that Trait not only achieves average state-of-the-art for few-shot on nine target datasets by evaluating only 4% of BERRI source data, but also is very competitive for zero-shot compared with LLM-based rankers.
Accept or Deny? Evaluating LLM Fairness and Performance in Loan Approval across Table-to-Text Serialization Approaches
PDF ↗Large Language Models (LLMs) are increasingly employed in high-stakes decision-making tasks, such as loan approvals. While their applications expand across domains, LLMs struggle to process tabular data, ensuring fairness and delivering reliable predictions. In this work, we assess the performance and fairness of LLMs on serialized loan approval datasets from three geographically distinct regions: Ghana, Germany, and the United States. Our evaluation focuses on the model’s zero-shot and in-context learning (ICL) capabilities. Our results reveal that the choice of serialization format significantly affects both performance and fairness in LLMs, with certain formats such as GReaT and LIFT yielding higher F1 scores but exacerbating fairness disparities. Notably, while ICL improved model performance by 4.9-59.6% relative to zero-shot baselines, its effect on fairness varied considerably across datasets. Our work underscores the importance of effective tabular data representation methods and fairness-aware models to improve the reliability of LLMs in financial decision-making.
A large body of research has found substantial gender bias in NLP systems. Most of this research takes a binary, essentialist view of gender: limiting its variation to the categories _men_ and _women_, conflating gender with sex, and ignoring different sexual identities. But gender and sexuality exist on a spectrum, so in this paper we study the biases of large language models (LLMs) towards sexual and gender minorities beyond binary categories. Grounding our study in a widely used social psychology model—the Stereotype Content Model—we demonstrate that English-language survey questions about social perceptions elicit more negative stereotypes of sexual and gender minorities from both humans and LLMs. We then extend this framework to a more realistic use case: text generation. Our analysis shows that LLMs generate stereotyped representations of sexual and gender minorities in this setting, showing that they amplify representational harms in creative writing, a widely advertised use for LLMs.