论文检索

Yibo Fan, Jingru Li, Huan Li

Large language models (LLMs) have been increasingly applied across a wide range of domains. However, recent studies have identified the presence of certain glitch tokens in their vocabularies, which can trigger hallucinations and lead to unpredictable or even harmful outputs. While various methods have been proposed to detect such tokens, effectively repairing them remains a key challenge for ensuring the reliability of LLMs. In this work, we propose GlitchCleaner, a lightweight yet effective approach to mitigate the adverse effects caused by glitch tokens. GlitchCleaner introduces auxiliary branches into specific components within selected layers of the model, enabling efficient and targeted token repair. These branches are implemented using the low-rank adaptation (LoRA) technique, adding less than 0.1% additional parameters to the original model. Furthermore, a gating mechanism dynamically controls the activation of these branches based on the model’s input, ensuring precise intervention without disrupting normal inference behavior. Experimental results across multiple mainstream models demonstrate that our method achieves an average repair rate of 86.88%, representing an improvement of over 30% compared to existing approaches, while ensuring lossless preservation of the model’s baseline capabilities and causing negligible impact on inference speed.

Xuanbo Fan, Tianqi Zhao, Yi Cheng, Chi Xiu, Jiaxin Guo, Boci Peng, Bingjing Xu, Jessica Zhang, Feng Sun, Yan Zhang

Retrieval-Augmented Generation (RAG) improves the factual accuracy of large language models by grounding responses in external content. However, most RAG systems assume access to static and well-organized corpora with fixed retrieval logic. In practice, real-world sources are heterogeneous and unlabeled, including user-uploaded documents, manuals, and datasets. Effective access in such settings requires adaptive and self-directed retrieval behavior. We present SegMem‑RAG, a memory-augmented RAG framework that learns to route queries across multiple unlabeled corpora based on experience. It incrementally updates a structured memory and uses self-reflection to guide retrieval over time without supervision. Experimental results demonstrate that SegMem‑RAG significantly outperforms recent baselines in generation quality on multi-corpus QA tasks.

Xiaoran Fan, Zhichao Sun, Tao Ji, Lixing Shen, Tao Gui

As vision-language models (VLMs) tackle increasingly complex and multimodal tasks, the rapid growth of Key-Value (KV) cache imposes significant memory and computational bottlenecks during inference. While Multi-Head Latent Attention (MLA) offers an effective means to compress the KV cache and accelerate inference, adapting existing VLMs to the MLA architecture without costly pretraining remains largely unexplored. In this work, we present \textbf{MHA2MLA-VLM}, a parameter-efficient and multimodal-aware framework for converting off-the-shelf VLMs to MLA. Our approach features two core techniques: (1) a modality-adaptive partial-RoPE strategy that supports both traditional and multimodal settings by selectively masking nonessential dimensions, and (2) a modality-decoupled low-rank approximation method that independently compresses the visual and textual KV spaces. Furthermore, we introduce parameter-efficient fine-tuning to minimize adaptation cost and demonstrate that minimizing output activation error, rather than parameter distance, substantially reduces performance loss. Extensive experiments on three representative VLMs show that MHA2MLA-VLM restores original model performance with minimal supervised data, significantly reduces KV cache footprint, and integrates seamlessly with KV quantization.

Xiaoran Fan, Zhichao Sun, Yangfan Gao, Jingfei Xiong, Hang Yan, Yifei Cao, Jiajun Sun, Shuo Li, Zhihao Zhang, Zhiheng Xi 等

Speech-language models (SLMs) offer a promising path toward unifying speech and text understanding and generation. However, challenges remain in achieving effective cross-modal alignment and high-quality speech generation. In this work, we systematically investigate the role of speech tokenizer designs in LLM-centric SLMs, augmented by speech heads and speaker modeling. We compare coupled, semi-decoupled, and fully decoupled speech tokenizers under a fair SLM framework and find that decoupled tokenization significantly improves alignment and synthesis quality. To address the information density mismatch between speech and text, we introduce multi-token prediction (MTP) into SLMs, enabling each hidden state to decode multiple speech tokens. This leads to up to 12× faster decoding and a substantial drop in word error rate (from 6.07 to 3.01). Furthermore, we propose a speaker-aware generation paradigm and introduce RoleTriviaQA, a large-scale role-playing knowledge QA benchmark with diverse speaker identities. Experiments demonstrate that our methods enhance both knowledge understanding and speaker consistency.

Qi Fan, An Zou, Yehan Ma

Large Language Models (LLMs) are increasingly deployed in time-critical systems, such as robotics, autonomous driving, embodied intelligence, and industrial automation, where generating accurate responses within a given time budget is crucial for decision-making, control, or safety-critical tasks. However, the auto-regressive generation process of LLMs makes it challenging to model and estimate the end-to-end execution time. Furthermore, existing efficient inference methods based on a fixed key-value (KV) cache eviction ratio struggle to adapt to varying tasks with diverse time budgets, where an improper eviction ratio may lead to incomplete inference or a drop in response performance. In this paper, we propose TimeBill, a novel time-budgeted inference framework for LLMs that balances the inference efficiency and response performance. To be more specific, we propose a fine-grained response length predictor (RLP) and an execution time estimator (ETE) to accurately predict the end-to-end execution time of LLMs. Following this, we develop a time-budgeted efficient inference approach that adaptively adjusts the KV cache eviction ratio based on execution time prediction and the given time budget. Finally, through extensive experiments, we demonstrate the advantages of TimeBill in improving task completion rate and maintaining response performance under various overrun strategies.

Jiejie Fan, Xiaojuan Ban, Zhiyan Zhang, Xi Sun

Graph Neural Networks (GNNs) offer superior modeling capabilities for text classification by capturing complex spatial features within semantic representations. However, existing graph-based approaches often suffer from computational inefficiency and limited ability to model both fine-grained local structures and the sequential nature of text. To address these challenges, we propose HC2-GNN, a Hierarchical Clustering and Coarsening Graph Neural Network, which introduces a novel lightweight graph clustering algorithm called Compromise Conductance Graph Clustering (C2GC). C2GC enables efficient graph clustering while simultaneously preserving both the textual order and the topological coherence of subgraphs. Furthermore, it incorporates a virtue cluster mechanism that expands each subgraph with semantically relevant neighbors, explicitly enabling cross-cluster information propagation without compromising local structural integrity. HC2-GNN aggregates local and global features by combining subgraph-level and full-graph representations, enhancing semantic discriminability for classification. Extensive experiments on benchmark datasets demonstrate that HC2-GNN consistently outperforms existing state-of-the-art text classification methods.

Zeying Duan, Youtian Du, Yuanlin Chang, Bowen Lin, Weijia Wu

Neuro-symbolic learning has emerged as a promising paradigm for interpretable visual reasoning, where mapping natural language questions to executable programs plays a central role. However, most existing methods focus exclusively on the forward program generation from questions while overlooking the reverse process of reconstructing questions from programs. In this paper, we propose BiPaR (Bidirectional Parsing and Reconstruction), a Transformer-based framework that jointly models both program parsing and question reconstruction within a unified architecture. Unlike previous approaches that only perform forward parsing, BiPaR introduces reverse program-to-question reconstruction as a powerful auxiliary signal, which improves program generation quality and accelerates convergence, particularly under limited supervision. We further provide a theoretical analysis showing how reverse reconstruction facilitates faster optimization during training. The bidirectional modeling makes BiPaR well-suited for both supervised and semi-supervised learning scenarios. We present two architectural variants: BiPaR-Full, which employs encoder-decoder Transformers for both modules; and BiPaR-DOnly, a lightweight variant that employs a decoder-only structure for question reconstruction, reducing model complexity. Experiments on CLEVR and a GQA subset demonstrate that BiPaR significantly outperforms standard Transformer baselines. Furthermore, in the semi-supervised learning setting, BiPaR achieves notable improvements by leveraging additional questions without program annotations.

Yong Du, Yuchen Yan, Fei Tang, Zhengxi Lu, Chang Zong, Weiming Lu, Shengpei Jiang, Yongliang Shen

Graphical User Interface (GUI) grounding, the task of mapping natural language instructions to precise screen coordinates, is fundamental to autonomous GUI agents. While existing methods achieve strong performance through extensive supervised training or reinforcement learning with labeled rewards, they remain constrained by the cost and availability of pixel-level annotations. We observe that when models generate multiple predictions for the same GUI element, the spatial overlap patterns reveal implicit confidence signals that can guide more accurate localization. Leveraging this insight, GUI-RC (Region Consistency), a test-time scaling method that constructs spatial voting grids from multiple sampled predictions to identify consensus regions where models show highest agreement. Without any training, GUI-RC improves accuracy by 2-3% across various architectures on ScreenSpot benchmarks. We further introduce GUI-RCPO (Region Consistency Policy Optimization), transforming these consistency patterns into rewards for test-time reinforcement learning. By computing how well each prediction aligns with the collective consensus, GUI-RCPO enables models to iteratively refine their outputs on unlabeled data during inference. Extensive experiments demonstrate the generality of our approach: using only 1,272 unlabeled data, GUI-RCPO achieves 3-6% accuracy improvements across various architectures on ScreenSpot benchmarks. Our approach reveals the untapped potential of test-time scaling and test-time reinforcement learning for GUI grounding, offering a promising path toward more data-efficient GUI agents.

Yiming Du, Bingbing Wang, Yang He, Bin Liang, Baojun Wang, Zhongyang Li, Lin Gui, Jeff Z. Pan, Ruifeng Xu, Kam-Fai Wong

Modern task-oriented dialogue (TOD) systems increasingly rely on large language model (LLM) agents, leveraging Retrieval-Augmented Generation (RAG) and long-context capabilities for long-term memory utilization. However, these methods prioritise semantic similarity over task intent, degrading multi-session coherence. We propose MemGuide, a two-stage intent-driven memory selection framework: (1) Intent‑Aligned Retrieval retrieves goal-consistent QA‑formatted memory units; (2) Missing‑Slot Guided Filtering reranks units by slot-completion gain via a chain‑of‑thought reasoner and fine‑tuned LLaMA‑8B filter. We also introduce the MS-TOD, the first multi-session TOD benchmark with 132 diverse personas, 956 task goals, and annotated intent-aligned memory targets. Evaluations on MS-TOD show that MemGuide boosts task success rate by 11% (88%→99%) and reduces dialogue length by 2.84 turns, and matches single‑session performance.

Yexing Du, Kaiyuan Liu, Youcheng Pan, Zheng Chu, Bo Yang, Xiaocheng Feng, Ming Liu, Yang Xiang

As Large Language Models (LLMs) are increasingly popularized in the multilingual world, ensuring hallucination-free factuality becomes markedly crucial. However, existing benchmarks for evaluating the reliability of Multimodal Large Language Models (MLLMs) predominantly focus on textual or visual modalities with a primary emphasis on English, which creates a gap in evaluation when processing multilingual input, especially in speech. To bridge this gap, we propose a novel Cross-lingual and Cross-modal Factuality benchmark (CCFQA). Specifically, the CCFQA benchmark contains parallel speech-text factual questions across 8 languages, designed to systematically evaluate MLLMs' cross-lingual and cross-modal factuality capabilities. Our experimental results demonstrate that current MLLMs still face substantial challenges on the CCFQA benchmark. Furthermore, we propose a few-shot transfer learning strategy that effectively transfers the Question Answering (QA) capabilities of LLMs in English to multilingual Spoken Question Answering (SQA) tasks, achieving competitive performance with GPT-4o-mini-Audio using just 5-shot training. We release CCFQA as a foundational research resource to promote the development of MLLMs with more robust and reliable speech understanding capabilities.

Li Du, Hanyu Zhao, Yiming Ju, Tengfei Pan

High-quality instructions are crucial for aligning pretrained models to improve their performance on downstream tasks. Although current instruction datasets have reached tens of millions of samples, models finetuned on them may still struggle with complex instruction following and tasks in rare domains. This is primarily due to limited expansion in both "coverage" (coverage of task types and knowledge areas) and "depth" (instruction complexity) of the instruction set. To address this issue, we propose a systematic instruction data construction framework, which integrates a hierarchical labeling system, an informative seed selection algorithm, an evolutionary data synthesis process, and a model deficiency diagnosis with targeted data generation. These components form an iterative closed-loop to continuously enhance the coverage and depth of instruction data. Based on this framework, we construct Infinity Instruct Subject, a high-quality dataset containing approximately 1.5 million instructions. Experiments on multiple foundation models and benchmark tasks demonstrate its effectiveness in improving instruction-following capabilities. Further analyses suggest that InfinityInstruct-Subject shows enlarged coverage and depth compared to comparable synthesized instruction datasets.

Haizhou Du, Jiujiu Li, Dongyang Li, Luobin Huang, Lisheng Wang

Multi-Hop Question Answering (MHQA) requires step-by-step reasoning across multiple pieces of information to answer complex questions. The cache-aided Retrieval-Augmented Generation (RAG) can accelerate the process of external knowledge retrieval at each reasoning step for MHQA. However, existing methods focus on the internal structure and ignore the misalignment between the queries’ arrival order and cache hit order. To tackle this, we propose Mnemosyne, a cache hit order fitting method designed to accelerate the RAG progress for MHQA. Specifically, our cache-aware order fitting strategy adjusts the order of queries arrival via graph reordering to better align with the cache hit order, thereby reducing the likelihood of failed or unproductive retrieval attempts. The multi-granularity caching storage mechanism is designed to loosen the strict hit condition to multiple similar semantic matching modes, facilitating that relevant documents can still be retrieved. Experiments conducted on four multi-hop QA datasets demonstrate that Mnemosyne effectively reduces retrieval latency while enhancing task answer F1 score, achieving a superior trade-off between efficiency and effectiveness.

Yulin Dou, Jiangming Liu

Humans increasingly query Large Language Models (LLMs) to accomplish personal tasks according to their individual preferences. However, these preferences are often unconsciously veiled during conversation. To address this, LLMs have to elicit human preferences through multi-turn dialogue, where tasks are accomplished via iterative clarifying questions and final response generated by LLMs as effective questioners. Existing approaches based on self-taught reasoning have two limitations: 1) they struggle to avoid generating irrelevant questions and 2) the final responses to tasks are misled by the conversations. To overcome these limitations, we propose TO-GATE, a novel framework that enhances question generation through trajectory optimization. TO-GATE comprises two key components: a clarification resolver, which generates optimal questioning trajectories to produce effective elicitation questions, and a summarizer, which ensures task-aligned final responses. Experimental results show that TO-GATE significantly outperforms baseline methods, achieving a 9.32% improvement on standard preference elicitation benchmarks.

Zhongren Dong, Bin Wang, Jing Han, Haotian Guo, Xiaojun Mo, Yimin Cao, Zixing Zhang

Neural Speech Codecs face a fundamental trade-off at low bitrates: preserving acoustic fidelity often compromises semantic richness. To address this, we introduce SACodec, a novel codec built upon an asymmetric dual-quantizer that employs our proposed Semantic Anchoring mechanism. This design strategically decouples the quantization of Semantic and Acoustic details. The semantic anchoring is achieved via a lightweight projector that aligns acoustic features with a frozen, large-scale mHuBERT codebook, injecting linguistic priors while guaranteeing full codebook utilization. Sequentially, for acoustic details, a residual activation module with SimVQ enables a single-layer quantizer (acoustic path) to faithfully recover fine-grained information. At just 1.5 kbps, SACodec establishes a new state of the art by excelling in both fidelity and semantics: subjective listening tests confirm that its reconstruction quality is perceptually highly comparable to ground-truth audio, while its tokens demonstrate substantially improved semantic richness in downstream tasks. This work suggests that assigning specialized semantic quantizers to distinct information streams offers an effective path to reconcile the long-standing trade-off between fidelity, semantics, and modeling simplicity in low-bitrate speech tokenization.

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.