论文检索

Weijie Li, Jin Wang, Liang-Chih Yu, Xuejie Zhang

Large reasoning models (LRMs) improve performance at test time by thinking longer, but this often leads to overthinking and high computational cost. To address this, recent reinforcement learning (RL) methods adopt outcome-level rewards, such as rule- or prompt-based signals, that favor shorter correct reasoning paths but often overlook reasoning quality. While such rewards neglect intermediate reasoning, dense supervision from process reward models (PRMs) has proven more effective in promoting coherent and high-quality reasoning. However, static PRM supervision introduces two challenges: reward hacking, since fixed rewards poorly capture global reasoning objectives, and the high training cost of obtaining dense reward labels at scale. To overcome these issues, we propose Step Group Relative Policy Optimization (Step-GRPO), a GRPO-based method that integrates step-level PRM signals into sparse trajectory-level feedback, avoiding costly step-level supervision while improving reasoning quality beyond accuracy. In addition, Step-GRPO employs a step-attention mechanism that captures inter-step dependencies and emphasizes critical reasoning steps, effectively mitigating reward hacking. We apply Step-GRPO to train large language models and observe consistent gains in reasoning quality, accuracy, and shorter reasoning traces across multiple math benchmarks, outperforming reinforcement learning baselines at substantially lower cost. Notably, the proposed model achieves 36.7 percent accuracy on AIME 2024 with 11,000 training samples and a training cost of 38 US dollars, surpassing baselines that require over 1,000 US dollars and more than 40,000 samples, demonstrating strong cost-effectiveness and scalability.

Shuocheng Li, Yihao Liu, Silin Du, Wenxuan Zeng, Zhe Xu, Mengyu Zhou, Yeye He, Haoyu Dong, Shi Han, Dongmei Zhang

Large language models (LLMs) have shown great promise in automating data science workflows. However, existing models still struggle with multi-step reasoning and tool use, limiting their effectiveness on complex data analysis tasks. To address this limitation, we propose a scalable pipeline that extracts high-quality, tool-based data analysis tasks and their executable multi-step solutions from real-world Jupyter notebooks and associated data files. Using this pipeline, we introduce NbQA, a large-scale dataset of standardized task–solution pairs that reflect authentic tool-use patterns in practical data science scenarios. To further enhance the multi-step reasoning capabilities, we present Jupiter, a framework that formulates data analysis as a search problem and applies Monte Carlo Tree Search (MCTS) to generate diverse solution trajectories for value model learning. During inference, Jupiter combines the value model and node visit counts to efficiently collect executable multi-step plans with minimal search steps. Experimental results show that Qwen2.5-7B and 14B-Instruct models on NbQA solve 77.82% and 86.38% of tasks on InfiAgent-DABench, respectively—matching or surpassing GPT-4o and advanced agent frameworks. Further evaluations demonstrate improved generalization and stronger tool-use reasoning across diverse multi-step reasoning tasks.

Shuaimin Li, Liyang Fan, Yufang Lin, Zeyang Li, Xian Wei, Shiwen Ni, Hamid Alinejad-Rokny, Min Yang

Existing paper review methods often rely on superficial manuscript features or directly on large language models (LLMs), which are prone to hallucinations, biased scoring, and limited reasoning capabilities. Moreover, these methods often fail to capture the complex argumentative reasoning and negotiation dynamics inherent in reviewer-author interactions. To address these limitations, we propose ReViewGraph (Reviewer-Author Debates Graph Reasoner), a novel framework that performs heterogeneous graph reasoning over LLM-simulated multi-round reviewer-author debates. In our approach, reviewer-author exchanges are simulated through LLM-based multi-agent collaboration. Diverse opinion relations (e.g., acceptance, rejection, clarification, and compromise) are then explicitly extracted and encoded as typed edges within a heterogeneous interaction graph. By applying graph neural networks to reason over these structured debate graphs, ReViewGraph captures fine-grained argumentative dynamics and enables more informed review decisions. Extensive experiments on three datasets demonstrate that ReViewGraph outperforms strong baselines with an average relative improvement of 15.73%, underscoring the value of modeling detailed reviewer–author debate structures.

Shike Li, Xiaokai Wang, Xiaofeng Liu, Xin Tong, Hu Zhang

Large language models (LLMs) frequently generate fluent yet factually inaccurate content, a phenomenon known as hallucination. Recent inference-time approaches aim to improve truthfulness by steering model activations toward semantically meaningful directions. While effective to some extent, these methods typically process activations independently, neglecting the internal coordination structure of multi-head attention (MHA), where attention heads interact to form semantic representations. In this work, we propose CoFact, an adaptive inference-time mechanism that improves factual consistency by dynamically coordinating attention head behaviors. Inspired by cooperative game theory, CoFact conceptualizes attention heads as collaborative agents. It models the semantic utility and redundancy of each head and adaptively modulates their contributions to the final attention output. Notably, rather than directly altering intermediate representations, CoFact performs token-level coordination to encourage diverse and complementary attention patterns across heads. CoFact is plug-and-play compatible with mainstream LLM architectures and requires no additional supervision or model retraining. Experimental results across multiple standard factuality benchmarks demonstrate that CoFact consistently enhances factual accuracy while maintaining generation fluency.

Shangyu Li, Juyong Jiang, Tiancheng Zhao, Jiasi Shen

We introduce OSVBench, a new benchmark for evaluating Large Language Models (LLMs) on the task of generating complete formal specifications for verifying the functional correctness of operating system kernels. This benchmark is built upon a real-world operating system kernel, Hyperkernel, and consists of 245 complex specification generation tasks in total, each of which is a long-context task of about 20k-30k tokens. The benchmark formulates the specification generation task as a program synthesis problem confined to a domain for specifying states and transitions. This formulation is provided to LLMs through a programming model. The LLMs must be able to understand the programming model and verification assumptions before delineating the correct search space for syntax and semantics and generating formal specifications. Guided by the operating system's high-level functional description, the LLMs are asked to generate a specification that fully describes all correct states and transitions for a potentially buggy code implementation of the operating system. Experimental results with 12 state-of-the-art LLMs indicate limited performance of existing LLMs on the specification generation task for operating system verification. Significant disparities in their performance highlight differences in their ability to handle long-context code generation tasks.

Qizhi Li, Xuyang Wang, Yingke Chen, Ming Yan, Dezhong Peng, Xi Peng, Xu Wang

Recent advancements in Pre-trained Language Models (PLMs) have significantly enhanced performance across various Natural Language Processing (NLP) tasks. However, the variability in data distributions across different domains presents challenges in generalizing these models to unseen domains. Domain generalization offers a promising solution, but existing text domain generalization methods typically rely on adversarial training to learn domain-invariant features, which often leads to models with high computational and memory overhead. To address this issue, this paper proposes a novel solution named Generalization via Prompts and Contrastive Learning (GenPromptCL) to enhance the generalization capability in domain generalization. GenPromptCL consists of two key components: Domain-Misleading Prompt Learning (DMPL) and Pseudo Label-based Contrastive Learning (PCL). Specifically, DMPL disrupts domain labels randomly, misleading the model into producing incorrect domain labels. This forces the model to learn domain-invariant features. Meanwhile, PCL generates pseudo labels within a single mini-batch, enabling the model to learn both intra-class and inter-class discriminative representations with low time and space complexity. Extensive experimental results demonstrate that GenPromptCL achieves state-of-the-art performance on three distinct text classification tasks (sentiment analysis, rumor detection, and natural language inference) while significantly improving model operation efficiency.

Qiyang Li, Rui Kong, Yuchen Li, Hengyi Cai, Shuaiqiang Wang, Linghe Kong, Guihai Chen, Dawei Yin

The integration of dynamic, sparse structures like Mixture-of-Experts (MoE) with parameter-efficient adapters (e.g., LoRA) is a powerful technique for enhancing Large Language Models (LLMs). However, this architectural enhancement comes at a steep cost: despite minimal increases in computational load, the inference latency often skyrockets, leading to decoding speeds slowing by over 2.5 times. Through a fine-grained performance analysis, we pinpoint the primary bottleneck not in the computation itself, but in the severe overhead from fragmented, sequential CUDA kernel launches required for conventional dynamic routing. To address this challenge, we introduce AdaFuse, a framework built on a tight co-design between the algorithm and the underlying hardware system to enable efficient dynamic adapter execution. Departing from conventional layer-wise or block-wise routing, AdaFuse employs a token-level pre-gating strategy, which makes a single, global routing decision for all adapter layers before a token is processed. This ``decide-once, apply-everywhere'' approach effectively staticizes the execution path for each token, creating an opportunity for holistic optimization. We capitalize on this by developing a custom CUDA kernel that performs a fused switching operation, merging the parameters of all selected LoRA adapters into the backbone model in a single, efficient pass. Experimental results on popular open-source LLMs show that AdaFuse achieves accuracy on par with state-of-the-art dynamic adapters while drastically cutting decoding latency by a factor of over 2.4x, thereby bridging the gap between model capability and inference efficiency.

Qingchuan Li, Mingyue Cheng, Zirui Liu, Daoyu Wang, Yuting Zeng, Tongxuan Liu

Logical reasoning is a core challenge in natural language understanding and a fundamental capability of artificial intelligence, underpinning scientific discovery, mathematical theorem proving, and complex decision-making. Despite the remarkable progress of large language models (LLMs), most current approaches still rely on forward reasoning paradigms, generating step-by-step rationales from premises to conclusions. However, such methods often suffer from redundant inference paths, hallucinated steps, and semantic drift, resulting in inefficient and unreliable reasoning. In this paper, we propose a novel framework, Hypothesis-driven Backward Logical Reasoning (HBLR). The core idea is to integrate confidence-aware symbolic translation with hypothesis-driven backward reasoning. In the translation phase, only high-confidence spans are converted into logical form, such as first-order logic (FOL), while uncertain content remains in natural language. A translation reflection module further ensures semantic fidelity by evaluating symbolic outputs and reverting lossy ones back to text when necessary. In the reasoning phase, HBLR simulates human deductive thinking by assuming the conclusion is true and recursively verifying its premises. A reasoning reflection module further identifies and corrects flawed inference steps, enhancing logical coherence. Extensive experiments on five reasoning benchmarks demonstrate that HBLR consistently outperforms strong baselines in both accuracy and efficiency.

Qinfeng Li, Miao Pan, Jintao Chen, Fu Teng, Zhiqiang Shen, Ge Su, Hao Peng, Xuhong Zhang

Model merging has emerged as an efficient technique for expanding large language models (LLMs) by integrating specialized expert models. However, it also introduces a new threat: model merging stealing, where free-riders exploit models through unauthorized model merging. Unfortunately, existing defense mechanisms fail to provide effective protection. Specifically, we identify three critical protection properties that existing methods fail to simultaneously satisfy: (1) proactively preventing unauthorized merging; (2) ensuring compatibility with general open-source settings; (3) achieving high security with negligible performance loss. To address the above issues, we propose MergeBarrier, a plug-and-play defense that proactively prevents unauthorized merging. The core design of MergeBarrier is to disrupt the Linear Mode Connectivity (LMC) between the protected model and its homologous counterparts, thereby eliminating the low-loss path required for effective model merging. Extensive experiments show that MergeBarrier effectively prevents model merging stealing with negligible accuracy loss.

Qinfeng Li, Miao Pan, Ke Xiong, Ge Su, Zhiqiang Shen, Yan Liu, Sun Bing, Hao Peng, Xuhong Zhang

Retrieval-Augmented Generation (RAG) systems deployed over proprietary knowledge bases face growing threats from reconstruction attacks that aggregate model responses to replicate knowledge bases. Such attacks exploit both intra-class and inter-class paths—progressively extracting fine-grained knowledge within topics and diffusing it across semantically related ones, thereby enabling comprehensive extraction of the original knowledge base. However, existing defenses target only one path, leaving the other unprotected. We conduct a systematic exploration to assess the impact of protecting each path independently and find that joint protection is essential for effective defense. Based on this, we propose RAGFort, a structure-aware dual-module defense combining contrastive reindexing for inter-class isolation and constrained cascade generation for intra-class protection. Experiments across security, performance, and robustness confirm that RAGFort significantly reduces reconstruction success while preserving answer quality, offering the first comprehensive defense against knowledge base extraction attacks.

Qiming Li, Zekai Ye, Xiaocheng Feng, Weihong Zhong, Weitao Ma, Xiachong Feng

Despite the remarkable advancements of Large Vision-Language Models (LVLMs), the mechanistic interpretability remains underexplored. Existing analyses are insufficiently comprehensive and lack examination covering visual and textual tokens, model components, and the full range of layers. This limitation restricts actionable insights to improve the faithfulness of model output and the development of downstream tasks, such as hallucination mitigation. To address this limitation, we introduce Fine-grained Cross-modal Causal Tracing (FCCT) framework, which systematically quantifies the causal effects on visual object perception. FCCT conducts fine-grained analysis covering the full range of visual and textual tokens, three core model components including multi-head self-attention (MHSA), feed-forward networks (FFNs), and hidden states, across all decoder layers. Our analysis is the first to demonstrate that MHSAs of the last token in middle layers play a critical role in aggregating cross-modal information, while FFNs exhibit a three-stage hierarchical progression for the storage and transfer of visual object representations. Building on these insights, we propose Intermediate Representation Injection (IRI), a training-free inference-time technique that reinforces visual object information flow by precisely intervening on cross-modal representations at specific components and layers, thereby enhancing perception and mitigating hallucination. Consistent improvements across five widely used benchmarks and LVLMs demonstrate IRI achieves state-of-the-art performance, while preserving inference speed and other foundational performance.

Mengfan Li, Xuanhua Shi, Yang Deng

Large Language models (LLMs) are revolutionizing the conversational recommender systems (CRS) through their impressive capabilities in instruction comprehension, reasoning, and human interaction. A core factor underlying effective dialogue is the ability to infer and reason about others' mental states (such as desire, intention, and belief), a cognitive capacity commonly referred to as Theory of Mind (ToM). Despite growing interest in evaluating ToM in LLMs, current benchmarks predominantly rely on synthetic narratives inspired by Sally-Anne test, which emphasize physical perception and fail to capture the complexity of mental state inference in real-world conversational settings. Moreover,existing benchmarks often overlook a critical component of human ToM: behavioral prediction, the ability to use inferred mental states to guide strategic decision-making and select appropriate conversational actions for future interactions. To better align LLM-based ToM evaluation with human-like social reasoning, we propose RecToM, a novel benchmark for evaluating ToM abilities in recommendation dialogues. RecToM focuses on two complementary dimensions: Cognitive Inference and Behavioral Prediction. The former focus on understanding what has been communicated by inferring the underlying mental states. The latter emphasizes what should be done next, evaluating whether LLMs can leverage these inferred mental states to predict, select, and assess appropriate dialogue strategies. Together, these dimensions enable a comprehensive assessment of ToM reasoning in CRS. Extensive experiments on state-of-the-art LLMs demonstrate that RecToM poses a significant challenge. While the models exhibit partial competence in recognizing mental states, they struggle to maintain coherent, strategic ToM reasoning throughout dynamic recommendation dialogues, particularly in tracking evolving intentions and aligning conversational strategies with inferred mental states.

Longhao Li, Zhao Guo, Hongjie Chen, Yuhang Dai, Ziyu Zhang, Hongfei Xue, Tianlun Zuo, Chengyou Wang, Shuiyuan Wang, Xin Xu 等

The development of speech understanding and generation has been significantly accelerated by the availability of large-scale, high-quality speech datasets. Among these, ASR and TTS are regarded as the most established and fundamental tasks. However, for Cantonese (Yue Chinese), spoken by approximately 84.9 million native speakers worldwide, limited annotated resources have hindered progress and resulted in suboptimal ASR and TTS performance. To address this challenge, we propose WenetSpeech-Pipe, an integrated pipeline for building large-scale speech corpus with multi-dimensional annotation tailored for speech understanding and generation. Based on this pipeline, we release WenetSpeech-Yue, the first large-scale Cantonese speech corpus with multi-dimensional annotation for ASR and TTS, covering 21,800 hours across 10 domains with annotations including ASR transcription, text confidence, speaker identity, age, gender, speech quality scores, among other annotations. We also release WSYue-eval, a comprehensive Cantonese benchmark with two components: WSYue-ASR-eval, a manually annotated set for evaluating ASR on short and long utterances, code-switching, and diverse acoustic conditions, and WSYue-TTS-eval, with base and coverage subsets for standard and generalization testing. Experimental results show that models trained on WenetSpeech-Yue achieve competitive results against state-of-the-art (SOTA) Cantonese ASR and TTS systems, including commercial and LLM-based models, highlighting the value of our dataset and pipeline.

Junsong Li, Jie Zhou, Bihao Zhan, Yutao Yang, Qianjun Pan, Shilian Chen, Tianyu Huai, Xin Li, Qin Chen, Liang He

Alignment plays a crucial role in Large Language Models (LLMs) in aligning with human preferences on a specific task/domain. Traditional alignment methods suffer from catastrophic forgetting, where models lose previously learned values when adapting to new preferences or domains. We introduce LifeAlign, a novel framework for lifelong alignment that enables LLMs to maintain consistent human preference alignment across sequential learning tasks without forgetting previously learned values. Our approach consists of two key innovations. First, we propose a focalized preference optimization strategy that aligns LLMs with new preferences while preventing the erosion of alignment acquired from previous tasks. Second, we develop a short-to-long memory consolidation mechanism that merges denoised short-term preference representations into stable long-term memory using intrinsic dimensionality reduction, enabling efficient storage and retrieval of alignment patterns across diverse domains. We evaluate LifeAlign across multiple sequential alignment tasks spanning different domains and preference types. Experimental results demonstrate that our method achieves superior performance in maintaining both preference alignment quality and knowledge retention compared to existing lifelong learning approaches.

Jiayi Li, Xiao Liu, Yansong Feng

Large Language Model (LLM)-based multi-agent systems are increasingly used to simulate human interactions and solve collaborative tasks. A common practice is to assign agents with personas to encourage behavioral diversity. However, this raises a critical yet underexplored question: do personas introduce biases into multi-agent interactions? This paper presents a systematic investigation into persona-induced biases in multi-agent interactions, with a focus on social traits like trustworthiness (how an agent's opinion is received by others) and insistence (how strongly an agent advocates for its opinion). Through a series of controlled experiments in collaborative problem-solving and persuasion tasks, we reveal that (1) LLM-based agents exhibit biases in both trustworthiness and insistence, with personas from historically advantaged groups (e.g., men and White individuals) perceived as less trustworthy and demonstrating less insistence; and (2) agents exhibit significant in-group favoritism, showing a higher tendency to conform to others who share the same persona. These biases persist across various LLMs, group sizes, and numbers of interaction rounds, highlighting an urgent need for awareness and mitigation to ensure the fairness and reliability of multi-agent systems.

Haoyu Li, Xuhong Li, Yiming Dong, Kun Liu

Dataset diversity plays a pivotal role for the successful training of many machine learning models, particularly in the supervised fine-tuning (SFT) stage of large language model (LLM) development. Despite increasing recognition of its importance, systematic analyses of dataset diversity still remain underexplored. To address this gap, this work presents a systematic taxonomy of existing diversity-control strategies, which primarily focus on the instruction component, operating at either macroscopic (entire instruction semantics) or mesoscopic levels (instruction units), and furthermore introduces a novel analysis of microscopic diversity within the response component, specifically analyzing the statistical distribution of tokens in SFT training samples. In the experimental evaluation, we construct fixed-size datasets (e.g., 10,000 samples each) from a corpus of 117,000 open-source SFT samples, incorporating six distinct diversity-control strategies spanning macro-, meso-, and microscopic levels applied to both instructions and responses. We then fine-tune LLMs on these datasets to assess the six diversity-control strategies. Results reveal that while macroscopic and mesoscopic strategies lead to higher performance with increasing diversity, the microscopic strategy in responses exhibits not only a stronger correlation between model performance and the degree of diversity, but also superior performance with maximum diversity across all strategies. These findings offer actionable insights for constructing high-performance SFT datasets.

Haoran Li, Zhiming Su, Junyan Yao, Enwei Zhang, Yang Ji, Yan Chen, Kan Zhou, Chao Feng, Jiao Ran

Synthetic data is widely adopted in embedding models to ensure diversity in training data distributions across dimensions such as difficulty, length, and language. However, existing prompt-based synthesis methods struggle to capture domain-specific data distributions, particularly in data-scarce domains, and often overlook fine-grained relevance diversity. In this paper, we present a Chinese short video dataset with 4-level relevance annotations, filling a critical resource void. Further, we propose a semi-supervised synthetic data pipeline where two collaboratively trained models generate domain-adaptive short video data with controllable relevance labels. Our method enhances relevance-level diversity by synthesizing samples for underrepresented intermediate relevance labels, resulting in a more balanced and semantically rich training data set. Extensive offline experiments show that the embedding model trained on our synthesized data outperforms those using data generated based on prompting or vanilla supervised fine-tuning(SFT). Moreover, we demonstrate that incorporating more diverse fine-grained relevance levels in training data enhances the model's sensitivity to subtle semantic distinctions, highlighting the value of fine-grained relevance supervision in embedding learning. In the search enhanced recommendation pipeline of Douyin's dual-column scenario, through online A/B testing, the proposed model increased click-through rate(CTR) by 1.45%, raised the proportion of Strong Relevance Ratio (SRR) by 4.9%, and improved the Image User Penetration Rate (IUPR) by 0.1054%.

Hao Li, Tao He, Jiafeng Liang, Zheng Chu, Ming Liu

Large Language Models (LLMs) have demonstrated remarkable performance across a wide range of tasks, yet they generally lack self-awareness, often displaying overconfidence when confronted with questions beyond their knowledge boundaries. This limitation severely hinders their trustworthiness in high-stakes scenarios. Existing calibration methods typically rely on sampling accuracy, derived from multiple outputs, as a proxy for model confidence. However, this coarse-grained metric fails to capture the model’s internal cognitive states, such as confusion, hallucination, or persistent belief in false knowledge. To address this, we propose CogConf (Cognitive Confidence), a cognitively grounded uncertainty signal that extends sampling accuracy by incorporating the semantic diversity of incorrect answers and the model’s abstention behaviors. By shifting the focus from sampling-based to cognition-oriented uncertainty modeling, CogConf offers a more faithful reflection of the model's internal beliefs. Building on this signal, we introduce CogAlign, a simple yet effective alignment framework that explicitly aligns the model’s verbalized confidence with CogConf, thereby producing uncertainty estimates that better reflect the model’s internal cognition. Experimental results on six knowledge-intensive in-domain and out-of-domain QA datasets demonstrate that CogConf robustly characterizes the model's internal uncertainty. Building on this foundation, CogAlign guides the model's expression to significantly enhance the trustworthiness and utility of its uncertainty calibration without compromising its underlying QA capabilities, while also demonstrating strong cross-task generalization and output stability. Offering a new pathway toward building more trustworthy LLMs.

Guozheng Li, Xinyu Zhang

Self-training large language models (LLMs) with generated reasoning paths has emerged as a promising approach to improve performance on complex reasoning tasks. However, most existing methods rely on correctness-based supervision, treating samples that reach the correct answer as high-quality despite potentially flawed intermediate steps, leading to noisy training signals. In this work, we propose K-STaR (Knowledge-aware Self-Taught Reasoner), a self-training framework that verifies reasoning paths through knowledge elicitation and integration as a proxy, without requiring any external reward models or dense step-by-step annotations. K-STaR models reasoning as a structured composition of knowledge units and automatically assigns process rewards to intermediate steps via consistency and frequency analysis, ensuring that only knowledge-grounded reasoning paths are retained. Experiments on mathematical and commonsense reasoning tasks show that K-STaR consistently discovers higher-quality reasoning paths and achieves superior self-training performance compared to prior methods. Our results highlight the importance of moving beyond correctness-centric supervision toward knowledge-grounded self-improvement.

Fei Li, Song Liu, Weiguo Wu, Shiqiang Nie, Jinyu Wang

The high memory demands of the Key-Value (KV) Cache during the inference of Large Language Models (LLMs) severely restrict their deployment in resource-constrained platforms. Quantization can effectively alleviate the memory pressure caused by KV Cache. However, existing methods either rely on static one-size-fits-all precision allocation or fail to dynamically prioritize critical KV in long-context tasks, forcing memory-accuracy-throughput tradeoffs. In this work, we propose a novel mixed-precision quantization method for KV Cache named KVmix. KVmix leverages gradient-based importance analysis to evaluate how individual Key and Value projection matrices affect the model loss, enabling layer-specific bit-width allocation for mix-precision quantization. It dynamically prioritizes higher precision for important layers while aggressively quantizing less influential ones, achieving a tunable balance between accuracy and efficiency. KVmix introduces a dynamic long-context optimization strategy that adaptively keeps full-precision KV pairs for recent pivotal tokens and compresses older ones, achieving high-quality sequence generation with low memory usage. Additionally, KVmix provides efficient low-bit quantization and CUDA kernels to optimize computational overhead. On LLMs such as Llama and Mistral, KVmix achieves near-lossless inference performance with extremely low quantization configuration (Key 2.19bit Value 2.38bit), while delivering a remarkable 4.9× memory compression and a 5.3× speedup in inference throughput.