论文检索

Radu Marinescu, Debarun Bhattacharjya, Junkyu Lee, Tigran T. Tchrakian, Javier Carnerero-Cano, Yufang Hou, Elizabeth M. Daly, Alessandra Pascale

Large language models (LLMs) have achieved remarkable success in generative tasks, yet they often fall short in ensuring the factual accuracy of their outputs thus limiting their reliability in real-world applications where correctness is critical. In this paper, we present FactReasoner, a novel neuro-symbolic based factuality assessment framework that employs probabilistic reasoning to evaluate the truthfulness of long-form generated responses. FactReasoner decomposes a response into atomic units, retrieves relevant contextual information from external knowledge sources, and models the logical relationships (e.g., entailment, contradiction) between these units and their contexts using probabilistic encodings. It then estimates the posterior probability that each atomic unit is supported by the retrieved evidence. Our experiments on both labeled and unlabeled benchmark datasets demonstrate that FactReasoner often outperforms state-of-the-art prompt-based methods in terms of factual precision and recall.

Taowen Liu, Marta Andronic, Deniz Gunduz, George Anthony Constantinides

LLM training is resource-intensive. Quantized training improves computational and memory efficiency but introduces quantization noise, which can hinder convergence and degrade model accuracy. Stochastic Rounding (SR) has emerged as a theoretically attractive alternative to deterministic rounding, offering unbiased gradient estimates. However, its interaction with other training factors—especially batch size—remains underexplored. In this paper, we present a theoretical and empirical study of mini-batch stochastic gradient descent (SGD) with SR, showing that increased batch sizes can compensate for reduced precision during backpropagation. Furthermore, we show that quantizing weights and activations impacts gradient variance in distinct ways. Our experiments validate these theoretical insights. Our experiments validate these theoretical insights.

Rem Hida, Masahiro Kaneko, Naoaki Okazaki

Warning: This paper contains examples of stereotypes and biases. Large Language Models (LLMs) exhibit considerable social biases, and various studies have tried to evaluate and mitigate these biases accurately. Previous studies use downstream tasks to examine the degree of social biases for evaluation and mitigation. While the output of LLMs highly depends on prompts, prior works evaluating and mitigating bias have often relied on a limited variety of prompts. In this paper, we investigate the sensitivity of LLMs when changing prompt variations (task instruction, few-shot examples, debias-prompt) by analyzing task performance and social bias of LLMs. Our experimental results reveal that LLM rankings fluctuate across prompts for both task performance and social bias. We also confirmed that the impact of format changes can differ for each bias category. Performance improvement from prompt settings may not result in reduced bias. Moreover, the ambiguity of instances is a common factor in LLM sensitivity to prompts across advanced LLMs. We recommend using diverse prompts, as in this study, to compare the effects of prompts on social bias in LLMs.

Hanghai Hong, Yibo Xie, Jiawei Zheng, Xiaoli Wang

Large language models (LLMs) have recently achieved remarkable progress in sentence-level machine translation, but scaling to document-level machine translation (DocMT) remains challenging, particularly in modeling long-range dependencies and discourse phenomena across sentences and paragraphs. Document translations generated by LLMs often suffer from poor consistency, weak coherence, and omission errors. To address these issues, we propose SubDocTrans, a novel DocMT framework that enables LLMs to produce high-quality translations through plug-and-play, multi-granularity knowledge extraction and integration. SubDocTrans first performs topic segmentation to divide a document into coherent topic sub-documents. For each sub-document, both global and local knowledge are extracted including bilingual summary, theme, proper nouns, topics, and transition hint. We then incorporate this multi-granularity knowledge into the prompting strategy, to guide LLMs in producing consistent, coherent, and accurate translations. We conduct extensive experiments across various DocMT tasks, and the results demonstrate the effectiveness of our framework, particularly in improving consistency and coherence, reducing omission errors, and mitigating hallucinations.

Guimin Hu, Daniel Hershcovich, Hasti Seifi

Haptic signals, from smartphone vibrations to virtual reality touch feedback, can effectively convey information and enhance realism, but designing signals that resonate meaningfully with users is challenging. To facilitate this, we introduce a multimodal dataset and task, of matching user descriptions to vibration haptic signals, and highlight two primary challenges: (1) lack of large haptic vibration datasets annotated with textual descriptions as collecting haptic descriptions is time-consuming, and (2) limited capability of existing tasks and models to describe vibration signals in text.To advance this area, we create HapticCap, the first fully human-annotated haptic-captioned dataset, containing 92,070 haptic-text pairs for user descriptions of sensory, emotional, and associative attributes of vibrations. Based on HapticCap, we propose the haptic-caption retrieval task and present the results of this task from a supervised contrastive learning framework that brings together text representations within specific categories and vibrations. Overall, the combination of language model T5 and audio model AST yields the best performance in the haptic-caption retrieval task, especially when separately trained for each description category. The dataset is available at https://huggingface.co/datasets/GuiminHu/HapticCap.

Sike Xiang, Shuang Chen, Amir Atapour-Abarghouei

As multimodal large language models (MLLMs) advance, their large-scale architectures pose challenges for deployment in resource-constrained environments. In the age of large models, where energy efficiency, computational scalability and environmental sustainability are paramount, the development of lightweight and high-performance models is critical for real-world applications. As such, we propose a lightweight MLLM framework for end-to-end visual question answering. Our proposed approach centres on BreezeCLIP, a compact yet powerful vision-language encoder optimised for efficient multimodal understanding. With only 1.2 billion parameters overall, our model significantly reduces computational cost while achieving performance comparable to standard-size MLLMs. Experiments conducted on multiple datasets further validate its effectiveness in balancing accuracy and efficiency. The modular and extensible design enables generalisation to broader multimodal tasks. The proposed lightweight vision-language framework is denoted as BcQLM (BreezeCLIP-enhanced Q-Gated Multimodal Language Model). It offers a promising path toward deployable MLLMs under practical hardware constraints. The source code is available at https://github.com/thico0224/BcQLM.

Nicola Arici, Luca Putelli, Ejdis Gjinika, Ivan Serina, Alfonso Emilio Gerevini

Figurative language and figures of speech, such as metaphors and hyperboles, are used every day in written and oral communication among human beings. Nonetheless, this imaginative use of words in a non literal way requires a solid understanding of semantics and a deep real-world knowledge. In the longstanding debate about whether Neural Language Models (NLMs) really have a full understanding of text, analysing how they can recognise figurative language can provide some intuition of their functioning, their capabilities and their limits. Therefore, in this paper, we exploit probing tasks to study how several NLMs of different sizes recognise four different figures of speech: hyperboles, metaphors, oxymorons and pleonasms. We analyse whether this information is learned and how it is acquired during the training of the model, describing its learning trajectory. Moreover, we analyse which layers have a better comprehension of figurative language and the influence of pre-training data. Datasets and code are available at https://github.com/nicolarici/learning-trajectories.

Bang Trinh Tran To, Thai Le

This work presents LURK (Latent Unlearned Knowledge), a novel framework that probes for undesired knowledge retention in unlearned LLMs through adversarial suffix prompting. LURK automatically generates adversarial prompt suffixes designed to elicit residual knowledge about the Harry Potter domain, a commonly used benchmark for unlearning. Our experiments reveal that even models deemed successfully unlearned can leak idiosyncratic information under targeted adversarial conditions, highlighting critical limitations of current unlearning evaluation standards. By uncovering implicit knowledge through indirect probing, LURK offers a more rigorous and diagnostic tool for assessing the robustness of unlearning algorithms. Code and data will be available at https://github.com/Rachel1809/LURK.

Qi Chai, Zhang Zheng, Junlong Ren, Deheng Ye, Zichuan Lin, Hao Wang

Minecraft, as an open-world virtual interactive environment, has become a prominent platform for research on agent decision-making and execution. Existing works primarily adopt a single Large Language Model (LLM) agent to complete various in-game tasks. However, for complex tasks requiring lengthy sequences of actions, single-agent approaches often face challenges related to inefficiency and limited fault tolerance. Despite these issues, research on multi-agent collaboration remains scarce. In this paper, we propose CausalMACE, a holistic causality planning framework designed to enhance multi-agent systems, in which we incorporate causality to manage dependencies among subtasks. Technically, our proposed framework introduces two modules: an overarching task graph for global task planning and a causality-based module for dependency management, where inherent rules are adopted to perform causal intervention. Experimental results demonstrate our approach achieves state-of-the-art performance in multi-agent cooperative tasks of Minecraft. The code will be open-sourced upon the acceptance of this paper.

Jiho Park, Jongyoon Song, Minjin Choi, Kyuho Heo, Taehun Huh, Ji Won Kim

Large language models (LLMs) are increasingly integral as productivity assistants, but existing benchmarks fall short in rigorously evaluating their real-world instruction-following capabilities. Current benchmarks often (i) lack sufficient multilinguality, (ii) fail to capture the implicit constraints inherent in user requests, and (iii) overlook the complexities of multi-turn dialogue. To address these critical gaps and provide a more realistic assessment, we introduce TRUEBench (Trustworthy Real-world Usage Evaluation Benchmark), a novel benchmark specifically designed for LLM-based productivity assistants. TRUEBench distinguishes itself by featuring input prompts across 12 languages, incorporating intra-instance multilingual instructions, employing rigorous evaluation criteria to capture both explicit and implicit constraints, and including complex multi-turn dialogue scenarios with both accumulating constraints and context switches. Furthermore, to ensure reliability in evaluation, we refined constraints using an LLM validator. Extensive experiments demonstrate that TRUEBench presents significantly greater challenges than existing benchmarks; for instance, a strong model like OpenAI o1 achieved only a 69.07% overall pass rate. TRUEBench offers a demanding and realistic assessment of LLMs in practical productivity settings, highlighting their capabilities and limitations.

Zhe Li, Wei Zhao, Yige Li, Jun Sun

Influence functions are important for quantifying the impact of individual training data points on a model’s predictions. Although extensive research has been conducted on influence functions in traditional machine learning models, their application to large language models (LLMs) has been limited. In this work, we conduct a systematic study to address a key question: do influence functions work on LLMs? Specifically, we evaluate influence functions across multiple tasks and find that they consistently perform poorly in most settings. Our further investigation reveals that their poor performance can be attributed to: (1) inevitable approximation errors when estimating the iHVP component due to the scale of LLMs, (2) uncertain convergence during fine-tuning, and, more fundamentally, (3) the definition itself, as changes in model parameters do not necessarily correlate with changes in LLM behavior. Thus, our study suggests the need for alternative approaches for identifying influential samples.

Zekun Wang, Jingjie Zeng, Yingxu Li, Liang Yang, Hongfei Lin

Large language models (LLMs) achieve remarkable performance across various domains, largely due to training on massive datasets. However, this also raises growing concerns over the exposure of sensitive and private information, making model unlearning increasingly critical.However, existing methods often struggle to balance effective forgetting with maintaining model utility. In this work, we propose HyperUnlearn, a human-inspired unlearning framework. We construct two types of fuzzy data—local and global—to simulate forgetting, and represent them in hyperbolic and Euclidean spaces, respectively. Unlearning is performed on a model with frozen early layers to isolate forgetting and preserve useful knowledge.Experiments demonstrate that HyperUnlearn effectively forgets sensitive content while maintaining the model’s language understanding, fluency, and benchmark performance, offering a practical trade-off between forgetting and capability preservation.

Ruilin Luo, Tianle Gu, Lin Wang, Yunfeng Zhou, Songtao Jiang, Lei Wang, Yujiu Yang

Text-Attributed Graphs (TAGs), which integrate text and graph structures, have recently gained traction, especially in web applications. However, as a graph structure, TAG representation learning (TAGRL) naturally inherits issues from Graph Neural Networks (GNNs), such as fairness. Moreover, previous TAGRL research has mainly focused on using LM-as-encoder to boost downstream task performance, with little consideration given to whether this process may raise additional concerns related to fairness and other safety-related issues. As the first work to explore fairness in TAGRL, this paper proposes the concept of evolving LM-as-encoder to LM-as-fair-encoder, developing a two-stage fairness-aware alignment process called FairTAG based on the observed issues. Specifically, we first mitigate the tendency of LMs to overfit to homophily during downstream tasks fine-tuning, followed by subgraph-level connection behavior preference optimization for selected anchor nodes. We provide theoretical support and demonstrate the feasibility of LM-as-fair-encoder through extensive experiments and ablation studies. We also show that FairTAG can be seamlessly integrated with fairness-enhancing strategies on the GNNs decoder side, thus innovatively constructing a plug-and-play learning framework.

Yunfei Wang, Yeqin Zhang, Yuyang Wu, Liang Lu, Phi Le Nguyen, Xiaoliang Wang, Cam-Tu Nguyen

As machine learning (ML) application continues to expand across diverse fields, there is a rising demand for ML code generation. In this paper, we aim at a critical research question: Can machines autonomously generate ML code for sophisticated, human-designed algorithms or solutions? To answer this question, we introduce a novel benchmark, MLAlgo-Bench, which includes two challenging tasks: 1) Generating code for ML algorithms including both traditional ML and modern deep learning-based methods, and 2) Giving humans solution sketches, writing ML code for solving practical tasks in Kaggle competitions. This benchmark is unique in its focus on the challenges of interpreting intricate human instructions and producing multi-step, high-complexity code, offering a rigorous test for current Large Language Model (LLM) capabilities. We introduce an automatic evaluation framework with comprehensive metrics such as task pass rate, relative performance metric, and time overhead. Currently, the top-performing models (Claude3.5-Sonet) achieve a 48.8% task completion rate on realizing machine learning algorithms, and a 21.6% rate for completing Kaggle competitions. Further analysis suggests substantial room for improvement.

Aysenur Kocak, Shuo Yang, Bardh Prenkaj, Gjergji Kasneci

Pre-trained language models have achieved remarkable success across diverse applications but remain susceptible to spurious, concept-driven correlations that impair robustness and fairness. In this work, we introduce CURE, a novel and lightweight framework that systematically disentangles and suppresses conceptual shortcuts while preserving essential content information. Our method first extracts concept-irrelevant representations via a dedicated content extractor reinforced by a reversal network, ensuring minimal loss of task-relevant information. A subsequent controllable debiasing module employs contrastive learning to finely adjust the influence of residual conceptual cues, enabling the model to either diminish harmful biases or harness beneficial correlations as appropriate for the target task. Evaluated on the IMDB and Yelp datasets using three pre-trained architectures, CURE achieves an absolute improvement of +10 points in F1 score on IMDB and +2 points on Yelp, while introducing minimal computational overhead. Our approach establishes a flexible, unsupervised blueprint for combating conceptual biases, paving the way for more reliable and fair language understanding systems.

Zhensheng Jin, Xinze Li, Yifan Ji, Chunyi Peng, Zhenghao Liu, Qi Shi, Yukun Yan, Shuo Wang, Furong Peng, Ge Yu

Recent advances in Chain-of-Thought (CoT) prompting have substantially improved the reasoning capabilities of Large Language Models (LLMs). However, these methods often suffer from overthinking, leading to unnecessarily lengthy or redundant reasoning traces. Existing approaches attempt to mitigate this issue through curating multiple reasoning chains for training LLMs, but their effectiveness is often constrained by the quality of the generated data and prone to overfitting. To address the challenge, we propose Reasoning Compression Through Stepwise Trials (ReCUT), a novel method aimed at balancing the accuracy and length of reasoning trajectory. Specifically, ReCUT employs a stepwise exploration mechanism and a long-short switched sampling strategy, enabling LLMs to incrementally generate diverse reasoning paths. These paths are evaluated and used to construct preference pairs to train two specialized models (Gemini LLMs)—one optimized for reasoning accuracy, the other for shorter reasoning. A final integrated model is obtained by interpolating the parameters of these two models. Experimental results across multiple math reasoning datasets and backbone models demonstrate that ReCUT significantly reduces reasoning lengths by approximately 30-50%, while maintaining or improving reasoning accuracy compared to various baselines. All codes and data will be released via https://github.com/NEUIR/ReCUT.

Hangyu He, Xin Yuan, Kai Wu, Ren Ping Liu, Wei Ni

Large Language Models (LLMs) can undergo hallucinations in specialized domains, and standard Retrieval-Augmented Generation (RAG) often falters due to general-purpose embeddings ill-suited for domain-specific terminology. Though domain-specific fine-tuning enhances retrieval, centralizing data introduces privacy risks. The use of federated learning (FL) can alleviate this to some extent, but faces challenges of data heterogeneity, poor personalization, and expensive training data generation. We propose pFedRAG, a novel Personalized Federated RAG framework, which enables efficient collaborative fine-tuning of embedding models to address these challenges. The key contribution is a new Depth-Adaptive Tiered Embedding (DATE) architecture, which comprises a Global Shared Layer, combined using FL to capture common knowledge, and a Personalized Layer with adjustable depth tailored for local data and training results of each client. The depth is locally controlled based on crafted metrics and scoring criteria. Also, pFedRAG incorporates a fully client-side pipeline leveraging local small LLMs and vector database filtering to construct high-quality query-document pairs. Experiments on diverse medical non-IID document datasets demonstrate that pFedRAG significantly reduces communication costs, handles data heterogeneity, and improves retrieval performance. Human evaluations confirm the enhanced response quality of pFedRAG.

Siskos Dimitrios, Stavros Papadopoulos, Pablo Peso Parada, Jisi Zhang, Karthikeyan Saravanan, Anastasios Drosou

This work investigates retrieval augmented generation as an efficient strategy for automatic context discovery in context-aware Automatic Speech Recognition (ASR) system, in order to improve transcription accuracy in the presence of rare or out-of-vocabulary terms. However, identifying the right context automatically remains an open challenge. This work proposes an efficient embedding-based retrieval approach for automatic context discovery in ASR. To contextualize its effectiveness, two alternatives based on large language models (LLMs) are also evaluated: (1) large language model (LLM)-based context generation via prompting, and (2) post-recognition transcript correction using LLMs. Experiments on the TED-LIUMv3, Earnings21 and SPGISpeech demonstrate that the proposed approach reduces WER by up to 17% (percentage difference) relative to using no-context, while the oracle context results in a reduction of up to 24.1%.

Wei Zhao, Zhe Li, Yige Li, Jun Sun

Large Vision-Language Models (LVLMs) have made significant strides in multimodal comprehension, thanks to extensive pre-training and fine-tuning on large-scale visual datasets. However, despite their robust textual safety mechanisms, they remain vulnerable to harmful visual inputs. Existing safeguards—typically relying on pre-filtering or fine-tuning—incur high costs and diminish overall utility. To address this critical vulnerability, we introduce SafeCLIP, a lightweight method that leverages LVLMs’ inherent multimodal alignment for zero-shot toxic image detection. By projecting CLIP’s discarded CLS token into its text space and matching it with toxic descriptors, SafeCLIP detects harmful content without any architectural changes—adding minimal latency and enabling dynamic safety corrections during inference and fine-tuning. Experiments show that SafeCLIP achieves a 66.9% defense success rate with only 3.2% false positive rate and 7.2% overhead. In contrast, state-of-the-art methods achieve 52.9% success but have a 10.7% false positive rate and 210% overhead. Our work demonstrates that leveraging inherent multimodal alignment can yield efficient, low-cost LVLM safety. Code is available at anonymous.4open.science/r/safeclip-2C01.

Xiaotian Han, Yiren Jian, Xuefeng Hu, Haogeng Liu, Yiqi Wang, Qihang Fan, Yuang Ai, Huaibo Huang, Ran He, Zhenheng Yang 等

Pre-training on large, high-quality datasets is essential for improving the reasoning abilities of Large Language Models (LLMs), particularly in specialized fields like mathematics. However, the field of Multimodal LLMs (MLLMs) lacks a comprehensive, open-source dataset for mathematical reasoning. To fill this gap, we present InfiMM-WebMath-40B, a high-quality dataset of interleaved image-text documents. It consists of 24 million web pages, 85 million image URLs, and 40 billion text tokens, all carefully extracted and filtered from CommonCrawl. We outline our data collection and processing pipeline in detail. Models trained on InfiMM-WebMath-40B demonstrate strong performance in both text-only and multimodal settings, setting a new state-of-the-art on multimodal math benchmarks such as MathVerse and We-Math.