Recent advances in long-context language models (LCLMs), designed to handle extremely long contexts, primarily focus on utilizing external contextual information, often leaving the influence of language models' parametric knowledge underexplored. In this work, we firstly investigate how this parametric knowledge affects content generation and demonstrate that its impact becomes increasingly pronounced as context length extends. Furthermore, we show that the model’s ability to utilize parametric knowledge, which we call parametric recall ability, does not improve simultaneously with its ability to leverage contextual knowledge through extrinsic retrieval ability. Moreover, better extrinsic retrieval ability can interfere with the model’s parametric recall ability, limiting its full potential. To bridge this gap, we design a simple yet effective Hybrid Needle-in-a-Haystack test that evaluates models based on their capabilities across both abilities, rather than solely emphasizing extrinsic retrieval ability. Our experimental results reveal that Qwen-2.5 models significantly outperform Llama-3.1 models, demonstrating superior potential to combine various abilities. Moreover, even the more powerful Llama-3.1-70B-Instruct model fails to exhibit better performance, highlighting the importance of evaluating models from a dual-ability perspective.

Yu Fu, Chen Luo, Josef Valvoda, Xin Zhang, Xuejing Lei, Xiao Pan, Hui Liu, Yue Dong. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Weiqi Wang, Limeng Cui, Xin Liu, Sreyashi Nag, Wenju Xu, Chen Luo, Sheikh Muhammad Sarwar, Yang Li, Hansu Gu, Hui Liu, Changlong Yu, Jiaxin Bai, Yifan Gao, Haiyang Zhang, Qi He, Shuiwang Ji, Yangqiu Song. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Process-supervised reward models serve as a fine-grained function that provides detailed step-wise feedback to model responses, facilitating effective selection of reasoning trajectories for complex tasks.Despite its advantages, evaluation on PRMs remains less explored, especially in the multimodal domain.To address this gap, this paper first benchmarks current vision large language models (VLLMs) as two types of reward models: output reward models (ORMs) and process reward models (PRMs) on multiple vision-language benchmarks, which reveal that neither ORM nor PRM consistently outperforms across all tasks, and superior VLLMs do not necessarily yield better rewarding performance.To further advance evaluation, we introduce VILBENCH, a vision-language benchmark designed to require intensive process reward signals.Notably, Ope-nAI's GPT-4o with Chain-of-Thought (CoT) achieves only 27.3% accuracy, challenging current VLLMs.Lastly, we preliminarily showcase a promising pathway towards bridging the gap between general VLLMs and reward models-by collecting 73.6K vision-language process reward data using an enhanced treesearch algorithm, our 3B model is able to achieve an average improvement of 3.3% over standard CoT and up to 2.5% compared to its untrained counterpart on VILBENCH by selecting OpenAI o1's generations.We will release our code, model, and data at https: //ucsc-vlaa.github.io/ViLBench.

Multimodal Large Language Models (MLLMs) trained on massive data may memorize sensitive personal information and photos, posing serious privacy risks.To mitigate this, MLLM unlearning methods are proposed, which finetune MLLMs to forget sensitive information.However, it remains unclear whether the knowledge has been truly forgotten or just hidden in the model.Therefore, we propose to study a novel problem of MLLM unlearning attack, which aims to recover the unlearned knowledge of an unlearned MLLM.To achieve the goal, we propose a novel framework-Stealthy Unlearning Attack (SUA)-that learns a universal noise pattern.When applied to input images, this noise can trigger the model to reveal unlearned content.While pixel-level perturbations may be visually subtle, they can be detected in the semantic embedding space, making such attacks vulnerable to potential defenses.To improve stealthiness, we introduce an embedding alignment loss that minimizes the difference between the perturbed and denoised image embeddings, ensuring that the attack remains semantically unnoticeable.Experimental results show that SUA can effectively recover unlearned information from MLLMs.Furthermore, the learned noise generalizes well-i.e., a single perturbation trained on a few samples can reveal forgotten contents in unseen images.

Multimodal Retrieval-Augmented Generation (MRAG) systems enhance LMMs by integrating external multimodal databases, but introduce unexplored privacy vulnerabilities.While text-based RAG privacy risks have been studied, multimodal data presents unique challenges.We provide the first systematic analysis of MRAG privacy vulnerabilities across vision-language and speech-language modalities.Using a novel compositional structured prompt attack in a black-box setting, we demonstrate how attackers can extract private information by manipulating queries.Our experiments reveal that LMMs can both directly generate outputs resembling retrieved content and produce descriptions that indirectly expose sensitive information, highlighting the urgent need for robust privacy-preserving MRAG techniques.The code is available at

Large language models (LLMs) have achieved remarkable performance on vari- ous natural language tasks. However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world. This moti- vates the development of knowledge editing methods designed to update certain knowledge in LLMs without changing unrelated others. To make selective edits, previous efforts often sought to update a small amount of parameters in some spe- cific layer(s) of a LLM. Nonetheless, in challenging scenarios, they still fall short in making successful edits while preserving knowledge irrelevant to the updates simultaneously, resulting in a notable editing-locality trade-off. In this work, we question if the trade-offs are caused by the fact that parameter-based updates have a global effect, i.e., edited parameters affect all inputs indiscriminately. In light of this, we explore the feasibility of representation fine-tuning, which applied some linear update to a few representations in a learned subspace, for knowledge edit- ing. While being effective to enhance an LLM’s general ability as demonstrated in the previous work, we theoretically show that this linear update imposes a tension in editing-locality trade-off. Subsequently, BaFT is proposed to break the linear- ity. BaFT computes a weight for each basis that spans a dimension of the subspace based on the input representation. This input-dependent weighting mechanism al- lows BaFT to manage different types of knowledge in an adaptive way, thereby achieving a better editing-locality trade-off. Experiments on three LLMs with five editing benchmarks in diverse scenarios show the superiority of our method.

Large language models (LLMs) have shown remarkable proficiency in generating text, benefiting from extensive training on vast textual corpora. However, LLMs may also acquire unwanted behaviors from the diverse and sensitive nature of their training data, which can include copyrighted and private content. Machine unlearning has been introduced as a viable solution to remove the influence of such problematic content without the need for costly and time-consuming retraining. This process aims to erase specific knowledge from LLMs while preserving as much model utility as possible. Despite the effectiveness of current unlearning methods, little attention has been given to whether existing unlearning methods for LLMs truly achieve forgetting or merely hide the knowledge, which current unlearning benchmarks fail to detect. This paper reveals that applying quantization to models that have undergone unlearning can restore the "forgotten" information. We conduct comprehensive experiments using various quantization techniques across multiple precision levels to thoroughly evaluate this phenomenon. We find that for unlearning methods with utility constraints, the unlearned model retains an average of 21\% of the intended forgotten knowledge in full precision, which significantly increases to 83\% after 4-bit quantization. Based on our empirical findings, we provide a theoretical explanation for the observed phenomenon and propose a quantization-robust unlearning strategy aimed at mitigating this intricate issue. Our results highlight a fundamental tension between preserving the utility of the unlearned model and preventing knowledge recovery through quantization, emphasizing the challenge of balancing these two objectives. Altogether, our study underscores a major failure in existing unlearning methods for LLMs, strongly advocating for more comprehensive and robust strategies to ensure authentic unlearning without compromising model utility. Our code is available at: https://github.com/zzwjames/FailureLLMUnlearning.

Large language models (LLMs) have achieved remarkable performance on various natural language tasks. However, they are trained on static corpora and their knowledge can become outdated quickly in the fast-changing world. This motivates the development of knowledge editing (KE) to update specific knowledge in LLMs without changing unrelated others or compromising their pre-trained capabilities. Previous efforts sought to update a small amount of parameters of a LLM and proved effective for making selective updates. Nonetheless, the edited LLM often exhibits degraded ability to reason about the new knowledge. In this work, we identify a key issue: heterogeneous token overfitting (HTO), where the LLM overfits different tokens in the provided knowledge at varying rates. To tackle this, we propose OVERTONE, a token-level smoothing method that mitigates HTO by adaptively refining the target distribution. Theoretically, OVERTONE offers better parameter updates with negligible computation overhead. It also induces an implicit DPO but does not require preference data pairs. Extensive experiments across four editing methods, two LLMs, and diverse scenarios demonstrate the effectiveness and versatility of our method.

Large language models (LLMs) based on Transformer architecture are powerful but face challenges with deployment, inference latency, and costly fine-tuning. These limitations highlight the emerging potential of small language models (SLMs), which can either replace LLMs through innovative architectures and technologies, or assist them as efficient proxy or reward models. Emerging architectures such as Mamba and xLSTM address the quadratic scaling of inference with window length in Transformers by enabling linear scaling. To maximize SLM performance, test-time compute scaling strategies reduce the performance gap with LLMs by allocating extra compute budget during test time. Beyond standalone usage, SLMs could also assist in LLMs via weak-to-strong learning, proxy tuning, and guarding, fostering secure and efficient LLM deployment. Lastly, the trustworthiness of SLMs remains a critical yet underexplored research area. However, there is a lack of tutorials on cutting-edge SLM technologies, prompting us to conduct one.

Ran Xu, Hui Liu, Sreyashi Nag, Zhenwei Dai, Yaochen Xie, Xianfeng Tang, Chen Luo, Yang Li, Joyce C. Ho, Carl Yang, Qi He. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Shenglai Zeng, Jiankun Zhang, Bingheng Li, Yuping Lin, Tianqi Zheng, Dante Everaert, Hanqing Lu, Hui Liu, Hui Liu, Yue Xing, Monica Xiao Cheng, Jiliang Tang. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Test-time scaling (TTS) enhances the performance of large language models (LLMs) by allocating additional compute resources during inference. However, existing research primarily investigates TTS in single-stage tasks; while many real-world problems are multi-stage complex tasks, composed of a sequence of heterogeneous subtasks with each subtask requires LLM of specific capability. Therefore, we study a novel problem: the test-time compute-optimal scaling in multi-stage complex tasks, aiming to select suitable models and allocate budgets per subtask to maximize overall performance. TTS in multi-stage tasks introduces two fundamental challenges: (i) The combinatorial search space of model and budget allocations, combined with the high cost of inference, makes brute-force search impractical. (ii) The optimal model and budget allocations across subtasks are interdependent, increasing the complexity of the compute-optimal search. To address this gap, we conduct extensive pilot experiments on four tasks across six datasets, deriving three empirical insights characterizing the behavior of LLMs in multi-stage complex tasks. Informed by these insights, we propose AgentTTS, an LLM-agent-based framework that autonomously searches for compute-optimal allocations through iterative feedback-driven interactions with the execution environment. Experimental results demonstrate that AgentTTS significantly outperforms traditional and other LLM-based baselines in search efficiency, and shows improved robustness to varying training set sizes and enhanced interpretability.

Prompting LLMs offers an efficient way to guide output generation without explicit model training. In the e-commerce domain, prompt-based applications are widely used in query understanding, recommender systems, and customer support. However, adapting LLMs to different tasks often requires extensive prompt engineering by domain experts, along with frequent updates to align with evolving business needs. Additionally, crafting fully unbiased natural language prompts remains a challenge for humans. To address these challenges, we propose a novel framework, Examples as the Prompt (EaP). Specifically, EaP automatically selects the most representative examples to maximize the few-shot capability of LLMs. It is efficient due to its unsupervised example selection and adaptive to potential data distribution shifts. We validate EaP on four real-world production use cases, demonstrating that it achieves comparable or even superior performance comparing to hand-crafted prompts designed by domain experts. Additionally, we introduce EaPlite, which entirely replaces the natural language components of prompts with labeled examples. EaPlite improves LLM inference speed by up to 70% without compromising performance. The online A/B test shows that using EaP and EaPlite for data labeling can bring significant composite revenue gain by 0.06%.

Over the past several years, legal applications of deep learning have been on the rise. However, as with other high-stakes decision making areas, the requirement for interpretability is of crucial importance. Current models utilized by legal practitioners are more of the conventional machine learning type, wherein they are inherently interpretable, yet unable to harness the performance capabilities of data-driven deep learning models. In this work, we utilize deep learning models in the area of trademark law to shed light on the issue of likelihood of confusion between trademarks. Specifically, we introduce a model-agnostic interpretable intermediate layer, a technique which proves to be effective for legal documents. Furthermore, we utilize weakly supervised learning by means of a curriculum learning strategy, effectively demonstrating the improved performance of a deep learning model. This is in contrast to the conventional models which are only able to utilize the limited number of expensive manually-annotated samples by legal experts. Although the methods presented in this work tackles the task of risk of confusion for trademarks, it is straightforward to extend them to other fields of law, or more generally, to other similar high-stakes application scenarios.

Conversation disentanglement aims to separate intermingled messages into detached sessions, which is a fundamental task in understanding multi-party conversations.Existing work on conversation disentanglement relies heavily upon human-annotated datasets, which are expensive to obtain in practice.In this work, we explore to train a conversation disentanglement model without referencing any human annotations.Our method is built upon a deep co-training algorithm, which consists of two neural networks: a messagepair classifier and a session classifier.The former is responsible for retrieving local relations between two messages while the latter categorizes a message to a session by capturing context-aware information.Both networks are initialized respectively with pseudo data built from an unannotated corpus.During the deep co-training process, we use the session classifier as a reinforcement learning component to learn a session assigning policy by maximizing the local rewards given by the messagepair classifier.For the message-pair classifier, we enrich its training data by retrieving message pairs with high confidence from the disentangled sessions predicted by the session classifier.Experimental results on the large Movie Dialogue Dataset demonstrate that our proposed approach achieves competitive performance compared to the previous supervised methods.Further experiments show that the predicted disentangled conversations can promote the performance on the downstream task of multi-party response selection.

Exploiting label hierarchies has become a promising approach to tackling the zero-shot multi-label text classification (ZS-MTC) problem.Conventional methods aim to learn a matching model between text and labels, using a graph encoder to incorporate label hierarchies to obtain effective label representations (Rios and Kavuluru, 2018).More recently, pretrained models like BERT (Devlin et al., 2018) have been used to convert classification tasks into a textual entailment task (Yin et al., 2019).This approach is naturally suitable for the ZS-MTC task.However, pretrained models are underexplored in the existing work because they do not generate individual vector representations for text or labels, making it unintuitive to combine them with conventional graph encoding methods.In this paper, we explore to improve pretrained models with label hierarchies on the ZS-MTC task.We propose a Reinforced Label Hierarchy Reasoning (RLHR) approach to encourage interdependence among labels in the hierarchies during training.Meanwhile, to overcome the weakness of flat predictions, we design a rollback algorithm that can remove logical errors from predictions during inference.Experimental results on three reallife datasets show that our approach achieves better performance and outperforms previous non-pretrained methods on the ZS-MTC task.

Persona can function as the prior knowledge for maintaining the consistency of dialogue systems. Most of previous studies adopted the self persona in dialogue whose response was about to be selected from a set of candidates or directly generated, but few have noticed the role of partner in dialogue. This paper makes an attempt to thoroughly explore the impact of utilizing personas that describe either self or partner speakers on the task of response selection in retrieval-based chatbots. Four persona fusion strategies are designed, which assume personas interact with contexts or responses in different ways. These strategies are implemented into three representative models for response selection, which are based on the Hierarchical Recurrent Encoder (HRE), Interactive Matching Network (IMN) and Bidirectional Encoder Representations from Transformers (BERT) respectively. Empirical studies on the Persona-Chat dataset show that the partner personas neglected in previous studies can improve the accuracy of response selection in the IMN- and BERT-based models. Besides, our BERT-based model implemented with the context-response-aware persona fusion strategy outperforms previous methods by margins larger than 2.7% on original personas and 4.6% on revised personas in terms of [email protected] (top-1 accuracy), achieving a new state-of-the-art performance on the Persona-Chat dataset.

Dialogue disentanglement aims to separate intermingled messages into detached sessions. The existing research focuses on two-step architectures, in which a model first retrieves the relationships between two messages and then divides the message stream into separate clusters. Almost all existing work puts significant efforts on selecting features for message-pair classification and clustering, while ignoring the semantic coherence within each session. In this paper, we introduce the first end-to- end transition-based model for online dialogue disentanglement. Our model captures the sequential information of each session as the online algorithm proceeds on processing a dialogue. The coherence in a session is hence modeled when messages are sequentially added into their best-matching sessions. Meanwhile, the research field still lacks data for studying end-to-end dialogue disentanglement, so we construct a large-scale dataset by extracting coherent dialogues from online movie scripts. We evaluate our model on both the dataset we developed and the publicly available Ubuntu IRC dataset [Kummerfeld et al., 2019]. The results show that our model significantly outperforms the existing algorithms. Further experiments demonstrate that our model better captures the sequential semantics and obtains more coherent disentangled sessions.

Building explainable systems is a critical problem in the field of Natural Language Processing (NLP), since most machine learning models provide no explanations for the predictions. Existing approaches for explainable machine learning systems tend to focus on interpreting the outputs or the connections between inputs and outputs. However, the fine-grained information (e.g. textual explanations for the labels) is often ignored, and the systems do not explicitly generate the human-readable explanations. To solve this problem, we propose a novel generative explanation framework that learns to make classification decisions and generate fine-grained explanations at the same time. More specifically, we introduce the explainable factor and the minimum risk training approach that learn to generate more reasonable explanations. We construct two new datasets that contain summaries, rating scores, and fine-grained reasons. We conduct experiments on both datasets, comparing with several strong neural network baseline systems. Experimental results show that our method surpasses all baselines on both datasets, and is able to generate concise explanations at the same time.

Quote is a language phenomenon of transcribing the statement of someone else, such as a proverb and a famous saying. An appropriate usage of quote usually equips the expression with more elegance and credibility. However, there are times when we are eager to stress our idea by citing a quote, while nothing relevant comes to mind. Therefore, it is exciting to have a recommender system which provides quote recommendations while we are writing. This article extends previous study of quote recommendation, the task that recommends the appropriate quote according to the context (i.e., the content occurring before and after the quote). In this article, a quote recommender system called QuoteRec is presented to tackle the task. We investigate two models to learn the vector representations of quotes and contexts, and then rank the candidate quotes based on the representations. The first model learns the quote representation according to the contexts of a quote. The second model is an extension of the neural network model in previous study, which learns the representation of a quote by concerning both its content and contexts. Experimental results demonstrate the effectiveness of the two models in learning the semantic representations of quotes, and the neural network model achieves state-of-the-art results on the quote recommendation task.