论文检索

Kyungmin Kim, Youngbin Choi, Hyounghun Kim, Dongwoo Kim, Sangdon Park

In this paper, we propose \texttt{ChronoBias}, a novel benchmark for evaluating \textit{time-conditional group bias} in the \textit{time-sensitive} knowledge of large language models (LLMs).Our benchmark is constructed via a template-based semi-automated generation method, balancing the quality-quantity trade-off in existing benchmark curation approaches.For knowledge that changes over time, \textit{time-conditional group bias} exhibits varying patterns across time intervals, evident in both the best- and worst-performing groups and in the bias metric itself.In addition to \textit{parametric knowledge bias}–which influences group bias across all time intervals–we identify \textit{time-sensitivity bias} as an additional factor after a model’s knowledge cutoff, accounting for much of the variation in \textit{time-conditional group bias} over time.Since both biases are irreducible, retrieval-augmented generation (RAG) can be a promising approach, as it can address post-cutoff knowledge and better leverage pretraining knowledge that is underrepresented in the model parameters.While RAG improves both overall performance and group bias, we observe that the disparate patterns of \textit{time-conditional group bias} still persist.Therefore, through extensive experiments with various model configurations, we illustrate how accurate and fair RAG-based LLMs should behave and provide actionable guidelines toward constructing such ideal models.

Sanjay Govindan, Maurice Pagnucco, Yang Song

Large Language Models (LLMs) are trained on diverse and often conflicting knowledge spanning multiple domains and time periods. Some of this knowledge is only valid within specific temporal contexts, such as answering the question, “Who is the President of the United States in 2022?” Ensuring LLMs generate time-appropriate responses is crucial for maintaining relevance and accuracy. In this work we explore activation engineering as a method for temporally aligning LLMs to improve factual recall without any training. Activation engineering has predominantly been used to steer subjective and qualitative outcomes such as toxicity or behavior. Our research is one of few that uncovers the bounds of activation engineering on objective outcomes. We explore an activation engineering technique to anchor LLaMA 2, LLaMA 3.1, Qwen 2 and Gemma 2 to specific points in time and examine the effects of varying injection layers and prompting strategies. Our experiments demonstrate up to a 44% and 16% improvement in relative and explicit prompting respectively, achieving comparable performance to the fine-tuning method proposed by Zhao et al. (2024). Notably, for LLaMA 2 and LLaMA 3.1 our approach achieves similar results to the fine-tuning baseline while being significantly more computationally efficient and requiring no pre-aligned datasets.

Ding-Chu Zhang, Xiaowen Zhang, Yue Fei, Renjun Hu, Xiao-Wen Yang, Zhi Zhou, Baixuan Li, Yu-Feng Li, Xing Shi, Wei Lin

Retrieval-augmented generation (RAG) enables large language models (LLMs) to address queries beyond their internal knowledge by integrating domain knowledge in specialized corpus, which necessitates the generation of benchmarks on specific corpus to evaluate RAG systems. However, existing automated generation methods exhibit Weak Applicability and Weak Scalability. Weak Applicability refers to the reliance on metadata from specific corpora for query generation, constraining applicability to other corpora. Weak Scalability is characterized by fixed query content after generation, unable to dynamically increase difficulty, limiting scalability of the query. To overcome these issues, we propose AutoEvolve, an applicable approach for dynamically evolving queries to construct scalable RAG benchmarks. Our approach is grounded in three key innovations: (i) a corpus-agnostic method for constructing the universal entity-document graph; (ii) a suite of evolution operations designed to dynamically update queries; and (iii) a difficulty-guided metric that directs query evolution process. Through experiments on three generated benchmarks, we demonstrate that AutoEvolve evolves queries that are significantly more challenging, paving the way for more applicable and scalable RAG evaluations.

Yanhong Li, Min Yang, Xiping Hu, Chengming Li

Recent studies have highlighted the remarkable knowledge retention capabilities of Large Language Models (LLMs) like GPT-4, while simultaneously revealing critical limitations in maintaining knowledge currency and accuracy. Existing knowledge editing methodologies, designed to update specific factual information without compromising general model performance, often encounter two fundamental challenges: parameter conflict during knowledge overwriting and excessive computational overhead. In this paper, we introduce ForGet (Forget for Get), a novel approach grounded in the principle of “forgetting before learning”. By pinpointing the location within the LLM that corresponds to the target knowledge, we first erase the outdated knowledge and then insert the new knowledge at this precise spot. ForGet is the first work to leverage a two-phase gradient-based process for knowledge editing, offering a lightweight solution that also delivers superior results. Experimental findings show that our method achieves more effective knowledge editing at a lower cost compared to previous techniques across various base models.

Chengbing Wang, Yang Zhang, Zhicheng Wang, Tianhao Shi, Keqin Bao, Fuli Feng, Tat-Seng Chua

Fine-tuning large language models (LLMs) for recommendation in a generative manner has delivered promising results, but encounters significant inference overhead due to autoregressive decoding in the language space. This work explores bypassing language-space decoding by directly matching candidate items with the LLM’s internal thought representations in the latent space, eliminating the time-consuming autoregressive process to reduce computational costs. Towards this, we introduce Light Latent-space Decoding (L2D), an effective and efficient latent-space decoding method. L2D represents user-preferred items by using the hidden states of test sequences reflecting the LLM’s internal thought, and obtains candidate item representations from the hidden states of training sequences labeled with the corresponding candidate items. It then matches the two types of representations to decode items, achieving latent-space decoding. In this way, it enables efficient decoding without altering the LLM’s generative tuning paradigm, thereby preserving performance. Extensive empirical results demonstrate that L2D is more than 10x faster than language-space decoding while maintaining or enhancing performance.

Euntae Choi, Sumin Song, Woosang Lim, Sungjoo Yoo

We propose Rotate, Clip, and Partition (RCP), a Quantization-Aware Training (QAT) approach that first realizes extreme compression of LLMs with W2A4KV4 (2-bit weight, 4-bit activation, and 4-bit KV-cache) configuration. RCP integrates recent rotation techniques with a novel non-uniform weight quantizer design by theoretically and empirically analyzing the impact of rotation on the non-uniformity of weight distribution. Our weight quantizer, Learnable Direct Partitioning (LDP), introduces learnable parameters to directly learn non-uniform intervals jointly with LLM weights. We also present a GPU kernel supporting GEMV on non-uniform W2A4 as proof of concept. Experiments show that RCP can compress LLaMA-2-7B to W2A4KV4 with a loss of only 2.84 WikiText2 PPL and 5.29 times reduced memory footprint. Furthermore, RCP can quantize challenging mobile-targeted LLaMA-3.2 models and domain-specific WizardCoder-7B and MetaMath-7B with no critical problems such as convergence failure and repetition. Code is available at https://github.com/songsm921/RCP.

Xinrong Chen, Hengyuan Zhang, Yingmin Qiu, Xiao Liang, Ziyue Li, Guanyu Wang, Weiping Li, Tong Mo, Hayden Kwok-Hay So, Ngai Wong

Parameter-efficient fine-tuning (PEFT) methods, particularly Low-Rank Adaptation (LoRA), offer an efficient way to adapt large language models with reduced computational costs. However, their performance is limited by the small number of trainable parameters. Recent work combines LoRA with the Mixture-of-Experts (MoE), i.e., LoRA-MoE, to enhance capacity, but two limitations remain in hindering the full exploitation of its potential: 1) the influence of downstream tasks when assigning expert numbers, and 2) the uniform rank assignment across all LoRA experts, which restricts representational diversity.To mitigate these gaps, we propose GuiLoMo, a fine-grained layer-wise expert numbers and ranks allocation strategy with GuidedSelection Vectors (GSVs). GSVs are learned via a prior bilevel optimization process to capture both model- and task-specific needs, and are then used to allocate optimal expert numbers and ranks.Experiments on three backbone models across diverse benchmarks show that GuiLoMo consistently achieves superior or comparable performance to all baselines. Further analysis offers key insights into how expert numbers and ranks vary across layers and tasks, highlighting the benefits of adaptive expert configuration. Our code is available at https://anonymous.4open.science/r/GuiLoMo-034.

Jinyuan Feng, ChaoPeng Wei, Tenghai Qiu, Tianyi Hu, Zhiqiang Pu

In parameter-efficient fine-tuning, mixture-of-experts (MoE), which involves specializing functionalities into different experts and sparsely activating them appropriately, has been widely adopted as a promising approach to trade-off between model capacity and computation overhead. However, current MoE variants fall short on heterogeneous datasets, ignoring the fact that experts may learn similar knowledge, resulting in the underutilization of MoE’s capacity. In this paper, we propose Contrastive Representation for MoE (CoMoE), a novel method to promote modularization and specialization in MoE, where the experts are trained along with a contrastive objective by sampling from activated and inactivated experts in top-k routing. We demonstrate that such a contrastive objective recovers the mutual-information gap between inputs and the two types of experts. Experiments on several benchmarks and in multi-task settings demonstrate that CoMoE can consistently enhance MoE’s capacity and promote modularization among the experts.

Yao Yan

Arithmetic offers a compact test of whether large language models compute or memorize. We study multi-digit addition in LLaMA-3-8B-Instruct using linear probes and the Logit Lens, and find a consistent four-stage, layer-wise ordering of probe-decodable signal types across depth: (1) early layers encode formula structure (operand/operator layout) while the gold next token is still far from top-1; (2) mid layers expose digit-wise sums and carry indicators; (3) deeper layers express result-level numerical abstractions that support near-perfect digit decoding from hidden states; and (4) near the output, representations align with final sequence generation, with the correct next token reliably ranked first. Across experiments, each signal family becomes linearly decodable with high accuracy (stage-wise peaks typically \geq95% on in-domain multi-digit addition, and up to 99%). Taken together, these observations—in our setting—are consistent with a hierarchical, computation-first account rather than rote pattern matching, and help explain why Logit Lens inspection is most informative mainly in later layers. Code and data are available at https://github.com/YaoToolChest/addition-in-four-movements.git.

Wenhao Zheng, Liaoyaqi Wang, Dongshen Peng, Hongxia Xu, Yun Li, Hongtu Zhu, Tianfan Fu, Huaxiu Yao

Clinical trials are pivotal yet costly processes, often spanning multiple years and requiring substantial expenses, motivating predictive models to identify likely-to-fail drugs early and save resources. Recent approaches leverage deep learning to integrate multimodal data for clinical outcome prediction; however, they rely heavily on manually designed modality-specific encoders, limiting their adaptability to new modalities and ability to effectively share information across modalities. To address these challenges, we propose a multimodal mixture-of-experts (LIFTED) framework. Specifically, LIFTED transforms modality-specific data into natural language descriptions, encoded via unified, noise-resilient encoders. A sparse Mixture-of-Experts mechanism then identifies shared patterns across modalities, extracting consistent representations. Finally, another mixture-of-experts module dynamically integrates these modality representations, emphasizing critical information. Experiments show that LIFTED significantly outperforms baseline methods in predicting clinical trial outcomes across all phases, highlighting the effectiveness of our proposed approach.

Xinpeng Ti, Wentao Ye, Zhifang Zhang, Junbo Zhao, Chang Yao, Lei Feng, Haobo Wang

With the continuous development of language models and the widespread availability of various types of accessible interfaces, large language models (LLMs) have been applied to an increasing number of fields. However, due to the vast amounts of data and computational resources required for model development, protecting the model’s parameters and training data has become an urgent and crucial concern. Due to the revolutionary training and application paradigms of LLMs, many new attacks on language models have emerged in recent years. In this paper, we define these attacks as “reverse engineering” (RE) techniques on LMs and aim to provide an in-depth analysis of reverse engineering of language models. We illustrate various methods of reverse engineering applied to different aspects of a model, while also providing an introduction to existing protective strategies. On the one hand, it demonstrates the vulnerabilities of even black box models to different types of attacks; on the other hand, it offers a more holistic perspective for the development of new protective strategies for models.

Chenlong Wang, Yuanning Feng, Dongping Chen, Zhaoyang Chu, Ranjay Krishna, Tianyi Zhou

Recent advances in large reasoning models have enabled complex, step-by-step reasoning but often introduce significant overthinking, resulting in verbose and redundant outputs that hinder efficiency. In this study, we examine whether explicit self-reflection, signaled by tokens such as “Wait” and “Hmm”, is necessary for advanced reasoning. We propose NoWait, a simple yet effective approach that disables explicit self-reflection by suppressing these tokens during inference. Extensive experiments on ten benchmarks across textual, visual, and video reasoning tasks show that NoWait reduces chain-of-thought trajectory length by up to 27%–51% in five R1-style model series, without compromising model utility. NoWait thus offers a plug-and-play solution for efficient and utility-preserving multimodal reasoning.

Ishan Jindal, Jayant Taneja, Badrinath Chandana, Vikas Kapur, Sachin Dev Sharma

Large language models (LLMs) demonstrate strong reasoning capabilities but are expensive to run at inference time, limiting their practical deployment. We propose Offloaded Reasoning (OR), a modular strategy where a lightweight model generates intermediate reasoning traces that are then used by a larger model to produce the final answer. We further introduce Offloaded Reasoning with Refinement (ORR), where the large model first edits or improves the reasoning trace before answering. Unlike token-level acceleration methods, OR and ORR operate at the reasoning level and require no retraining of the large model. Experiments on GSM8K and Math500 show that OR achieves up to 8x faster inference than full large-model reasoning with minimal accuracy loss, while ORR recovers or exceeds full accuracy at substantially lower cost. Our results highlight the potential of modular, delegation-based reasoning for building more efficient and adaptable LLM systems.

Yutong Gao, Maoyuan Shao, Xinyang Huang, Chuang Zhu, Yu Weng, Xuan Liu, Lijuan Sun, Guoshun Nan

CLIP’s success has demonstrated that prompt tuning can achieve robust cross-modal semantic alignment for tasks ranging from open-domain recognition to fine-grained classification. However, redundant or weakly relevant feature components introduce noise and incur unnecessary computational costs. In this work, we propose Spotlighter, a lightweight token-selection framework that simultaneously enhances accuracy and efficiency in prompt tuning. Spotlighter evaluates each visual token’s activation from both sample-wise and semantic-wise perspectives and retains only the top-scoring tokens for downstream prediction. A class-specific semantic memory bank of learned prototypes refines this selection, ensuring semantic representativeness and compensating for discarded features. To further prioritize informative signals, we introduce a two-level ranking mechanism that dynamically weights token–prototype interactions. Across 11 few-shot benchmarks, Spotlighter outperforms CLIP by up to 11.19% in harmonic mean accuracy and achieves up to 0.8K additional FPS, with only 21 extra parameters. These results establish Spotlighter as an effective and scalable baseline for prompt tuning.

Xiangfeng Luo, Ruoxin Zheng, Jianqiang Huang, Hang Yu

Although Hyper-relational Knowledge Graphs (HKGs) can model complex facts better than traditional KGs, the Hyper-relational Knowledge Graph Completion (HKGC) is more sensitive to inherent noise, particularly struggling with two prevalent HKG-specific noise types: Intra-fact Inconsistency and Cross-fact Association Noise.To address these challenges, we propose **HDiff**, a novel conditional denoising diffusion framework for robust HKGC that learns to reverse structured noise corruption. HDiff integrates a **Consistency-Enhanced Global Encoder (CGE)** using contrastive learning to enforce intra-fact consistency and a **Context-Guided Denoiser (CGD)** performing iterative refinement. The CGD features dual conditioning leveraging CGE’s global context and local confidence estimates, effectively combatting both noise types. Extensive experiments demonstrate that HDiff substantially outperforms state-of-the-art HKGC methods, highlighting its effectiveness and significant robustness, particularly under noisy conditions.

Mansour Al Ghanim, Jiaqi Xue, Rochana Prih Hastuti, Mengxin Zheng, Yan Solihin, Qian Lou

We present a study to benchmark representative watermarking methods in cross-lingual settings. The current literature mainly focuses on the evaluation of watermarking methods for the English language. However, the literature for evaluating watermarking in cross-lingual settings is scarce. This results in overlooking important adversary scenarios in which a cross-lingual adversary could be in, leading to a gray area of practicality over cross-lingual watermarking. In this paper, we evaluate four watermarking methods in four different and vocabulary rich languages. Our experiments investigate the quality of text under different watermarking procedure and the detectability of watermarks with practical translation attack scenarios. Specifically, we investigate practical scenarios that an adversary with cross-lingual knowledge could take, and evaluate whether current watermarking methods are suitable for such scenarios. Finally, from our findings, we draw key insights about watermarking in cross-lingual settings.

Yuchen Su, Yonghua Zhu, Ruofan Wang, Zijian Huang, Diana Benavides-Prado, Michael Witbrock

Pun generation seeks to creatively modify linguistic elements in text to produce humour or evoke double meanings. It also aims to preserve coherence and contextual appropriateness, making it useful in creative writing and entertainment across various media and contexts. This field has been widely studied in computational linguistics, while there are currently no surveys that specifically focus on pun generation. To bridge this gap, this paper provides a comprehensive review of pun generation datasets and methods across different stages, including traditional approaches, deep learning techniques, and pre-trained language models. Additionally, we summarise both automated and human evaluation metrics used to assess the quality of pun generation. Finally, we discuss the research challenges and propose promising directions for future work.

Abhijit Chakraborty, Chahana Dahal, Vivek Gupta

Federated Retrieval-Augmented Generation (Federated RAG) combines Federated Learning (FL),which enables distributed model training without exposing raw data, with Retrieval-Augmented Generation (RAG), which improves the factual accuracy of language models by grounding outputs in external knowledge. As large language models are increasingly deployed in privacy-sensitive domains such as healthcare, finance, and personalized assistance, Federated RAG offers a promising framework for secure, knowledge-intensive natural language processing (NLP). To the best of our knowledge, this paper presents the first systematic mapping study of Federated RAG, covering literature published between 2020 and 2025. Following Kitchenham’s guidelines for evidence-based software engineering, we develop a structured classification of research focuses, contribution types, and application domains. We analyze architectural patterns, temporal trends, and key challenges, including privacy-preserving retrieval, cross-client heterogeneity, and evaluation limitations. Our findings synthesize a rapidly evolving body of research, identify recurring design patterns, and surface open questions, providing a foundation for future work at the intersection of RAG and federated systems.

Zahidul Islam, Mrigank Rochan

Recent advances in Large Multimodal Models (LMMs) have showcased impressive visual understanding and vision-language reasoning capabilities, yet their computational cost hinders practical deployment, especially in resource-constrained settings. A key bottleneck is the large number of visual tokens generated by its vision encoders, which increases latency and memory demands. Existing token reduction methods often require costly fine-tuning or apply fixed token reduction ratios, ignoring image complexity and vision-language interactions. We propose AdaptMerge, a training-free, inference-time token merging strategy that adaptively reduces visual tokens by leveraging feature diversity and language-guided relevance. By dynamically adjusting to image complexity and ensuring multimodal coherence, AdaptMerge significantly lowers floating-point operations while improving performance. Extensive experiments on Google’s latest Gemma 3 models (4B and 12B parameters) across four challenging benchmarks demonstrate that AdaptMerge outperforms state-of-the-art token reduction techniques, achieving both reduced computational costs and improved performance, thereby providing a practical pathway to more efficient LMMs.

Mahammed Kamruzzaman, Amanda Cercas Curry, Alba Cercas Curry, Flor Miriam Plaza-del-Arco

Large language models (LLMs) are increasingly used to predict human emotions, but previous studies show that these models reproduce gendered emotion stereotypes. Emotion stereotypes are also tightly tied to race and skin tone (consider for example the trope of the angry black woman), but previous work has thus far overlooked this dimension. In this paper, we address this gap by introducing the first large-scale multimodal study of racial, gender, and skin-tone bias in emotion attribution, revealing how modality (text, images) and their combination shape emotion stereotypes in Multimodal LLMs (MLLMs). We evaluate four open-source MLLMs using 2.1K emotion-related events paired with 400 neutral face images across three different prompt strategies. Our findings reveal varying biases in MLLMs representations of different racial groups: models reproduce racial stereotypes across modalities, with textual cues being particularly noticeable. Models also reproduce colourist trends, with darker skin tones showing more skew. Our research highlights the need for future rigorous evaluation and mitigation strategies that account for race, colorism, and gender in MLLMs.