论文检索

Dinh-Truong Do, Nguyen-Khang Le, Le-Minh Nguyen

Speculative decoding accelerates large language model (LLM) inference by using a lightweight drafter to propose multiple tokens, which are then verified in parallel by the base model. While effective in English, existing methods often struggle in multilingual scenarios due to static vocabularies and the lack of language-specific instruction data. To address these limitations, we present AdaSpec, a multilingual speculative decoding framework that dynamically adapts both the drafter and vocabulary at decoding time. AdaSpec generates language-specific instruction data using the LLM itself, enabling training of drafters for low-resource languages. It also constructs adaptive vocabularies tailored to each language's characteristics. In addition, we introduce Multi-SpecBench, a comprehensive multilingual benchmark covering seven languages and seven generation tasks, to evaluate multilingual speculative decoding performance. Extensive experiments show that AdaSpec achieves up to 2.3× speedup over the state-of-the-art method of EAGLE-2, even in English, demonstrating its effectiveness across diverse languages and tasks.

Zhenyu Ding, Yuhao Wang, Tengyue Xiao, Haoying Wang, Guojun Ma, Mingyang Wan, Caigui Jiang, Ning Ding

Large Language Models (LLMs) demonstrate impressive capabilities, yet their outputs often suffer from misalignment with human preferences due to the inadequacy of weak supervision and a lack of fine-grained control. Training-time alignment methods like Reinforcement Learning from Human Feedback (RLHF) face prohibitive costs in expert supervision and inherent scalability limitations, offering limited dynamic control during inference. Consequently, there is an urgent need for scalable and adaptable alignment mechanisms. To address this, we propose W2S-AlignTree, a pioneering plug-and-play inference-time alignment framework that synergistically combines Monte Carlo Tree Search (MCTS) with the Weak-to-Strong Generalization paradigm for the first time. W2S-AlignTree formulates LLM alignment as an optimal heuristic search problem within a generative search tree. By leveraging weak model's real-time, step-level signals as alignment proxies and introducing an Entropy-Aware exploration mechanism, W2S-AlignTree enables fine-grained guidance during strong model's generation without modifying its parameters. The approach dynamically balances exploration and exploitation in high-dimensional generation search trees. Experiments across controlled sentiment generation, summarization, and instruction-following show that W2S-AlignTree consistently outperforms strong baselines. Notably, W2S-AlignTree raises the performance of Llama3-8B from 1.89 to 2.19, a relative improvement of 15.9% on the summarization task.

Ruomeng Ding, Wei Cheng, Minglai Shao, Chen Zhao

Large language models (LLMs) are increasingly applied to sequential decision-making through in-context learning (ICL), yet their effectiveness is highly sensitive to prompt quality. Effective prompts should meet three principles: focus on decision-critical information, provide step-level granularity, and minimize reliance on expert annotations through label efficiency. However, existing ICL methods often fail to satisfy all three criteria simultaneously. Motivated by these challenges, we introduce SkillGen, a skill-based ICL framework for structured sequential reasoning. It constructs an action-centric, domain-level graph from sampled trajectories, identifies high-utility actions via temporal-difference credit assignment, and retrieves step-wise skills to generate fine-grained, context-aware prompts. We further present a theoretical analysis showing that focusing on high-utility segments supports task identifiability and informs more effective ICL prompt design. Experiments on ALFWorld, BabyAI, and ScienceWorld, using both open-source and proprietary LLMs, show that SkillGen achieves consistent gains, improving progress rate by 5.9%–16.5% on average across models.

Jérémie Dentan, Davide Buscaldi, Sonia Vanier

Verbatim memorization in Large Language Models (LLMs) is a multifaceted phenomenon involving distinct underlying mechanisms. We introduce a novel method to analyze the different forms of memorization described by the existing taxonomy. Specifically, we train Convolutional Neural Networks (CNNs) on the attention weights of the LLM and evaluate the alignment between this taxonomy and the attention weights involved in decoding. We find that the existing taxonomy performs poorly and fails to reflect distinct mechanisms within the attention blocks. We propose a new taxonomy that maximizes alignment with the attention weights, consisting of three categories: memorized samples that are guessed using language modeling abilities, memorized samples that are recalled due to high duplication in the training set, and non-memorized samples. Our results reveal that few-shot verbatim memorization does not correspond to a distinct attention mechanism. We also show that a significant proportion of extractable samples are in fact guessed by the model and should therefore be studied separately. Finally, we develop a custom visual interpretability technique to localize the regions of the attention weights involved in each form of memorization.

Shuangyan Deng, Zhongsheng Wang, Rui Mao, Ciprian Doru Giurcăneanu, Jiamou Liu

Recent advances in Multimodal Large Language Models (MLLMs) have enabled joint reasoning over financial textual and visual inputs. However, they still struggle with financial terminology, logical consistency, and numerical computations. Moreover, while commercial large models perform well on reasoning tasks, their high inference costs limit their scalable usage in real world financial applications. We thus propose a cost-effective framework, CLER, that combines contrastive retrieval with step-wise reflection to improve reasoning performance. Also, the reasoning cost is only generated in the test stage when using commercial large models. CLER leverages FinErrorSet, a dataset of 8,000+ mistake correction pairs from diverse open-source MLLMs. A fine grained retriever is trained to identify structurally relevant errors for self-correction through individual reflection. Experiments on three benchmarks show that CLER consistently outperforms other baselines. To our knowledge, CLER is the first framework to use cross-model errors for financial reasoning.

Cui Danxin, Sihang Jiang, Keyi Wang, Zhiyi Duan, Yanghua Xiao, Bi Yude, Jiaqing Liang, Minggui He, Shimin Tao, Yilun Liu

As large language models (LLMs) are increasingly deployed in high-stakes domains such as education, healthcare, and law, accurately evaluating their nuanced reasoning process becomes essential to ensure their safety, reliability, and trustworthiness. However, most existing benchmarks evaluate LLMs at a coarse granularity. Current benchmarks lack a unified framework and rely on single‐task datasets, overlooking the intermediate steps of complex reasoning. This results in redundant overlap across benchmarks, poor generalization to multifaceted real-world tasks, and underutilizes the rich reasoning traces generated by advanced LLMs.

Chenhao Dang, Jing Ma

A persistent challenge in text classification (TC) is that enhancing model robustness against adversarial attacks typically degrades performance on clean data. We argue that this challenge can be resolved by modeling the distribution of clean samples in the encoder’s embedding manifold. To this end, we propose the Manifold-Correcting Causal Flow (MC²F), a two-module system that operates directly on sentence embeddings. A Stratified Riemannian Continuous Normalizing Flow (SR-CNF) learns the density of the clean data manifold. It identifies out-of-distribution embeddings, which are then corrected by a Geodesic Purification Solver. This solver projects adversarial points back onto the learned manifold via the shortest path, restoring a clean, semantically coherent representation. We conducted extensive evaluations on text classification (TC) across three datasets and multiple adversarial attacks. The results demonstrate that our method, MC²F, not only establishes a new state-of-the-art in adversarial robustness but also fully preserves performance on clean data, even yielding modest gains in Accuracy.

Zhenlong Dai, Zhuoluo Zhao, Hengning Wang, Xiu Tang, Sai Wu, Chang Yao, Zhipeng Gao, Jingyuan Chen

With the development of large language models (LLMs) in the field of programming, intelligent programming coaching systems have gained widespread attention. However, most research focuses on repairing the buggy code of programming learners without providing the underlying causes of the bugs. To address this gap, we introduce a novel task, namely LPR (Learner-Tailored Program Repair). We then propose a novel and effective framework, LSGen (Learner-Tailored Solution Generator), to enhance program repair while offering the bug descriptions for the buggy code. In the first stage, we utilize a repair solution retrieval framework to construct a solution retrieval database and then employ an edit-driven code retrieval approach to retrieve valuable solutions, guiding LLMs in identifying and fixing the bugs in buggy code. In the second stage, we propose a solution-guided program repair method, which fixes the code and provides explanations under the guidance of retrieval solutions. Moreover, we propose an Iterative Retrieval Enhancement method that utilizes evaluation results of the generated code to iteratively optimize the retrieval direction and explore more suitable repair strategies, improving performance in practical programming coaching scenarios. The experimental results show that our approach outperforms a set of baselines by a large margin, validating the effectiveness of our framework for the newly proposed LPR task.

Yuqin Dai, Shuo Yang, Guoqing Wang, Yong Deng, Zhanwei Zhang, Jun Yin, Pengyu Zeng, Zhenzhe Ying, Changhua Meng, Can Yi 等

Retrieval-Augmented Generation (RAG) enhances large language models (LLMs) by integrating up-to-date external knowledge, yet real-world web environments present unique challenges. These limitations manifest as two key challenges: pervasive misinformation in the web environment, which introduces unreliable or misleading content that can degrade retrieval accuracy, and the underutilization of web tools, which, if effectively employed, could enhance query precision and help mitigate this noise, ultimately improving retrieval results in RAG systems. To address these issues, we propose WebFilter, a novel RAG framework that generates source-restricted queries and filters out unreliable content. This approach combines a retrieval filtering mechanism with a behavior- and outcome-driven reward strategy, optimizing both query formulation and retrieval outcomes. Extensive experiments demonstrate that WebFilter improves answer quality and retrieval precision, outperforming existing RAG methods on both in-domain and out-of-domain benchmarks.

Lingling Dai, Andong Li, Cheng Chi, Yifan Liang, Xiaodong Li, Chengshi Zheng

In the field of audio generation, signal-to-noise ratio (SNR) has long served as an objective metric for evaluating audio quality. Nevertheless, recent studies have shown that SNR and its variants are not always highly correlated with human perception, prompting us to raise the questions: Why does SNR fail in measuring audio quality? And how to improve its reliability as an objective metric? In this paper, we identify the inadequate measurement of phase distance as a pivotal factor and propose to reformulate SNR with specially designed phase-distance terms, yielding an improved metric named GOMPSNR. We further extend the newly proposed formulation to derive two novel categories of loss function, corresponding to magnitude-guided phase refinement and joint magnitude-phase optimization, respectively. Besides, extensive experiments are conducted for an optimal combination of different loss functions. Experimental results on advanced neural vocoders demonstrate that our proposed GOMPSNR exhibits more reliable error measurement than SNR. Meanwhile, our proposed loss functions yield substantial improvements in model performance, and our well-chosen combination of different loss functions further optimizes the overall model capability.

Fangqi Dai, Xingjian Jiang, Zizhuang Deng

To prevent misinformation and social issues arising from trustworthy-looking content generated by LLMs, it is crucial to develop efficient and reliable methods for identifying the source of texts. Previous approaches have demonstrated exceptional performance in detecting texts fully generated by LLMs. However, these methods struggle when confronting more advanced LLM output or text with adversarial multi task machine-revision, especially in the black-box setting, where the generating model is unknown. To address this challenge, grounded in the hypothesis that human writing possesses consistent, distinctive stylistic patterns, we propose Human Language Preference Detection (HLPD). HLPD employs a reward‐based alignment process, Human Language Preference Optimization (HLPO), to shift the scoring model’s token distribution toward human‐like writing, making the model more sensitive to human writing, therefore enhancing the identification of machine-revised text. We test HLPD in an adversarial multi‑task evaluation framework that leverages a five‑dimensional prompt generator and multiple advanced LLMs to create diverse revision scenarios. When detecting texts revised by GPT-series models, HLPD achieves a 15.11% relative improvement in AUROC over ImBD, surpassing Fast-DetectGPT by 45.56%. When evaluated in texts generated by advanced LLMs, HLPD achieves the highest average AUROC, exceeding ImBD by 5.53% and Fast-DetectGPT by 34.14%.

Shiyao Cui, Xijia Feng, Yingkang Wang, Junxiao Yang, Zhexin Zhang, Biplab Sikdar, Hongning Wang, Han Qiu, Minlie Huang

Emojis are globally used non-verbal cues in digital communication, and extensive research has examined how large language models (LLMs) understand and utilize emojis across contexts. While usually associated with friendliness or playfulness, it is observed that emojis may trigger toxic content generation in LLMs. Motivated by such a observation, we aim to investigate: (1) whether emojis can clearly enhance the toxicity generation in LLMs and (2) how to interpret this phenomenon.* We begin with a comprehensive exploration of emoji-triggered LLM toxicity generation by automating the construction of prompts with emojis to subtly express toxic intent. Experiments across 5 mainstream languages on 7 famous LLMs along with jailbreak tasks demonstrate that prompts with emojis could easily induce toxicity generation. To understand this phenomenon, we conduct model-level interpretations spanning semantic cognition, sequence generation and tokenization, suggesting that emojis can act as a heterogeneous semantic channel to bypass the safety mechanisms. To pursue deeper insights, we further probe the pre-training corpus and uncover potential correlation between the emoji-related data polution with the toxicity generation behaviors.

Jing Cui, Yufei Han, Jianbin Jiao, Junge Zhang

Backdoor attacks embed malicious behaviors into Large Language Models (LLMs), enabling adversaries to trigger harmful outputs or bypass safety controls. However, the persistence of the implanted backdoors under user-driven post-deployment continual fine-tuning has been rarely examined. Most prior works evaluate the effectiveness and generalization of implanted backdoors only at releasing and empirical evidence shows that naively injected backdoor persistence degrades after updates. In this work, we study whether and how implanted backdoors persist through a multi‑stage post-deployment fine‑tuning. We propose P‑Trojan, a trigger‑based attack algorithm that explicitly optimizes for backdoor persistence across repeated updates. By aligning poisoned gradients with those of clean tasks on token embeddings, the implanted backdoor mapping is less likely to be suppressed or forgotten during subsequent updates. Theoretical analysis shows the feasibility of such persistent backdoor attacks after continual fine-tuning. And experiments conducted on the Qwen2.5 and LLaMA3 families of LLMs, as well as diverse task sequences, demonstrate that P‑Trojan achieves over \textbf{99\%} persistence while preserving clean‑task accuracy. Our findings highlight the need for persistence-aware evaluation and stronger defenses in realistic model adaptation pipelines.

Seungho Choi, Sihyun Park, Minsang Kim, Chansol Park, Bongsu Kim

Large language models (LLMs) often show poor performance in low-resource languages like Korean, partly due to unique linguistic challenges such as homophonous Sino-Korean words that are indistinguishable in Hangul script. To address this semantic ambiguity, we propose HanjaBridge, a novel meaning-injection technique integrated into a continual pre-training (CPT) framework. Instead of deterministically mapping a word to a single Hanja (Chinese character), HanjaBridge presents the model with all possible Hanja candidates for a given homograph, encouraging the model to learn contextual disambiguation. This process is paired with token-level knowledge distillation to prevent catastrophic forgetting. Experimental results show that HanjaBridge significantly improves Korean language understanding, achieving a 21% relative improvement on the KoBALT benchmark. Notably, by reinforcing semantic alignment between Korean and Chinese through shared Hanja, we observe a strong positive cross-lingual transfer. Furthermore, these gains persist even when Hanja augmentation is omitted at inference time, ensuring practical efficiency with no additional run-time cost.

Jonghyeon Choi, Yeonjun Choi, Hyun-chul Kim, Beakcheol Jang

This paper provides an early effort to systematically examine nation-level biases exhibited by Large Language Models (LLMs) within the domain of International Relations (IR), a dimension that has remained largely unexplored in prior research. Leveraging historical records from the United Nations Security Council (UNSC), we developed a bias evaluation framework comprising three distinct tests to explore nation-level bias in various LLMs, with a particular focus on the five permanent members of the UNSC. Experimental results show that, even with the general bias patterns across models (e.g., favorable biases toward the western nations, and unfavorable biases toward Russia), these still vary based on the LLM. Notably, even within the same LLM, the direction and magnitude of bias for a nation change depending on the evaluation context. This observation suggests that LLM biases are fundamentally multidimensional, varying across models and tasks. We also observe that models with stronger reasoning abilities show reduced bias and better prediction performance. Building on this finding, we introduce a debiasing framework that improves LLMs’ factual reasoning combining Retrieval-Augmented Generation with Reflexion-based self-reflection techniques. Experiments show it effectively reduces nation-level bias, and improves performance, particularly in GPT-4o-mini and LLama-3.3-70B. Our findings emphasize the need to assess nation-level bias alongside prediction performance when applying LLMs in the IR domain.

Jinwei Chi, Ke Wang, Yu Chen, Xuanye Lin, Qiang Xu

Automated essay scoring (AES) is a challenging task in cross-prompt settings due to the diversity of scoring criteria. While previous studies have focused on the output of large language models (LLMs) to improve scoring accuracy, we believe activations from intermediate layers may also provide valuable information. To explore this possibility, we evaluated the discriminative power of LLMs’ activations in cross-prompt essay scoring task. Specifically, we used activation to fit probes and further analyzed the effects of different models and input content of LLMs on this discriminative power. By computing the directions of essays across various trait dimensions under different prompts, we analyzed the variation in evaluation perspectives of large language models concerning essay types and traits. Results show that the activations possess strong discriminative power in evaluating essay quality and that LLMs can adapt their evaluation perspectives to different traits and essay types, effectively handling the diversity of scoring criteria in cross-prompt settings.

Yao Cheng, Yibo Zhao, Jiapeng Zhu, Yao Liu, Xing Sun, Xiang Li

Large Language Models (LLMs) have demonstrated significant potential across various domains. However, they often struggle with integrating external knowledge and performing complex reasoning, leading to hallucinations and unreliable outputs. Retrieval Augmented Generation (RAG) has emerged as a promising paradigm to mitigate these issues by incorporating external knowledge. Yet, conventional RAG approaches, especially those based on vector similarity, fail to effectively capture relational dependencies and support multi-step reasoning. In this work, we propose CogGRAG, a human cognition-inspired, graph-based RAG framework designed for Knowledge Graph Question Answering (KGQA). CogGRAG models the reasoning process as a tree-structured mind map that decomposes the original problem into interrelated subproblems and explicitly encodes their semantic relationships. This structure not only provides a global view to guide subsequent retrieval and reasoning but also enables self-consistent verification across reasoning paths. The framework operates in three stages: (1) top-down problem decomposition via mind map construction, (2) structured retrieval of both local and global knowledge from external Knowledge Graphs (KGs), and (3) bottom-up reasoning with dual-process self-verification. Unlike previous tree-based decomposition methods such as MindMap or Graph-CoT, CogGRAG unifies problem decomposition, knowledge retrieval, and reasoning under a single graph-structured cognitive framework, allowing early integration of relational knowledge and adaptive verification. Extensive experiments demonstrate that CogGRAG achieves superior accuracy and reliability compared to existing methods.

Daixuan Cheng, Shaohan Huang, Xuekai Zhu, Bo Dai, Xin Zhao, Zhenliang Zhang, Furu Wei

Balancing exploration and exploitation is a central goal in reinforcement learning (RL). Despite recent advances in enhancing language model (LM) reasoning, most methods lean toward exploitation, and increasingly encounter performance plateaus. In this work, we revisit entropy -- a signal of exploration in RL -- and examine its relationship to exploratory reasoning in LMs. Through empirical analysis, we uncover positive correlations between high-entropy regions and three types of exploratory reasoning actions: (1) pivotal tokens that determine or connect logical steps, (2) reflective actions such as self-verification and correction, and (3) rare behaviors under-explored by the base LMs. Motivated by this, we introduce a minimal modification to standard RL with only one line of code: augmenting the advantage function with an entropy-based term. Unlike traditional maximum-entropy methods which encourage exploration by promoting uncertainty, we encourage exploration by promoting deeper and longer reasoning chains. Notably, our method achieves significant gains on the Pass@K metric -- an upper-bound estimator of LM reasoning capabilities -- even when evaluated with extremely large K values, pushing the boundaries of LM reasoning.

Zhengyu Chen, Yudong Wang, Teng Xiao, Ruochen Zhou, Xusheng Yang, Wei Wang, Zhifang Sui, Jingang Wang

Recent advancements in improving the reasoning capabilities of Large Language Models have underscored the efficacy of Process Reward Models (PRMs) in addressing intermediate errors through structured feedback mechanisms. This study analyzes PRMs from multiple perspectives, including training methodologies, scalability, and generalization capabilities. We investigate the interplay between pre-training and reward model training FLOPs to assess their influence on PRM efficiency and accuracy in complex reasoning tasks. Our analysis reveals a pattern of diminishing returns in performance with increasing PRM scale, highlighting the importance of balancing model size and computational cost. Furthermore, the diversity of training datasets significantly impacts PRM performance, emphasizing the importance of diverse data to enhance both accuracy and efficiency. We further examine test-time scaling strategies, identifying Monte Carlo Tree Search as the most effective method when computational resources are abundant, while Best-of-N Sampling serves as a practical alternative under resource-limited conditions. Notably, our findings indicate that PRMs trained on mathematical datasets exhibit performance comparable to those tailored for code generation, suggesting robust cross-domain generalization. Employing a gradient-based metric, we observe that PRMs exhibit a preference for selecting responses with similar underlying patterns, further informing their optimization.

Zehao Chen, Rong Pan, Haoran Li

Human writers often begin their stories with an overarching mental scene, where they envision the interactions between characters and their environment. Inspired by this creative process, we propose a novel approach to long-form story generation, termed hybrid bottom-up long-form story generation, using multi-agent simulations. In our method, agents interact within a dynamic sandbox environment, where their behaviors and interactions with one another and the environment generate emergent events. These events form the foundation for the story, enabling organic character development and plot progression. Unlike traditional top-down approaches that impose rigid structures, our hybrid bottom-up approach allows for the natural unfolding of events, fostering more spontaneous and engaging storytelling. The system is capable of generating stories exceeding 10,000 words while maintaining coherence and consistency, addressing some of the key challenges faced by current story generation models. We achieve state-of-the-art performance across several metrics. This approach offers a scalable and innovative solution for creating dynamic, immersive long-form stories that evolve organically from agent-driven interactions.