论文检索

Xu Zhang, Hangcheng Liu, Shangwei Guo, Shudong Zhang, Tianwei Zhang, Tao Xiang

Recent studies reveal that adversaries can manipulate the internal knowledge of large language models (LLMs) on selected topics through model editing, causing attacker-specified harmful or biased outputs when queried about the edited content. Once such tampered LLMs are distributed, they can mislead users on the targeted topics, thereby potentially propagating misinformation or reinforcing stereotypes. However, existing knowledge manipulation attacks rely on the ability to redistribute compromised models, which is infeasible in constrained settings like Federated Instruction Tuning (FedIT), where a central server controls LLM's training and distribution. In this work, we introduce ShadeEdit, the first attack framework that leverages strengthened model editing to enable knowledge manipulation in FedIT scenarios. ShadeEdit introduces two key components to address two challenges posed by the training process of FedIT: (1) a paraphrase-based editing dataset selection strategy to mitigate the dilution from benign updates on malicious ones by constructing a high-quality editing dataset, and (2) an adaptive manipulation mechanism to evade aggregation-based defenses via an adaptive clipping strategy. ShadeEdit achieves an average 99.5% attack success rate over eight robust aggregation algorithms while preserving instruction-following accuracy, demonstrating its strong attack effectiveness and model-utility preservation.

Xin Zhang, Victor S. Sheng

Despite the rapid progress in large language models (LLMs), even sub-billion-scale systems perform at chance level on challenging natural language inference (NLI) benchmarks such as Adversarial Natural Language Inference (ANLI), while training larger models is often impractical due to limited computational resources. We address this parameter-efficiency bottleneck in NLI with a Complex-Vector Token Representation that explicitly decouples each token from its context, and a Token-Context Attention mechanism that updates each token based on the most informative contextual semantics. On ANLI, a 0.8B-parameter Token-Context Attention model achieves higher parameter efficiency (accuracy per parameter) than all 1B and comparable 0.8B self-attention baselines; it also suffers smaller performance degradation under Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) attacks and achieves the largest few-shot gains on SNLI and MNLI while exhibiting no significant degradation in ANLI accuracy after adaptation. These results suggest that explicitly disentangling token and context offers a viable alternative to standard self-attention for NLI tasks.

Xiaoyun Zhang, Zhengyue Zhao, Wenxuan Shi, Kaidi Xu, Di Huang, Xing Hu

With the widespread application of Large Language Models (LLMs), it has become a significant concern to ensure their safety and prevent harmful responses. While current safe-alignment methods based on instruction fine-tuning and Reinforcement Learning from Human Feedback (RLHF) can effectively reduce harmful responses from LLMs, they often require high-quality datasets and heavy computational overhead during model training. Another way to align language models is to modify the logit of tokens in model outputs without heavy training. Recent studies have shown that contrastive decoding can enhance the performance of language models by reducing the likelihood of confused tokens. However, these methods require the manual selection of contrastive models or instruction templates, limiting the degree of contrast. To this end, we propose Adversarial Contrastive Decoding (ACD), an optimization-based framework to generate two opposite soft system prompts, the Safeguarding Prompt (SP) and the Adversarial Prompt (AP), for prompt-based contrastive decoding. The SP aims to promote safer outputs while the AP aims to exploit the harmful parts of the model, providing a strong contrast to align the model with safety. ACD only needs to apply a lightweight prompt tuning on a rather small anchor dataset without training the target model. Experiments conducted on extensive models and benchmarks demonstrate that the proposed method achieves much better safety performance than previous model training-free decoding methods without sacrificing its original generation ability.

Xiang Zhang, Rui Xie, Shikun Zhang

Fine-tuning large language models (LLMs) in a parameter-efficient manner while preserving their pre-trained world knowledge remains a significant challenge. While Low-Rank Adaptation (LoRA) and its variants effectively mitigate catastrophic forgetting, they do not fully eliminate the loss of critical pre-trained knowledge. In this work, we first analyze the layer-wise distribution of domain-specific knowledge within LLMs through knowledge localization, and empirically identify a clear layer-specific pattern: pre-trained world knowledge predominantly resides in lower layers, whereas knowledge relevant to downstream tasks is more concentrated in higher layers. Motivated by this observation, we propose L2-LoRA, a simple yet effective variant of LoRA that applies layer-specific L2 regularization to the LoRA weights during fine-tuning. Specifically, L2-LoRA imposes stronger regularization on lower layers to preserve pre-trained world knowledge, while allowing greater adaptation in higher layers to better align with downstream tasks. Experiments across multiple benchmarks show that L2-LoRA not only consistently outperforms vanilla LoRA in downstream performance, but also effectively mitigates catastrophic forgetting by retaining more pre-trained knowledge.

Shuyi Zhang, Wei Shi, Sihang Li, Jiayi Liao, Tao Liang, Hengxing Cai, Xiang Wang

Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-Enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models.

Shimao Zhang, Zhejian Lai, Xiang Liu, Shuaijie She, Xiao Liu, Yeyun Gong, Shujian Huang, Jiajun Chen

Multilingual Alignment is an effective and representative paradigm to enhance LLMs' multilingual capabilities, which transfers the capabilities from the high-resource languages to the low-resource languages. Meanwhile, some research on language-specific neurons provides a new perspective to analyze and understand LLMs' mechanisms. However, we find that there are many neurons that are shared by multiple but not all languages and cannot be correctly classified. In this work, we propose a ternary classification methodology that categorizes neurons into three types, including language-specific neurons, language-related neurons, and general neurons. And we propose a corresponding identification algorithm to distinguish these different types of neurons. Furthermore, based on the distributional characteristics of different types of neurons, we divide the LLMs' internal process for multilingual inference into four parts: (1) multilingual understanding, (2) shared semantic space reasoning, (3) multilingual output space transformation, and (4) vocabulary space outputting. Additionally, we systematically analyze the models before and after alignment with a focus on different types of neurons. We also analyze the phenomenon of ''Spontaneous Multilingual Alignment''. Overall, our work conducts a comprehensive investigation based on different types of neurons, providing empirical results and valuable insights to better understand multilingual alignment and multilingual capabilities of LLMs.

Qingyu Zhang, Chunlei Xin, Xuanang Chen, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun, Qing Ye, Qianlong Xie, Xingxing Wang

Goal-driven persuasive dialogue, exemplified by applications like telemarketing, requires sophisticated multi-turn planning and strict factual faithfulness, which remains a significant challenge for even state-of-the-art Large Language Models (LLMs). A lack of task-specific data often limits previous works, and direct LLM application suffers from strategic brittleness and factual hallucination. In this paper, we first construct and release TeleSalesCorpus, the first real-world-grounded dialogue dataset for this domain. We then propose AI-Salesman, a novel framework featuring a dual-stage architecture. For the training stage, we design a Bayesian-supervised reinforcement learning algorithm that learns robust sales strategies from noisy dialogues. For the inference stage, we introduce the Dynamic Outline-Guided Agent (DOGA), which leverages a pre-built script library to provide dynamic, turn-by-turn strategic guidance. Moreover, we design a comprehensive evaluation framework that combines fine-grained metrics for key sales skills with the LLM-as-a-Judge paradigm. Experimental results demonstrate that our proposed AI-Salesman significantly outperforms baseline models in both automatic metrics and comprehensive human evaluations, showcasing its effectiveness in complex persuasive scenarios.

Mingjie Zhang, Xiaoling Zhou, Yuxiao Luo, Yiyu Liu, Shikun Zhang, Wei Ye

Knowledge distillation (KD) is a widely adopted technique for transferring the capabilities of large teacher models to smaller student models, thereby significantly reducing inference costs and memory consumption. However, existing KD methods are all constrained by an inherent greedy optimization objective, rooted in the assumption of teacher superiority: "Trust all teacher-generated outputs (TGOs)" and "Distrust any student-generated outputs (SGOs) unsupported by the teacher". We propose ASKD, a novel KD method with adaptive skewness determined by sample quality, refining this objective to: "Learn TGOs proportionally to their quality, and distrust only low-quality unsupported SGOs". ASKD comprises three key components: (1) A reinforcement learning-style optimization formulation to mitigate the inherent approximation bias in sample-based Kullback-Leibler (KL) divergence approximations used by previous KD methods; (2) Well-designed quality supervision signals to map and achieve adaptive skewness in skewed KL loss, pioneering the usage of sample quality to adjust learning magnitudes; (3) A gradient-clip function on high-quality SGOs for findings that high-quality SGOs in KL loss fail to yield positive updates and even cause adverse effects on some samples. Extensive experiments indicate that ASKD builds high-performance student models across various tasks, including instruction following, mathematical reasoning, and code generation, outperforming state-of-the-art methods comprehensively and surpassing GRPO-like approaches that use advantages as multiplicative factors. We also provide detailed mathematical proofs demonstrating properties such as Lipschitz continuity of the update coefficient and uniform convergence of the loss function, ensuring theoretical rigor for key components of ASKD.

Min Zhang, Yuzhe Lu, Yun Zhou, Panpan Xu, Lin Lee Cheong, Chang-Tien Lu, Haozhu Wang

Numerical reasoning over documents, which demands both contextual understanding and logical inference, is challenging for low-capacity local models deployed on computation-constrained devices. Although such complex reasoning queries could be routed to powerful remote models like GPT-4, exposing local data raises significant data leakage concerns. Existing mitigation methods generate problem descriptions or examples for remote assistance. However, the inherent complexity of numerical reasoning hinders the local model from generating logically equivalent queries and accurately inferring answers with remote guidance. In this paper, we present a model collaboration framework with two key innovations: (1) a context-aware synthesis strategy that shifts the query topics while preserving reasoning patterns; and (2) a tool-based answer reconstruction approach that reuses the remote-generated plug-and-play solution with code snippets. Experimental results demonstrate that our method achieves better reasoning accuracy than solely using local models while providing stronger data protection than fully relying on remote models. Furthermore, our method improves accuracy by 16.2% - 43.6% while reducing data leakage by 2.3% - 44.6% compared to existing data protection approaches.

Mengfei Zhang, Zhenglin Wang

While Large Reasoning Models (LRMs) exhibit remarkable capabilities in complex tasks, they often suffer from excessive redundancy in their chain-of-thought reasoning. This significantly reduces inference efficiency and increases computational costs. We identify that LRM redundancy is not uniformly homogeneous but can be taxonomized according to whether it is destructive to the final answer: destructive redundancy (e.g., logical drift, hallucination amplification) versus non-destructive redundancy (e.g., repetition, over-elaboration). Moreover, LRM's redundant and concise responses exhibit a significant distinction in their hidden layer representation spaces. Based on these insights, we propose CATS (Category-Aware Token-level Steering), a training-free and lightweight method to reduce the redundancy phenomenon. CATS decomposes redundancy into six semantically interpretable characteristic dimensions. By flexibly weighting and combining the differential vectors corresponding to these dimensions, CATS synthesizes a composite intervention vector, enabling zero-parameter intervention in the hidden layers. Experiments across three LRM models and five mathematical reasoning datasets demonstrate that CATS reduces reasoning length by an average of 25% while maintaining or even slightly improving task accuracy. CATS offers a pluggable, training-free, and lightweight solution, making it particularly beneficial for users in low-resource environments.

Lanxue Zhang, Yuqiang Xie, Fang Fang, Fanglong Dong, Rui Liu, Yanan Cao

Large Language Models demonstrate strong reasoning capabilities, which can be effectively compressed into smaller models. However, existing datasets and fine-tuning approaches still face challenges that lead to catastrophic forgetting, particularly for models smaller than 8B. First, most datasets typically ignore the relationship between training data knowledge and the model's inherent abilities, making it difficult to preserve prior knowledge. Second, conventional training objectives often fail to constrain inherent knowledge preservation, which can result in forgetting of previously learned skills. To address these issues, we propose a comprehensive solution that alleviates catastrophic forgetting from both the data and fine-tuning approach perspectives. On the data side, we construct a dataset of 5K instances that covers multiple reasoning tasks and incorporates metacognitive knowledge, making it more tolerant and effective for distillation into smaller models. We annotate the metacognitive knowledge required to solve each question and filter the data based on task knowledge and the model's inherent skills. On the training side, we introduce GDPO (Group Direction Preference Optimization), which is better suited for resource-limited scenarios and can efficiently approximate the performance of GRPO. Guided by the large model and by implicitly constraining the optimization path through a reference model, GDPO enables more effective knowledge transfer from the large model and constrains excessive parameter drift. Extensive experiments demonstrate that our approach significantly alleviates catastrophic forgetting and improves reasoning performance on smaller models.

Jusheng Zhang, Ningyuan Liu, Yijia Fan, Zihao Huang, Qinglin Zeng, Kaitong Cai, Jian Wang, Keze Wang

Large language models (LLMs) often generate hallucinated content lacking factual or contextual grounding, hindering their reliability in critical applications. Traditional methods like supervised fine-tuning and reinforcement learning from human feedback are data-intensive and computationally expensive, while static parameter editing struggles with context-dependent errors and catastrophic forgetting. To overcome these limitations, we introduce LLM-CAS, a framework that formulates real-time hallucination correction as a hierarchical reinforcement learning (HRL) problem. LLM-CAS trains an agent to learn a sophisticated policy, dynamically selecting optimal, temporary neuron perturbations during inference based on the immediate context. This learned, policy-driven approach provides greater adaptability than prior dynamic methods that rely on heuristic or pre-defined adjustments. As a result, LLM-CAS achieves significant performance gains across various LLMs, improving accuracy by 10.98 percentage points on StoryCloze, 2.71 points on TriviaQA, and 2.06 points on TruthfulQA's MC1 score, thereby outperforming static methods like ITI and CAA, as well as the dynamic SADI framework. This context-aware, efficient approach promises enhanced reliability for LLMs in high-stakes domains, with future potential for multimodal extensions.

Junan Zhang, Xueyao Zhang, Jing Yang, Yuancheng Wang, Fan Fan, Zhizheng Wu

Recent generative models have significantly advanced speech restoration tasks, yet their training objectives often misalign with human perceptual preferences, resulting in suboptimal quality. While post-training alignment has proven effective in other generative domains like text and image generation, its application to generative speech restoration remains largely under-explored. This work investigates the challenges of applying preference-based post-training to this task, focusing on how to define a robust preference signal and curate high-quality data to avoid reward hacking. To address these challenges, we propose a multi-metric preference alignment strategy. We construct a new dataset, GenSR-Pref, comprising 80K preference pairs, where each chosen sample is unanimously favored by a complementary suite of metrics covering perceptual quality, signal fidelity, content consistency, and timbre preservation. This principled approach ensures a holistic preference signal. Applying Direct Preference Optimization (DPO) with our dataset, we observe consistent and significant performance gains across three diverse generative paradigms: autoregressive models (AR), masked generative models (MGM), and flow-matching models (FM) on various restoration benchmarks, in both objective and subjective evaluations. Ablation studies confirm the superiority of our multi-metric strategy over single-metric approaches in mitigating reward hacking. Furthermore, we demonstrate that our aligned models can serve as powerful ''data annotators'', generating high-quality pseudo-labels to serve as a supervision signal for traditional discriminative models in data-scarce scenarios like singing voice restoration.

Jie Zhang, Bo Tang, Wanzi Shao, Wenqiang Wei, Jihao Zhao, Jianqing Zhu, Zhiyu Li, Wen Xi, Zehao Lin, Feiyu Xiong 等

Retrieval-Augmented Generation (RAG) improves large language models by retrieving external knowledge, often truncated into smaller chunks due to the input context window, which leads to information loss, resulting in response hallucinations and broken reasoning chains. Moreover, traditional RAG retrieves unstructured knowledge, introducing irrelevant details that hinder accurate reasoning. To address these issues, we propose TAdaRAG, a novel RAG framework for on-the-fly task-adaptive knowledge graph construction from external sources. Specifically, we design an intent-driven routing mechanism to a domain-specific extraction template, followed by supervised fine-tuning and a reinforcement learning-based implicit extraction mechanism, ensuring concise, coherent, and non-redundant knowledge integration. Evaluations on six public benchmarks and a real-world business benchmark (NowNewsQA) across three backbone models demonstrate that TAdaRAG outperforms existing methods across diverse domains and long-text tasks, highlighting its strong generalization and practical effectiveness.

Jie Zhang, Chenxu Niu, Zhefeng Nan, Yangyan Xu, Jinta Weng

Privacy concerns have long been a critical issue in AI models. With the rapid advancement of generative AI, the privacy awareness of models has drawn attention, raising new challenges for privacy protection that is independent of data and tasks. This paper introduces a novel framework for enhancing privacy protection through directional steering in representation space, which seamlessly integrates with both language and vision-language models. Specifically, we first construct a comprehensive privacy-related dataset based on the Solove taxonomy of privacy. Then, we leverage this dataset to enhance model privacy awareness in the representation space, steering the model to protect privacy during inference. Experiments on 12 models validate the effectiveness and generalization of our method. Moreover, we demonstrate the transferability of privacy-enhanced representations between same-source large language models (LLMs) and vision-language models (VLMs), offering a scalable solution for privacy protection in frontier AI models.

Jiayuan Zhang, Kaiquan Chen, Zhihao Lu, Enshen Zhou, Qian Yu, Jing Zhang

Web automation uses intelligent agents to perform high-level tasks by mimicking human interactions with webpages. Despite recent advances in LLM-based web agents, efficiently navigating complex, real-world webpages remains challenging due to massive DOM structures (10,000 ~ 100,000 tokens). Current approaches either truncate DOMs—losing vital information—or use inefficient heuristics and separate ranking models, failing to balance precision and scalability. We introduce Prune4Web, a novel paradigm that transforms DOM processing from LLM-based filtering to programmatic pruning. Our key innovation is DOM Tree Pruning Programming, where an LLM generates executable Python scoring programs to dynamically filter DOM elements based on semantic clues from decomposed sub-tasks. This approach eliminates the need for LLMs to process full DOMs, instead delegating traversal and scoring to lightweight, interpretable programs. The result is a 25 ~ 50 times reduction in candidate elements for grounding, enabling precise action localization without attention dilution. Additionally, we propose a data annotation method and a two-turn dialogue training strategy that jointly optimizes Planner, Programmatic Filter, and Grounder in a unified framework. Experiments demonstrate state-of-the-art performance. On our low-level task grounding task, our approach dramatically increases grounding accuracy from 46.80% to 88.28%, highlighting its effectiveness.

Jianqing Zhang, Wei Xia, Hande Dong, Qiang Lin, Jian Cao

LLM's code generation capabilities have yielded substantial improvements in the effectiveness of programming tasks. However, LLM-generated code still suffers from compilation and runtime errors. Existing offline preference optimization methods primarily focus on enhancing LLMs' coding abilities using pass/fail signals in the preference data, overlooking the deep-level error types in the failed codes. To address this, we propose Adaptively Progressive Preference Optimization (AP2O) for coding (i.e., AP2O-Coder), a method that guides LLMs adaptively and methodically to reduce code errors for code generation. Specifically, we construct an error notebook from failed codes and progressively optimize the LLM to correct errors type by type. Furthermore, we adaptively replay error types to tailor to the LLM's evolving weaknesses throughout training. Through extensive experiments on both code and general LLMs (Llama, Qwen, and DeepSeek series) with parameters ranging from 0.5B to 34B, our AP2O-Coder improves code generation performance by up to 3% in pass@k while using less preference data.

Hui Zhang, Po Hu, Wei Emma Zhang

The proliferation of multi-modal data on the internet has intensified the need for structured event understanding across textual and visual modalities. However, existing multi-modal event extraction models suffer from three major limitations: the absence of explicit event schema guidance, coarse-grained multi-modal alignment strategies, and reliance on heterogeneous, misaligned multi-modal training datasets. To address these issues, we propose LLaVA-MS-PIT, a Multi-modal Schema-Guided Progressive Instruction Tuning Framework that explicitly injects structured multi-modal event schema knowledge into the model before event extraction. Specifically, we introduce the textual event schema to establish the model’s prior knowledge of event concepts and enhance its ability to reason about event structures, while the visual event schema is employed to bridge the representation gap between textual and visual modalities at the event level, enabling unified and semantically aligned event representations across modalities. Moreover, to alleviate data scarcity and modality misalignment inherent in current benchmarks, we construct imSitu-MEE, a high-quality multi-modal parallel dataset generated and annotated through schema-guided procedures. Extensive experiments demonstrate that LLaVA-MS-PIT achieves competitive performance on multi-modal event extraction benchmarks, underscoring the effectiveness and necessity of schema-guided progressive instruction tuning.

Hao Zhang, Yaqi Chen, Nianwen Si, XuKui Yang, Wenlin Zhang, Dan Qu

Recently, End-to-End Speech Translation (E2E-ST) methods leveraging large language models (LLMs) have demonstrated strong generalization capabilities and excellent scalability by integrating pre-trained speech encoders with LLMs, where Low-Rank Adaptation (LoRA) is commonly used for parameter-efficient fine-tuning to reduce training costs. However, LoRA's low-rank assumption often fails in multilingual tasks, as the inherent complexity of cross-lingual semantic relationships and syntactic variations exceeds the representational capacity of low-rank matrices. This leads to parameter conflicts across languages, resulting in suboptimal performance. To address this issue, we propose Mixture of Low-Rank Adaptations (MoLoRA), which integrates the Mixture of Experts (MoE) mechanism with LoRA. MoLoRA effectively enhances the model's expressive capacity while maintaining parameter efficiency during training. Specifically, we treat multiple LoRA modules as low-rank experts and introduce a routing mechanism to dynamically activate language-specific experts. Additionally, shared experts are incorporated and consistently activated to model cross-lingual general knowledge. Furthermore, to enhance the robustness and accuracy of speech representations, we propose a Multi-Granularity Representation Fusion module (MGRF). This module mitigates local distortions in frame-level speech representations caused by noise by fusing frame-level and sentence-level features, thereby providing the LLM with more accurate high-level semantic information. We conduct multilingual experiments on the MuST-C and CoVoST-2 datasets. Our method achieves an average BLEU score of 32.2 across eight language pairs on the MuST-C dataset and an average of 36.3 across three language pairs on the CoVoST-2 dataset, establishing a new state-of-the-art (SOTA) performance.

Guo-Biao Zhang, Zhijing Wu, Tian Lan, Ding-Yuan Liu, Yu-Shi Zhu, Xian-Ling Mao

As a knowledge-intensive and challenging task, automatic generation of long-form wiki-style articles has garnered increasing attention from researchers due to its ability to efficiently integrate, organize and present vast amounts of both structured and unstructured knowledge. To the best of our knowledge, most of the existing mainstream state-of-the-art methods for automatic wiki-style article generation typically follow a "one-shot generation" paradigm: given a topic, (1) first generating a structured outline, (2) then independently and in parallel generating the content of each outline chapter in a one-shot using the chapter title and references. However, the core limitation of the paradigm lies in its disregards inter-chapter correlation and lacks post-generation revision and refinement, resulting in content redundancy, weak relevance and logical inconsistency. To address these issues, we propose WikiREVIEW, a novel multi-perspective review framework for automatic wiki-style article generation. Specifically, our proposed method introduces multi-perspective experts to review the content of each outline chapter at both chapter and paragraph levels following the initial generation, offering evaluation feedback and continuously refining the numerous deficiencies in the initial long-form article, ultimately achieving high-quality wiki-style article generation. Extensive experimental results on the public English dataset FreshWiki and our own constructed high-quality Chinese dataset ChineseWiki, demonstrate that our proposed WikiREVIEW significantly outperforms existing state-of-the-art automatic wiki-style article generation methods across all automatic evaluation metrics and human evaluation.