Although large language models (LLMs) have achieved remarkable advancements, their security remains a pressing concern. One major threat is jailbreak attacks, where adversarial prompts bypass model safeguards to generate harmful or objectionable content. Researchers study jailbreak attacks to understand security and robustness of LLMs. However, existing jailbreak attack methods face two main challenges: (1) an excessive number of iterative queries, and (2) poor generalization across models. In addition, recent jailbreak evaluation datasets focus primarily on question-answering scenarios, lacking attention to text generation tasks that require accurate regeneration of toxic content.To tackle these challenges, we propose two contributions:(1) **ICE**, a novel black-box jailbreak method that employs **I**ntent **C**oncealment and div**E**rsion to effectively circumvent security constraints. **ICE** achieves high attack success rates (ASR) with a single query, significantly improving efficiency and transferability across different models.(2) **BiSceneEval**, a comprehensive dataset designed for assessing LLM robustness in question-answering and text-generation tasks. Experimental results demonstrate that **ICE** outperforms existing jailbreak techniques, revealing critical vulnerabilities in current defense mechanisms. Our findings underscore the necessity of a hybrid security strategy that integrates predefined security mechanisms with real-time semantic decomposition to enhance the security of LLMs.
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
Dialogue Topic Segmentation (DTS) aims to divide dialogues into coherent segments. DTS plays a crucial role in various NLP downstream tasks, but suffers from chronic problems: data shortage, labeling ambiguity, and incremental complexity of recently proposed solutions. On the other hand, Despite advances in Large Language Models (LLMs) and reasoning strategies, these have rarely been applied to DTS. This paper introduces Def-DTS: Deductive Reasoning for Open-domain Dialogue Topic Segmentation, which utilizes LLM-based multi-step deductive reasoning to enhance DTS performance and enable case study using intermediate result. Our method employs a structured prompting approach for bidirectional context summarization, utterance intent classification, and deductive topic shift detection. In the intent classification process, we propose the generalizable intent list for domain-agnostic dialogue intent classification. Experiments in various dialogue settings demonstrate that Def-DTS consistently outperforms traditional and state-of-the-art approaches, with each subtask contributing to improved performance, particularly in reducing type 2 error. We also explore the potential for autolabeling, emphasizing the importance of LLM reasoning techniques in DTS.
MMS-LLaMA: Efficient LLM-based Audio-Visual Speech Recognition with Minimal Multimodal Speech Tokens
PDF ↗Audio-Visual Speech Recognition (AVSR) achieves robust speech recognition in noisy environments by combining auditory and visual information. However, recent Large Language Model (LLM) based AVSR systems incur high computational costs due to the high temporal resolution of audio-visual speech processed by LLMs. In this work, we introduce an efficient multimodal speech LLM framework that minimizes token length while preserving essential linguistic content. Our approach employs an early AV-fusion module for streamlined feature integration, an audio-visual speech Q-Former that dynamically allocates tokens based on input duration, and a refined query allocation strategy with a speech rate predictor to adjust token allocation according to speaking speed of each audio sample. Extensive experiments on the LRS3 dataset show that our method achieves state-of-the-art performance with a WER of 0.72% while using only 3.5 tokens per second. Moreover, our approach not only reduces token usage by 86% compared to the previous multimodal speech LLM framework, but also improves computational efficiency by reducing FLOPs by 35.7%. The code and models are available https://github.com/JeongHun0716/MMS-LLaMA.
Instruction following (IF) is a critical capability for large language models (LLMs). However, handling complex instructions with multiple constraints remains challenging. Previous methods typically select preference pairs based on the number of constraints they satisfy, introducing noise where chosen examples may fail to follow some constraints and rejected examples may excel in certain respects over the chosen ones. To address the challenge of aligning with multiple preferences, we propose a simple yet effective method called Reverse Preference Optimization (RPO). It mitigates noise in preference pairs by dynamically reversing the constraints within the instruction to ensure the chosen response is perfect, alleviating the burden of extensive sampling and filtering to collect perfect responses. Besides, reversal also enlarges the gap between chosen and rejected responses, thereby clarifying the optimization direction and making it more robust to noise. We evaluate RPO on two multi-turn IF benchmarks, Sysbench and Multi-IF, demonstrating average improvements over the DPO baseline of 4.6 and 2.5 points (on Llama-3.1 8B), respectively. Moreover, RPO scales effectively across model sizes (8B to 70B parameters), with the 70B RPO model surpassing GPT-4o.
While integrating external tools into large language models (LLMs) enhances their ability to access real-time information and domain-specific services, existing approaches focus narrowly on functional tool selection following user instructions while overlooking the critical role of context-aware personalization in tool selection. This oversight leads to suboptimal user satisfaction and inefficient tool utilization, particularly when overlapping toolsets require nuanced selection based on contextual factors. To bridge this gap, we introduce ToolSpectrum, a benchmark designed to evaluate LLMs’ capabilities in personalized tool utilization. Specifically, we formalize two key dimensions of personalization, user profile and environmental factors, and analyze their individual and synergistic impacts on tool selection. Through extensive experiments on ToolSpectrum, we demonstrate that personalized tool selection significantly improves user experience across diverse scenarios. However, even state-of-the-art LLMs exhibit the limited ability to reason jointly about user profiles and environmental factors, often prioritizing one dimension at the expense of the other. Our findings underscore the necessity of context-aware personalization in tool-augmented LLMs and reveal critical limitations for current models. Our data and code will be released soon.
Uncertainty Unveiled: Can Exposure to More In-context Examples Mitigate Uncertainty for Large Language Models?
PDF ↗Recent advances in handling long sequences have unlocked new possibilities for long-context in-context learning (ICL). While existing research predominantly focuses on performance gains driven by additional in-context examples, the impact on the trustworthiness of generated responses remains underexplored. This paper addresses this gap by investigating how increased examples influence predictive uncertainty—an essential aspect in trustworthiness. We begin by systematically quantifying uncertainty across different “shot” configurations in ICL, emphasizing the role of example quantity. Through uncertainty decomposition, we introduce a novel perspective on performance enhancement, focusing on epistemic uncertainty (EU). Our results reveal that additional examples reduce total uncertainty in both simple and complex tasks by injecting task-specific knowledge, thereby diminishing EU and enhancing performance. For complex tasks, these advantages emerge only after addressing the increased noise and uncertainty associated with longer inputs. Finally, we investigate the progression of internal confidence across layers, uncovering the underlying mechanisms that drive the reduction in uncertainty.
Recent advancements in large language models (LLMs) have significantly enhanced their knowledge and generative capabilities, leading to a surge of interest in leveraging LLMs for high-quality data synthesis. However, synthetic data generation via prompting LLMs remains challenging due to LLMs’ limited understanding of target data distributions and the complexity of prompt engineering, especially for structured formatted data. To address these issues, we introduce DiffLM, a controllable data synthesis framework based on variational autoencoder (VAE), which further (1) leverages diffusion models to reserve more information of original distribution and format structure in the learned latent distribution and (2) decouples the learning of target distribution knowledge from the LLM’s generative objectives via a plug-and-play latent feature injection module. As we observed significant discrepancies between the VAE’s latent representations and the real data distribution, the latent diffusion module is introduced into our framework to learn a fully expressive latent distribution. Evaluations on seven real-world datasets with structured formatted data (i.e., Tabular, Code, and Tool data) demonstrate that DiffLM generates high-quality data, with performance on downstream tasks surpassing that of real data by 2%–7% in certain cases. Data and code are available at https://github.com/bytedance/DiffLM.
This paper presents elaborate testing of various LLMs on their generalization capacities. We finetune six encoder models that have been pretrained with very different data (varying in size, language, and period) on a challenging event detection task in Early Modern Dutch archival texts. Each model is finetuned with 5 seeds on 15 different data splits, resulting in 450 finetuned models. We also pre-train a domain-specific Language Model on the target domain and fine-tune and evaluate it in the same way to provide an upper bound. Our experimental setup allows us to look at underresearched aspects of generalizability, namely i) shifts at multiple places in a modeling pipeline, ii) temporal and crosslingual shifts and iii) generalization over different initializations. The results show that none of the models reaches domain-specific model performance, demonstrating their incapacity to generalize. mBERT reaches highest F1 performance, and is relatively stable over different seeds and datasplits, contrary to XLM-R. We find that contemporary Dutch models do not generalize well to Early Modern Dutch as they underperform compared to crosslingual as well as historical models. We conclude that encoder LLMs lack temporal generalization capacities and that bigger models are not better, since even a model pre-trained with five hundred GPUs on 2.5 terabytes of training data (XLM-R) underperforms considerably compared to our domain-specific model, pre-trained on one GPU and 6 GB of data. All our code, data, and the domain-specific model are openly available.
Assessing the Reasoning Capabilities of LLMs in the context of Evidence-based Claim Verification
PDF ↗Although LLMs have shown great performance on Mathematics and Coding related reasoning tasks, the reasoning capabilities of LLMs regarding other forms of reasoning are still an open problem. Here, we examine the issue of reasoning from the perspective of claim verification. We propose a framework designed to break down any claim paired with evidence into atomic reasoning types that are necessary for verification. We use this framework to create RECV, the first claim verification benchmark, incorporating real-world claims, to assess the deductive and abductive reasoning capabilities of LLMs. The benchmark comprises of three datasets, covering reasoning problems of in creasing complexity. We evaluate three state of-the-art proprietary LLMs under multiple prompt settings. Our results show that while LLMs can address deductive reasoning prob lems, they consistently fail in cases of abductive reasoning. Moreover, we observe that enhancing LLMs with rationale generation is not always beneficial. Nonetheless, we find that generated rationales are semantically similar to those provided by humans, especially in deduc tive reasoning cases.
Machine Unlearning (MU) is critical for removing private or hazardous information from deep learning models. While MU has advanced significantly in unimodal (text or vision) settings, multimodal unlearning (MMU) remains underexplored due to the lack of open benchmarks for evaluating cross-modal data removal. To address this gap, we introduce CLEAR, the first open-source benchmark designed specifically for MMU. CLEAR contains 200 fictitious individuals and 3,700 images linked with corresponding question-answer pairs, enabling a thorough evaluation across modalities. We conduct a comprehensive analysis of 11 MU methods (e.g., SCRUB, gradient ascent, DPO) across four evaluation sets, demonstrating that jointly unlearning both modalities outperforms single-modality approaches. The dataset is available at [link](https://huggingface.co/datasets/therem/CLEAR)
MWPO: Enhancing LLMs Performance through Multi-Weight Preference Strength and Length Optimization
PDF ↗Direct Preference Optimization (DPO) have proposed offline alternatives to Reinforcement Learning from Human Feedback (RLHF). In DPO, each preference pair, which serves as the foundation for learning, is typically constructed by first generating multiple responses to the same instruction and then annotating them to indicate the preferred choice. However, when the responses are highly similar, the weak preference signal can introduce annotation noise, which may hinder model optimization. Additionally, DPO suffers from the drawback of over-optimizing for verbose generation. A potential reason is the presence of length bias in preference datasets, which can lead to length exploitation. To address these issues, we propose a DPO-based **m**ulti-**w**eight **p**reference strength and length **o**ptimization (MWPO) method. Specifically, we propose to reweight preference pairs based on implicit reward margins and response length margins, unifying them through a geometric mixture to generate synthetic weights for optimization. This method allows preference pairs with stronger preference signals or more favorable length feature to have a more pronounced impact on model parameters. Moreover, our method does not require additional annotators. We validate our method on models of four different scales across multiple benchmarks. Our method surpasses state-of-the-art (SOTA) baselines, outperforming DPO by up to 8.7% on AlpacaEval 2 while reducing generation length by 9.4% in the Mistral setting. Our code is available at https://github.com/AIR-hl/MWPO.
A Multi-Expert Structural-Semantic Hybrid Framework for Unveiling Historical Patterns in Temporal Knowledge Graphs
PDF ↗Temporal knowledge graph reasoning aims to predict future events with knowledge of existing facts and plays a key role in various downstream tasks. Previous methods focused on either graph structure learning or semantic reasoning, failing to integrate dual reasoning perspectives to handle different prediction scenarios. Moreover, they lack the capability to capture the inherent differences between historical and non-historical events, which limits their generalization across different temporal contexts. To this end, we propose a **M**ulti-**E**xpert **S**tructural-**S**emantic **H**ybrid (MESH) framework that employs three kinds of expert modules to integrate both structural and semantic information, guiding the reasoning process for different events. Extensive experiments on three datasets demonstrate the effectiveness of our approach.
Large Language Models (LLMs) face computational inefficiencies and redundant processing when handling long context inputs, prompting a focus on compression techniques. While existing semantic vector-based compression methods achieve promising performance, these methods fail to account for the intrinsic information density variations between context chunks, instead allocating soft tokens uniformly across context chunks. This uniform distribution inevitably diminishes allocation to information-critical regions. To address this, we propose Dynamic Allocation of Soft Tokens (DAST), a simple yet effective method that leverages the LLM’s intrinsic understanding of contextual relevance to guide compression. DAST combines perplexity-based local information with attention-driven global information to dynamically allocate soft tokens to the informative-rich chunks, enabling effective, context-aware compression. Experimental results across multiple benchmarks demonstrate that DAST surpasses state-of-the-art methods.
Research on holistic Automated Essay Scoring (AES) is long-dated; yet, there is a notable lack of attention for assessing essays according to individual traits. In this work, we propose TRATES, a novel trait-specific and rubric-based cross-prompt AES framework that is generic yet specific to the underlying trait. The framework leverages a Large Language Model (LLM) that utilizes the trait grading rubrics to generate trait-specific features (represented by assessment questions), then assesses those features given an essay. The trait-specific features are eventually combined with generic writing-quality and prompt-specific features to train a simple classical regression model that predicts trait scores of essays from an unseen prompt. Experiments show that TRATES achieves a new state-of-the-art performance across all traits on a widely-used dataset, with the generated LLM-based features being the most significant.
MELABenchv1: Benchmarking Large Language Models against Smaller Fine-Tuned Models for Low-Resource Maltese NLP
PDF ↗Large Language Models (LLMs) have demonstrated remarkable performance across various Natural Language Processing (NLP) tasks, largely due to their generalisability and ability to perform tasks without additional training. However, their effectiveness for low-resource languages remains limited. In this study, we evaluate the performance of 55 publicly available LLMs on Maltese, a low-resource language, using a newly introduced benchmark covering 11 discriminative and generative tasks. Our experiments highlight that many models perform poorly, particularly on generative tasks, and that smaller fine-tuned models often perform better across all tasks. From our multidimensional analysis, we investigate various factors impacting performance. We conclude that prior exposure to Maltese during pre-training and instruction-tuning emerges as the most important factor. We also examine the trade-offs between fine-tuning and prompting, highlighting that while fine-tuning requires a higher initial cost, it yields better performance and lower inference costs. Through this work, we aim to highlight the need for more inclusive language technologies and recommend for researchers working with low-resource languages to consider more “traditional” language modelling approaches.
In team communication, dialogue acts play a crucial role in helping team members understand each other’s intentions and revealing the roles and communication patterns within interactions. Although existing studies have focused on using Dialogue Act classification to capture the speaker’s intentions, few have explored the underlying power dynamics reflected by these dialogue acts. To this end, we present an online Dialogue Act Classification and Dynamic Power Analysis framework—Act2P, which is based on large language model. The framework combines the zero-shot learning capability of LLMs and introduces an online feedback classification method that allows for online classification with iterative feedback to previous stages, achieving efficient and accurate classification without the labeled data. Additionally, we also propose the PowerRank algorithm, which quantifies power dynamics through a graph-based structure. Through comparative experiments with existing methods, we demonstrate the significant superiority of Act2P in online scenarios and successfully visualize dialogue power in online, clearly presenting the distribution and dynamic transfer of power. This framework provides new scientific insights and practical tools for optimizing team collaboration.
Recent advancements in speech-to-speech dialogue systems leverage LLMs for multimodal interactions, yet they remain hindered by fine-tuning requirements, high computational overhead, and text-speech misalignment. Existing speech-enabled LLMs often degrade conversational quality by modifying the LLM, thereby compromising its linguistic capabilities. In contrast, we propose LLMVoX, a lightweight 30M-parameter, LLM-agnostic, autoregressive streaming TTS system that generates high-quality speech with low latency, while fully preserving the capabilities of the base LLM. Our approach achieves a significantly lower Word Error Rate compared to speech-enabled LLMs, while operating at comparable latency. By decoupling speech synthesis from LLM processing via a multi-queue token streaming system, LLMVoX enables seamless, infinite-length dialogues. Its plug-and-play design also facilitates extension to various tasks with different backbones. Furthermore, LLMVoX generalizes to new languages with minimal dataset adaptation, attaining a low Character Error Rate on an Arabic speech task. Evaluations demonstrate that LLMVoX matches or surpasses existing speech-enabled LLMs in both speech quality and latency, while maintaining the original linguistic strengths of the LLM. Additionally, we have integrated LLMVoX with a Vision-Language Model to create an omni-model with speech, text, and vision capabilities, without requiring additional multimodal training.
EMGLLM: Data-to-Text Alignment for Electromyogram Diagnosis Generation with Medical Numerical Data Encoding
PDF ↗Electromyography (EMG) tables are crucial for diagnosing muscle and nerve disorders, and advancing the automation of EMG diagnostics is significant for improving medical efficiency. EMG tables contain extensive continuous numerical data, which current Large Language Models (LLMs) often struggle to interpret effectively. To address this issue, we propose EMGLLM, a data-to-text model specifically designed for medical examination tables. EMGLLM employs the EMG Alignment Encoder to simulate the process that doctors compare test values with reference values, aligning the data into word embeddings that reflect health degree. Additionally, we construct ETM, a dataset comprising 17,250 real cases and their corresponding diagnostic results, to support medical data-to-text tasks. Experimental results on ETM demonstrate that EMGLLM outperforms various baseline models in understanding EMG tables and generating high-quality diagnoses, which represents an effective paradigm for automatic diagnosis generation from medical examination table.
Long-form question answering (LFQA) aims to provide thorough and in-depth answers to complex questions, enhancing comprehension. However, such detailed responses are prone to hallucinations and factual inconsistencies, challenging their faithful evaluation. This work introduces HaluQuestQA, the first hallucination dataset with localized error annotations for human-written and model-generated LFQA answers. HaluQuestQA comprises 698 QA pairs with 1.8k span-level error annotations for five different error types by expert annotators, along with preference judgments. Using our collected data, we thoroughly analyze the shortcomings of long-form answers and find that they lack comprehensiveness and provide unhelpful references. We train an automatic feedback model on this dataset that predicts error spans with incomplete information and provides associated explanations. Finally, we propose a prompt-based approach, Error-Informed Refinement, that uses signals from the learned feedback model to refine generated answers, which we show reduces errors and improves the quality of the answers across multiple models. Furthermore, humans find the answers generated by our approach comprehensive and highly prefer them (84%) over the baseline answers.
SynapticRAG: Enhancing Temporal Memory Retrieval in Large Language Models through Synaptic Mechanisms
PDF ↗Existing retrieval methods in Large Language Models show degradation in accuracy when handling temporally distributed conversations, primarily due to their reliance on simple similarity-based retrieval. Unlike existing memory retrieval methods that rely solely on semantic similarity, we propose SynapticRAG, which uniquely combines temporal association triggers with biologically-inspired synaptic propagation mechanisms. Our approach uses temporal association triggers and synaptic-like stimulus propagation to identify relevant dialogue histories. A dynamic leaky integrate-and-fire mechanism then selects the most contextually appropriate memories. Experiments on four datasets of English, Chinese and Japanese show that compared to state-of-the-art memory retrieval methods, SynapticRAG achieves consistent improvements across multiple metrics up to 14.66% points. This work bridges the gap between cognitive science and language model development, providing a new framework for memory management in conversational systems.