论文检索

Jiabei He, Yanzhe Zhang, Jiaming Zhou, Hui Wang, Haoqin Sun, Yong Qin

Speech signals convey abundant speaker-related metadata, yet current privacy research predominantly focuses on identity-centric voiceprint protection, leaving sensitive Speaker Attribute Privacy (SAP) largely underexplored. This paper introduces AudioPrivacy, a large-scale Chinese dataset designed to systematically evaluate SAP leakage in realistic, everyday scenarios. Comprising 227.3 hours of audio from 1,000 speakers, it uniquely encompasses four parallel modalities: speech, singing, paralinguistic expressions, and non-vocal acoustic signals (e.g., footsteps). Annotated with 11 diverse attributes, including fine-grained physiological traits often overlooked in traditional corpora, AudioPrivacy enables a granular analysis of acoustic privacy risks. Our evaluations reveal significant leakage across multiple attributes, even when inferred from non-vocal signals. Furthermore, we demonstrate that state-of-the-art Multimodal Large Language Models (MM LLMs) can precisely profile speakers and exacerbate these risks, underscores the urgent need to rethink privacy-preserving mechanisms in the era of powerful audio foundation models.

Zihao Yi, Qingxuan Jiang, Ruotian Ma, Xingyu Chen, Qu Yang, Mengru Wang, Fanghua Ye, Ying Shen, Zhaopeng Tu, Xiaolong Li 等

Large Language Models (LLMs) are increasingly tasked with creative generation, including the simulation of fictional characters. However, their ability to portray non-prosocial, antagonistic personas remains largely unexamined. We hypothesize that the safety alignment of modern LLMs creates a fundamental conflict with the task of authentically role-playing morally ambiguous or villainous characters. To investigate this, we introduce the Moral RolePlay benchmark, a new dataset featuring a four-level moral alignment scale and a balanced test set for rigorous evaluation. We task state-of-the-art LLMs with role-playing characters from moral paragons to pure villains. Our large-scale evaluation reveals a consistent, monotonic decline in role-playing fidelity as character morality decreases. We find that models struggle most with traits directly antithetical to safety principles, such as ”Deceitful” and ”Manipulative”, often substituting nuanced malevolence with superficial aggression. Furthermore, we demonstrate that general chatbot proficiency is a poor predictor of villain role-playing ability, with highly safety-aligned models performing particularly poorly. Our work provides the first systematic evidence of this critical limitation, highlighting a key tension between model safety and creative fidelity. Our benchmark and findings pave the way for developing more nuanced, context-aware alignment methods.

Hongyuan Yuan, Xinran He, Run Shao, Bolei He, Xianwei Xue, Mengke Chen, Qiutong Pan, Haiwei Wang, Haifeng Li

Extending CoT through RL has been widely used to enhance the reasoning capabilities of LLMs. However, due to the sparsity of reward signals, it can also induce undesirable thinking patterns such as overthinking, i.e., generating redundant intermediate reasoning content. In this work, we argue that a major source of such redundancy is inefficient reflection, which often manifests in two problematic patterns: Indiscriminate Reflection, where the model performs broad, low-impact checks throughout reasoning, and Repetitive Reflection, where it repeatedly re-verifies an already established conclusion. To address this, we introduce a graph-based CoT optimization framework. Specifically, we convert each linear CoT into a directed acyclic graph (DAG) with explicit dependency edges, and design a dual pruning strategy: branch-level pruning removes weakly contributing reflection branches, while depth-level pruning eliminates late-stage re-verification. We distill this behavior via a three-stage pipeline: (1) SFT to initialize the policy on pruned concise traces, (2) DPO to prefer correct but less redundant trajectories, and (3) GRPO with length penalty to jointly optimize answer correctness and efficiency. Experiments show that our approach reduces the average reasoning tokens by 42% while maintaining or improving accuracy.

Junbo Fu, Guoshuai Zhao, Linkang Yang, Yunqi Mi, Xueming Qian

The increasing integration of large language models (LLMs) in code generation has raised critical copyright concerns, particularly regarding the verbatim repetition of copyrighted code. To address this challenge, we propose a novel task: Duplicate-Aware Controlled Code Generation (DACCG), which aims to mitigate verbatim repetition while preserving the quality of generated code. To this end, we introduce Targeted Reordering Beam Search (TRBS), a plug-and-play decoding method that dynamically reorders beam candidates to reduce direct copying. TRBS leverages the FM-index for efficient substring detection and employs a spike-entropy-based protection mechanism to safeguard structural anchors critical to code coherence. Experimental results on a multi-language code generation benchmark demonstrate that TRBS effectively reduces verbatim repetition while maintaining functional adequacy. Our research represents a pioneering effort in code copyright protection from the model user’s perspective, offering novel insights into responsible code generation practices.

Xiaoyu Hu, Jinman Zhao

Large language models (LLMs) are increasingly deployed in financial contexts, raising critical concerns about reliability, alignment, and susceptibility to adversarial manipulation. While prior finance-related benchmarks assess LLMs’ capabilities in stock trading, they are often restricted to small sample and fail to demonstrate LLM susceptibility to context with potential human bias. We introduce Fin-Herding (financial herding under long and uncertain financial context), a benchmark for evaluating LLM investment decision-making when faced with uncertainty and possible human-biased opinions. Fin-Herding includes 8868 long firm-specific analyst reports, including firm aspects summarized and analyzed by sophisticated analysts with investment ratings (Bullish/Neutral/Bearish) spanning from various industries. We present large language models with firm analyst reports with/without analyst investment ratings and even with ’fake’ rating, to get investment ratings generated by LLMs. Our results reveal that LLMs tend to herd the explicit bias in context. We also develop a method to detect potential human opinions, which can encourage LLMs to think independently, some models even exceed human performance in predicting future stock return.

Chenye Zou, Difan Jiao, Lijie Hu

Large Language Models (LLMs) are increasingly deployed worldwide, yet they exhibit strong Western-centric biases, and the internal mechanisms governing their cultural behaviors remain poorly understood. Prior work has identified so-called cultural neurons, but individual neurons are often polysemous, conflating abstract cultural knowledge with surface-level lexical cues due to superposition. We apply Sparse Autoencoders (SAEs) to decompose intermediate LLM activations into sparse, interpretable feature representations that disentangle these factors. This analysis reveals culturally selective features that remain invariant across paraphrasing and task formats, indicating abstraction beyond lexical correlations. Through targeted feature ablation, we provide causal evidence that these features are necessary for cultural reasoning: their removal selectively degrades performance on culturally conditioned tasks. Furthermore, we show that steering model activations along these feature directions is sufficient to systematically modulate cultural-related knowledge generation, without retraining. Together, our results offer the first causal evidence that LLMs encode cultural knowledge as decoupled semantic structures rather than surface patterns, enabling a scalable pathway toward cultural alignment through mechanistic intervention. Code is available at https://github.com/IAN-YE/Cultural-features-SAE.

Jiajun Wu, Jian Yang, Wei Zhang, Linzheng Chai, Yuchi Ma, Ensheng Shi, Yuqing Ma, Zhoujun Li, Xianglong Liu

Large language models (LLMs) have demonstrated remarkable capabilities in code generation tasks. However, their effectiveness heavily relies on supervised training with extensive labeled (e.g., question-answering pairs) or unlabeled datasets (e.g., code snippets), which are often expensive and difficult to obtain at scale. To address this limitation, this paper introduces a method IPC, an unsupervised framework that leverages Internal Probing of LLMs for Code generation without any external corpus, even unlabeled code snippets. We introduce the problem space probing, test understanding probing, solution space probing, and knowledge consolidation and reinforcement to probe the internal knowledge and confidence patterns existing in LLMs. Further, IPC identifies reliable code candidates through self-consistency mechanisms and representation-based quality estimation to train UCoder (coder with unsupervised learning). We validate the proposed approach across multiple code benchmarks, demonstrating that unsupervised methods can achieve competitive performance compared to supervised approaches while significantly reducing the dependency on labeled data and computational resources. Analytic experiments reveal that internal model states contain rich signals about code quality and correctness, and that properly harnessing these signals enables effective unsupervised learning for code generation tasks, opening new directions for training code LLMs in resource-constrained scenarios.

Wei Zhang, Jian Yang, Renshuai Tao, Linzheng Chai, Shuyue Guo, Jiajun Wu, Xiaoming Chen, Ganqu Cui, Ning Ding, Xander Xu 等

Code large language models have demonstrated remarkable capabilities in programming tasks, yet current benchmarks primarily focus on single modality rather than visual game development. Most existing code-related benchmarks evaluate syntax correctness and execution accuracy, overlooking critical game-specific metrics such as playability, visual aesthetics, and user engagement that are essential for real-world deployment. To address the gap between current LLM capabilities in algorithmic problem-solving and competitive programming versus the comprehensive requirements of practical game development, we present V-GameGym, a comprehensive benchmark comprising 2,219 high-quality samples across 100 thematic clusters derived from real-world repositories, adopting a novel clustering-based curation methodology to ensure both diversity and structural completeness. Further, we introduce a multimodal evaluation framework with an automated LLM-driven pipeline for visual code synthesis using complete UI sandbox environments. Our extensive analysis reveals that V-GameGym effectively bridges the gap between code generation accuracy and practical game development workflows, providing quantifiable quality metrics for visual programming and interactive element generation.

Soyeon Kim, Cheongwoong Kang, Myeongjin Lee, Eun-Chul Chang, Lee Jaedeok, Jaesik Choi

The development of practical (multimodal) large language model assistants for Korean weather forecasters is hindered by the absence of a multidimensional, expert-level evaluation framework grounded in authoritative sources. To address this, we introduce K-MetBench, a diagnostic benchmark grounded in national qualification exams. It exposes critical gaps across four dimensions: expert visual reasoning of charts, logical validity via expert-verified rationales, Korean-specific geo-cultural comprehension, and fine-grained domain analysis. Our evaluation of 55 models reveals a profound *modality gap* in interpreting specialized diagrams and a *reasoning gap* where models hallucinate logic despite correct predictions. Crucially, Korean models outperform significantly larger global models in local contexts, demonstrating that parameter scaling alone cannot resolve cultural dependencies. K-MetBench serves as a roadmap for developing reliable, culturally aware expert AI agents.

Donghang Duan, Xu Zheng, Yuefeng He, Chong Mu, Leyi Cai, Lizong Zhang

Current LLM-based frameworks for text anonymization usually rely on remote API services from powerful LLMs, which creates an inherent privacy paradox: users must disclose the raw data to untrusted third parties for guaranteed privacy preservation. Moreover, directly migrating current solutions to local small-scale models (LSMs) offers a suboptimal solution with severe utility collapse. Our work argues that this failure stems not merely from the capability deficits of LSMs, but significantly from the inherent irrationality of the greedy adversarial strategies employed by current state-of-the-art (SOTA) methods. To address this, we propose Rational Localized Adversarial Anonymization (RLAA), a fully localized and training-free framework featuring an Attacker-Arbitrator-Anonymizer architecture. We model the anonymization process as a trade-off between Marginal Privacy Gain (MPG) and Marginal Utility Cost (MUC), demonstrating that greedy strategies tend to drift into an irrational state. Instead, RLAA introduces an arbitrator that acts as a rationality gatekeeper, validating the attacker’s inference to filter out ghost leaks. This mechanism promotes a rational early-stopping criterion, and structurally prevents utility collapse. Extensive experiments on different benchmarks demonstrate that RLAA achieves a superior privacy-utility trade-off compared to strong baselines.

Xiaofan Zheng, Xinghao Wang, Xiaojun Wan

The proliferation of Large Vision-Language Models (LVLMs) has exacerbated concerns regarding model misappropriation and license violations. Malicious users may deploy open-source models as black boxes and falsely claim ownership, sparking significant community interest in fingerprinting techniques for copyright authentication. Current fingerprinting methods largely follow a backdoor-based paradigm, employing specific inputs to elicit predetermined abnormal text outputs. However, such direct distortion of the model’s original predictions compromises modality alignment and inevitably degrades multimodal capabilities, leading to an inherent trade-off between robustness and harmlessness. To address these challenges, we investigate whether it is possible to embed robust fingerprints while maximally preserving the original normal outputs of the model. We propose a Synonym-Aware Logit Shaping Fingerprint (SALSF). The core insight of SALSF lies in reshaping the probability distribution of semantically similar long-tail tokens within the logits space while ensuring the original top-1 prediction token and its probability remain approximately invariant. By elevating the overall prediction probability of the semantic cluster to a level distinctly higher than the natural baseline, our approach stealthily embeds the fingerprint and mitigates the disruption to modality alignment. Experimental results demonstrate that SALSF maintains multimodal performance and substantially enhances fingerprint robustness, offering a novel paradigm for the intellectual property protection of LVLMs.

Jie Sun, Yu Liu, Lu Han, Qiwen Deng, Xiang Shu, Yang Xiao, Lintao Ma, Xingyu Lu, Jun Zhou, Pengfei Liu 等

While transformer-based Large Language Models (LLMs) theoretically support massive context windows, they suffer from severe performance degradation when processing long numerical sequences. We attribute this failure to the attention dispersion in the Softmax mechanism, which prevents the model from concentrating attention. To overcome this, we propose Separate Sequence (SepSeq), a training-free, plug-and-play framework to mitigate dispersion by strategically inserting separator tokens. Mechanistically, we demonstrate that separator tokens act as an attention anchor, recalibrating attention to focus on local segments while preserving global context. Extensive evaluations on 9 widely-adopted LLMs confirm the effectiveness of our approach: SepSeq yields an average relative accuracy improvement of 35.6% across diverse domains while reducing 16.4% inference token consumption.

Tianyu Pan, Damon L. Woodard

Large language models (LLMs) demonstrate strong performance, but they often lack transparency. We introduce GeoLAN, a training framework that treats token representations as geometric trajectories and applies stickiness conditions inspired by recent developments related to the Kakeya Conjecture. We have developed two differentiable regularizers, Katz-Tao Convex Wolff (KT-CW) and Katz-Tao Attention (KT-Attn), that promote isotropy and encourage diverse attention. Our experiments with Gemma-3 (1B, 4B, 12B) and Llama-3-8B show that GeoLAN frequently maintains task accuracy while improving geometric metrics and reducing certain fairness biases. These benefits are most significant in mid-sized models. Our findings reveal scale-dependent trade-offs between geometric precision and performance, suggesting that geometry-aware training is a promising approach to enhance mechanistic interpretability.

Haidong Xin, Xinze Li, Zhenghao Liu, Yukun Yan, Shuo Wang, Cheng Yang, Yu Gu, Ge Yu, Maosong Sun

Existing memory systems enable Large Language Models (LLMs) to support long-horizon human-LLM interactions by persisting historical interactions beyond limited context windows. However, while recent approaches have succeeded in constructing effective memories, they often disrupt the inherent logical and temporal relationships within interaction sessions, resulting in fragmented memory units and degraded reasoning performance. In this paper, we propose MetaMem, a novel framework that augments memory systems with a self-evolving meta-memory, aiming to teach LLMs how to effectively utilize memorized knowledge. During meta-memory optimization, MetaMem iteratively distills transferable knowledge utilization experiences across different tasks by self-reflecting on reasoning processes and performing actions to update the current meta-memory state. The accumulated meta-memory units serve as explicit knowledge utilization experiences, guiding the LLM to systematically identify and integrate critical evidence from scattered memory fragments. Extensive experiments demonstrate the effectiveness of MetaMem, which significantly outperforms strong baselines by over 3.6%. All codes and datasets are available at https://github.com/OpenBMB/MetaMem.

Zhihao Xu, Yongqi Tong, Xin Zhang, Jun Zhou, Xiting Wang

Multi-objective preference alignment often faces alignment conflicts, where optimizing for one objective (e.g., helpfulness) degrades performance on others (e.g., harmlessness). While prior work focuses on algorithmic solutions, the intrinsic conflict within data and its theoretical impact on training remain underexplored. To bridge this gap, we introduce the principle of Reward Consistency (RC), a theory-grounded criterion that approximates the alignment conflicts via reward models. We prove that a sample mitigates conflicts if and only if it satisfies RC, thereby ensuring improvement across all objectives during optimization. Building on this, we propose Reward Consistency Sampling (RCS), an automated framework for constructing pairwise data that adheres to RC, supplemented by a relaxation strategy to enhance flexibility. Extensive experiments show that RCS brings significant and consistent performance gains, achieving an average improvement of 23.07% in both harmlessness and helpfulness during simultaneous optimization comparde to the vanilla dataset. Our data-centric approach is complementary to existing alignment algorithms and effective in both sequential and simultaneous optimization scenarios.

Yingjian Zhu, Xinming Wang, Kun Ding, Ying Wang, Bin Fan, Shiming Xiang

Multi-modal Retrieval-Augmented Generation (RAG) has emerged as a highly effective paradigm for Knowledge-Based Visual Question Answering (KB-VQA). Despite recent advancements, prevailing methods still primarily depend on images as the retrieval key, and often overlook or misplace the role of Vision-Language Models (VLMs), thereby failing to leverage their potential fully. In this paper, we introduce WikiSeeker, a novel multi-modal RAG framework that bridges these gaps by implementing a multi-modal retriever and redefining the role of VLMs. Rather than serving merely as answer generators, we assign VLMs two specialized agents: a Refiner and an Inspector. The Refiner utilizes the capability of VLMs to rewrite the textual query according to the input image, significantly improving the performance of the multimodal retriever. The Inspector facilitates a decoupled generation strategy by selectively routing reliable retrieved context to another LLM for answer generation, while relying on the VLM’s internal knowledge when retrieval is unreliable. Extensive experiments on EVQA, InfoSeek, and M2KR demonstrate that WikiSeeker achieves state-of-the-art performance, with substantial improvements in both retrieval accuracy and answer quality.

Pengxiang Lan, Enneng Yang, Yuliang Liang, Jianzhe Zhao, Linying Jiang, Guibing Guo

Next location prediction aims to infer the next location users are likely to visit based on their historical check-in data. However, existing methods assume that check-in data is complete, overlooking the subjective nature of users’ check-in behavior, leading to inaccurate capture of user preferences. Recently, Large Language Models (LLMs) have offered a promising approach to location completion due to their extensive world knowledge. Nevertheless, our experiments reveal that LLMs struggle to interpret raw geographic coordinate information. To address these challenges, we propose LaMDA, an LLM-driven Multi-perspective Data Augmentation framework that employs dual completion agents to complement user mobility behaviors. Driven by our empirical findings that natural language descriptions align more closely with real-world geographic logic, LaMDA translates geographic coordinates into text to enhance spatial reasoning. Leveraging these semantic descriptions, LaMDA constructs dual agents to complement user mobility: "Micro-Level Completion" fills short-term omissions, while "Macro-Level Completion" infers unrecorded locations based on periodic preferences. Reliability is ensured through tailored real-world point-of-interest (POI) pools and a self-verification mechanism. Finally, a collaborative dual-graph module leverages this augmented data for fine-grained preference modeling. Extensive experiments on three real-world datasets demonstrate that LaMDA significantly outperforms state-of-the-art methods.

Guo Tang, Ke Cheng, Huiming Fan, Heng Chang, Wenxiang Zheng, Xianhao Ou, Junjia Xiang, Ming Liu, Yujun Zhou, Li Lanyu 等

Temporal Knowledge Graph (TKG) forecasting faces significant challenges due to distribution shifts and poor inductive generalization in parametric models. While Large Language Models (LLMs) offer potent semantic reasoning, existing LLM-based approaches struggle with implicit modality alignment and suboptimal graph linearization, failing to capture complex topologies without retraining. To bridge this gap, we propose ExE-LLM, a training-free, test-time adaptive framework that reframes TKG prediction as explicit evidence-driven reasoning. Our core philosophy is to decouple topological calculation from semantic reasoning: a heuristic module translates latent graph signals into natural language evidence, enabling the LLM to perform multi-source judgment. ExE-LLM incorporates a task-aware scheduler for test-time adaptation, a heuristic synthesizer for explicit modality alignment, and a self-diagnosis module for iterative optimization. Extensive experiments on four benchmarks demonstrate that ExE-LLM achieves SOTA performance in inductive settings, significantly outperforming fully trained graph neural networks without updating LLM parameters. The source code is available at https://github.com/JENLISA4EVER/ExE-LLM.

Huopu Zhang, Yanguang Liu, Miao Zhang, Zirui He, Mengnan Du

Predicting earnings surprises through the analysis of earnings conference call transcripts has attracted increasing attention from the financial research community. Conference calls serve as critical communication channels between company executives, analysts, and shareholders, offering valuable forward-looking information. However, these transcripts present significant analytical challenges, typically containing over 5,000 words with substantial redundancy and industry-specific terminology that creates obstacles for language models. In this work, we propose the Sparse Autoencoder for Financial Representation Enhancement (SAE-FiRE) framework to address these limitations by extracting key information while eliminating redundancy. SAE-FiRE employs Sparse Autoencoders (SAEs) to efficiently identify patterns and filter out noises, and focusing specifically on capturing nuanced financial signals that have predictive power for earnings surprises. Experimental results indicate that the proposed method can significantly outperform comparing baselines.

Yiheng Zhao, Jun Yan

Although large language models (LLMs) excel in complex reasoning tasks, they suffer from severe causal hallucination in event causality identification (ECI), particularly in smaller models (\leq1.5B parameters). A promising approach to address this issue is to fine-tune them with Chain-of-Thought (CoT) traces. However, there is currently a lack of CoT trace dataset available for ECI. In this paper, we first investigate the essential criteria that effective CoT traces should possess to mitigate causal hallucination in smaller models. We then design a pipeline to generate CoT traces that meet these criteria. Moreover, since there is currently no metric for quantifying causal hallucination, we also introduce a new metric, the Causal Hallucination Rate (CHR), to quantify causal hallucination, guide the formulation of effective CoT trace criteria, and validate the effectiveness of our pipeline. Our experiments show that fine-tuning with the CoT traces generated by our pipeline not only substantially reduces causal hallucination in smaller LLMs but also improves mean accuracy. Moreover, the fine-tuned models exhibit strong cross-dataset and cross-difficulty generalization, as well as robustness under intentionally misleading intervention prompts.