Zuchao Li

Luohe Shi, Hongyi Zhang, Yao Yao et al.

Large Language Models (LLMs), epitomized by ChatGPT's release in late 2022, have revolutionized various industries with their advanced language comprehension. However, their efficiency is challenged by the Transformer architecture's struggle with handling long texts. KV Cache has emerged as a pivotal solution to this issue, converting the time complexity of token generation from quadratic to linear, albeit with increased GPU memory overhead proportional to conversation length. With the development of the LLM community and academia, various KV Cache compression methods have been proposed. In this review, we dissect the various properties of KV Cache and elaborate on various methods currently used to optimize the KV Cache space usage of LLMs. These methods span the pre-training phase, deployment phase, and inference phase, and we summarize the commonalities and differences among these methods. Additionally, we list some metrics for evaluating the long-text capabilities of large language models, from both efficiency and capability perspectives. Our review thus sheds light on the evolving landscape of LLM optimization, offering insights into future advancements in this dynamic field. Links to the papers mentioned in this review can be found in our Github Repo https://github.com/zcli-charlie/Awesome-KV-Cache.

Cong Yang, Zuchao Li, Hongzan Jiao et al.

Recently, while significant progress has been made in remote sensing image change captioning, existing methods fail to filter out areas unrelated to actual changes, making models susceptible to irrelevant features. In this article, we propose a novel multimodal framework for remote sensing image change captioning, guided by Key Change Features and Instruction-tuned (KCFI). This framework aims to fully leverage the intrinsic knowledge of large language models through visual instructions and enhance the effectiveness and accuracy of change features using pixel-level change detection tasks. Specifically, KCFI includes a ViTs encoder for extracting bi-temporal remote sensing image features, a key feature perceiver for identifying critical change areas, a pixel-level change detection decoder to constrain key change features, and an instruction-tuned decoder based on a large language model. Moreover, to ensure that change description and change detection tasks are jointly optimized, we employ a dynamic weight-averaging strategy to balance the losses between the two tasks. We also explore various feature combinations for visual fine-tuning instructions and demonstrate that using only key change features to guide the large language model is the optimal choice. To validate the effectiveness of our approach, we compare it against several state-of-the-art change captioning methods on the LEVIR-CC dataset, achieving the best performance. Our code will be available at https://github.com/yangcong356/KCFI.git.

Zihong Zhang, Zuchao Li, Lefei Zhang et al.

Autoregressive decoding in Large Language Models (LLMs) generates one token per step, causing high inference latency. Speculative decoding (SD) mitigates this through a guess-and-verify strategy, but existing training-free variants face trade-offs: retrieval-based drafts break when no exact match exists, while logits-based drafts lack structural guidance. We propose $\textbf{RACER}$ ($\textbf{R}$etrieval-$\textbf{A}$ugmented $\textbf{C}$ont$\textbf{e}$xtual $\textbf{R}$apid Speculative Decoding), a lightweight and training-free method that integrates retrieved exact patterns with logit-driven future cues. This unification supplies both reliable anchors and flexible extrapolation, yielding richer speculative drafts. Experiments on Spec-Bench, HumanEval, and MGSM-ZH demonstrate that RACER consistently accelerates inference, achieving more than $2\times$ speedup over autoregressive decoding, and outperforms prior training-free methods, offering a scalable, plug-and-play solution for efficient LLM decoding. Our source code is available at $\href{https://github.com/hkr04/RACER}{https://github.com/hkr04/RACER}$.

Yi Zhao, Yajuan Peng, Cam-Tu Nguyen et al.

Large Reasoning Models (LRMs) achieve strong performance on complex tasks through extended chains of thought but suffer from high inference latency due to autoregressive reasoning. Recent work explores using Small Reasoning Models (SRMs) to accelerate LRM inference. In this paper, we systematically characterize the capability boundaries of SRMs and identify three common types of reasoning risks: (1) path divergence, where SRMs lack the strategic ability to construct an initial plan, causing reasoning to deviate from the most probable path; (2) cognitive overload, where SRMs fail to solve particularly difficult steps; and (3) recovery inability, where SRMs lack robust self-reflection and error correction mechanisms. To address these challenges, we propose TrigReason, a trigger-based collaborative reasoning framework that replaces continuous polling with selective intervention. TrigReason delegates most reasoning to the SRM and activates LRM intervention only when necessary-during initial strategic planning (strategic priming trigger), upon detecting extraordinary overconfidence (cognitive offload trigger), or when reasoning falls into unproductive loops (intervention request trigger). The evaluation results on AIME24, AIME25, and GPQA-D indicate that TrigReason matches the accuracy of full LRMs and SpecReason, while offloading 1.70x - 4.79x more reasoning steps to SRMs. Under edge-cloud conditions, TrigReason reduces latency by 43.9\% and API cost by 73.3\%. Our code is available at \href{https://github.com/QQQ-yi/TrigReason}{https://github.com/QQQ-yi/TrigReason}

Tianshuo Peng, Zuchao Li, Lefei Zhang et al.

Universal Information Extraction (UIE) has been introduced as a unified framework for various Information Extraction (IE) tasks and has achieved widespread success. Despite this, UIE models have limitations. For example, they rely heavily on span boundaries in the data during training, which does not reflect the reality of span annotation challenges. Slight adjustments to positions can also meet requirements. Additionally, UIE models lack attention to the limited span length feature in IE. To address these deficiencies, we propose the Fuzzy Span Universal Information Extraction (FSUIE) framework. Specifically, our contribution consists of two concepts: fuzzy span loss and fuzzy span attention. Our experimental results on a series of main IE tasks show significant improvement compared to the baseline, especially in terms of fast convergence and strong performance with small amounts of data and training epochs. These results demonstrate the effectiveness and generalization of FSUIE in different tasks, settings, and scenarios.

Yineng Chen, Zuchao Li, Lefei Zhang et al.

Optimizer is an essential component for the success of deep learning, which guides the neural network to update the parameters according to the loss on the training set. SGD and Adam are two classical and effective optimizers on which researchers have proposed many variants, such as SGDM and RAdam. In this paper, we innovatively combine the backward-looking and forward-looking aspects of the optimizer algorithm and propose a novel \textsc{Admeta} (\textbf{A} \textbf{D}ouble exponential \textbf{M}oving averag\textbf{E} \textbf{T}o \textbf{A}daptive and non-adaptive momentum) optimizer framework. For backward-looking part, we propose a DEMA variant scheme, which is motivated by a metric in the stock market, to replace the common exponential moving average scheme. While in the forward-looking part, we present a dynamic lookahead strategy which asymptotically approaches a set value, maintaining its speed at early stage and high convergence performance at final stage. Based on this idea, we provide two optimizer implementations, \textsc{AdmetaR} and \textsc{AdmetaS}, the former based on RAdam and the latter based on SGDM. Through extensive experiments on diverse tasks, we find that the proposed \textsc{Admeta} optimizer outperforms our base optimizers and shows advantages over recently proposed competitive optimizers. We also provide theoretical proof of these two algorithms, which verifies the convergence of our proposed \textsc{Admeta}.

Zhuosheng Zhang, Kehai Chen, Rui Wang et al.

Representation learning is the foundation of natural language processing (NLP). This work presents new methods to employ visual information as assistant signals to general NLP tasks. For each sentence, we first retrieve a flexible number of images either from a light topic-image lookup table extracted over the existing sentence-image pairs or a shared cross-modal embedding space that is pre-trained on out-of-shelf text-image pairs. Then, the text and images are encoded by a Transformer encoder and convolutional neural network, respectively. The two sequences of representations are further fused by an attention layer for the interaction of the two modalities. In this study, the retrieval process is controllable and flexible. The universal visual representation overcomes the lack of large-scale bilingual sentence-image pairs. Our method can be easily applied to text-only tasks without manually annotated multimodal parallel corpora. We apply the proposed method to a wide range of natural language generation and understanding tasks, including neural machine translation, natural language inference, and semantic similarity. Experimental results show that our method is generally effective for different tasks and languages. Analysis indicates that the visual signals enrich textual representations of content words, provide fine-grained grounding information about the relationship between concepts and events, and potentially conduce to disambiguation.

Meng Lan, Fu Rong, Zuchao Li et al.

Referring video object segmentation (RVOS) aims to segment the target object in a video sequence described by a language expression. Typical multimodal Transformer based RVOS approaches process video sequence in a frame-independent manner to reduce the high computational cost, which however restricts the performance due to the lack of inter-frame interaction for temporal coherence modeling and spatio-temporal representation learning of the referred object. Besides, the absence of sufficient cross-modal interactions results in weak correlation between the visual and linguistic features, which increases the difficulty of decoding the target information and limits the performance of the model. In this paper, we propose a bidirectional correlation-driven inter-frame interaction Transformer, dubbed BIFIT, to address these issues in RVOS. Specifically, we design a lightweight and plug-and-play inter-frame interaction module in the Transformer decoder to efficiently learn the spatio-temporal features of the referred object, so as to decode the object information in the video sequence more precisely and generate more accurate segmentation results. Moreover, a bidirectional vision-language interaction module is implemented before the multimodal Transformer to enhance the correlation between the visual and linguistic features, thus facilitating the language queries to decode more precise object information from visual features and ultimately improving the segmentation performance. Extensive experimental results on four benchmarks validate the superiority of our BIFIT over state-of-the-art methods and the effectiveness of our proposed modules.

Luohe Shi, Yao Yao, Zuchao Li et al.

Large language models (LLMs) have rapidly advanced and demonstrated impressive capabilities. In-Context Learning (ICL) and Parameter-Efficient Fine-Tuning (PEFT) are currently two mainstream methods for augmenting LLMs to downstream tasks. ICL typically constructs a few-shot learning scenario, either manually or by setting up a Retrieval-Augmented Generation (RAG) system, helping models quickly grasp domain knowledge or question-answering patterns without changing model parameters. However, this approach involves trade-offs, such as slower inference speed and increased space occupancy. PEFT assists the model in adapting to tasks through minimal parameter modifications, but the training process still demands high hardware requirements, even with a small number of parameters involved. To address these challenges, we propose Reference Trustable Decoding (RTD), a paradigm that allows models to quickly adapt to new tasks without fine-tuning, maintaining low inference costs. RTD constructs a reference datastore from the provided training examples and optimizes the LLM's final vocabulary distribution by flexibly selecting suitable references based on the input, resulting in more trustable responses and enabling the model to adapt to downstream tasks at a low cost. Experimental evaluations on various LLMs using different benchmarks demonstrate that RTD establishes a new paradigm for augmenting models to downstream tasks. Furthermore, our method exhibits strong orthogonality with traditional methods, allowing for concurrent usage. Our code can be found at https://github.com/ShiLuohe/ReferenceTrustableDecoding

Xin Yan, Zuchao Li, Lefei Zhang

Masked Image Modeling (MIM) is a new self-supervised vision pre-training paradigm using a Vision Transformer (ViT). Previous works can be pixel-based or token-based, using original pixels or discrete visual tokens from parametric tokenizer models, respectively. Our proposed centroid-based approach, CCViT, leverages k-means clustering to obtain centroids for image modeling without supervised training of the tokenizer model, which only takes seconds to create. This non-parametric centroid tokenizer only takes seconds to create and is faster for token inference. The centroids can represent both patch pixels and index tokens with the property of local invariance. Specifically, we adopt patch masking and centroid replacing strategies to construct corrupted inputs, and two stacked encoder blocks to predict corrupted patch tokens and reconstruct original patch pixels. Experiments show that our CCViT achieves 84.4% top-1 accuracy on ImageNet-1K classification with ViT-B and 86.0% with ViT-L. We also transfer our pre-trained model to other downstream tasks. Our approach achieves competitive results with recent baselines without external supervision and distillation training from other models.

Zuchao Li, Shitou Zhang, Hai Zhao et al.

BatGPT is a large-scale language model designed and trained jointly by Wuhan University and Shanghai Jiao Tong University. It is capable of generating highly natural and fluent text in response to various types of input, including text prompts, images, and audio. In the modeling level, we employ a bidirectional autoregressive architecture that allows the model to efficiently capture the complex dependencies of natural language, making it highly effective in tasks such as language generation, dialog systems, and question answering. Moreover, the bidirectional autoregressive modeling not only operates from left to right but also from right to left, effectively reducing fixed memory effects and alleviating model hallucinations. In the training aspect, we propose a novel parameter expansion method for leveraging the pre-training of smaller models and employ reinforcement learning from both AI and human feedback, aimed at improving the model's alignment performance. Overall, these approaches significantly improve the effectiveness of BatGPT, and the model can be utilized for a wide range of natural language applications.

Shitou Zhang, Jingrui Hou, Siyuan Peng et al.

Archives play a crucial role in preserving information and knowledge, and the exponential growth of such data necessitates efficient and automated tools for managing and utilizing archive information resources. Archival applications involve managing massive data that are challenging to process and analyze. Although LLMs have made remarkable progress in diverse domains, there are no publicly available archives tailored LLM. Addressing this gap, we introduce ArcGPT, to our knowledge, the first general-purpose LLM tailored to the archival field. To enhance model performance on real-world archival tasks, ArcGPT has been pre-trained on massive and extensive archival domain data. Alongside ArcGPT, we release AMBLE, a benchmark comprising four real-world archival tasks. Evaluation on AMBLE shows that ArcGPT outperforms existing state-of-the-art models, marking a substantial step forward in effective archival data management. Ultimately, ArcGPT aims to better serve the archival community, aiding archivists in their crucial role of preserving and harnessing our collective information and knowledge.

Qiwei Li, Zuchao Li, Xiantao Cai et al.

In recent years, the use of multi-modal pre-trained Transformers has led to significant advancements in visually-rich document understanding. However, existing models have mainly focused on features such as text and vision while neglecting the importance of layout relationship between text nodes. In this paper, we propose GraphLayoutLM, a novel document understanding model that leverages the modeling of layout structure graph to inject document layout knowledge into the model. GraphLayoutLM utilizes a graph reordering algorithm to adjust the text sequence based on the graph structure. Additionally, our model uses a layout-aware multi-head self-attention layer to learn document layout knowledge. The proposed model enables the understanding of the spatial arrangement of text elements, improving document comprehension. We evaluate our model on various benchmarks, including FUNSD, XFUND and CORD, and achieve state-of-the-art results among these datasets. Our experimental results demonstrate that our proposed method provides a significant improvement over existing approaches and showcases the importance of incorporating layout information into document understanding models. We also conduct an ablation study to investigate the contribution of each component of our model. The results show that both the graph reordering algorithm and the layout-aware multi-head self-attention layer play a crucial role in achieving the best performance.

Yifei Yang, Runhan Shi, Zuchao Li et al.

Retrosynthesis analysis is pivotal yet challenging in drug discovery and organic chemistry. Despite the proliferation of computational tools over the past decade, AI-based systems often fall short in generalizing across diverse reaction types and exploring alternative synthetic pathways. This paper presents BatGPT-Chem, a large language model with 15 billion parameters, tailored for enhanced retrosynthesis prediction. Integrating chemical tasks via a unified framework of natural language and SMILES notation, this approach synthesizes extensive instructional data from an expansive chemical database. Employing both autoregressive and bidirectional training techniques across over one hundred million instances, BatGPT-Chem captures a broad spectrum of chemical knowledge, enabling precise prediction of reaction conditions and exhibiting strong zero-shot capabilities. Superior to existing AI methods, our model demonstrates significant advancements in generating effective strategies for complex molecules, as validated by stringent benchmark tests. BatGPT-Chem not only boosts the efficiency and creativity of retrosynthetic analysis but also establishes a new standard for computational tools in synthetic design. This development empowers chemists to adeptly address the synthesis of novel compounds, potentially expediting the innovation cycle in drug manufacturing and materials science. We release our trial platform at \url{https://www.batgpt.net/dapp/chem}.

Jiani Guo, Zuchao Li, Jie Wu et al.

Large Language Models (LLMs), constrained by limited context windows, often face significant performance degradation when reasoning over long contexts. To address this, Retrieval-Augmented Generation (RAG) retrieves and reasons over chunks but frequently sacrifices logical coherence due to its reliance on similarity-based rankings. Similarly, divide-and-conquer frameworks (DCF) split documents into small chunks for independent reasoning and aggregation. While effective for local reasoning, DCF struggles to capture long-range dependencies and risks inducing conflicts by processing chunks in isolation. To overcome these limitations, we propose ToM, a novel Tree-oriented MapReduce framework for long-context reasoning. ToM leverages the inherent hierarchical structure of long documents (e.g., main headings and subheadings) by constructing a DocTree through hierarchical semantic parsing and performing bottom-up aggregation. Using a Tree MapReduce approach, ToM enables recursive reasoning: in the Map step, rationales are generated at child nodes; in the Reduce step, these rationales are aggregated across sibling nodes to resolve conflicts or reach consensus at parent nodes. Experimental results on 70B+ LLMs show that ToM significantly outperforms existing divide-and-conquer frameworks and retrieval-augmented generation methods, achieving better logical coherence and long-context reasoning. Our code is available at https://github.com/gjn12-31/ToM .

Qiwei Li, Zuchao Li, Ping Wang et al.

Semantic entity recognition is an important task in the field of visually-rich document understanding. It distinguishes the semantic types of text by analyzing the position relationship between text nodes and the relation between text content. The existing document understanding models mainly focus on entity categories while ignoring the extraction of entity boundaries. We build a novel hypergraph attention document semantic entity recognition framework, HGA, which uses hypergraph attention to focus on entity boundaries and entity categories at the same time. It can conduct a more detailed analysis of the document text representation analyzed by the upstream model and achieves a better performance of semantic information. We apply this method on the basis of GraphLayoutLM to construct a new semantic entity recognition model HGALayoutLM. Our experiment results on FUNSD, CORD, XFUND and SROIE show that our method can effectively improve the performance of semantic entity recognition tasks based on the original model. The results of HGALayoutLM on FUNSD and XFUND reach the new state-of-the-art results.

Letian Peng, Zuchao Li, Hai Zhao

We report the performance of DeBERTaV3 on CommonsenseQA in this report. We simply formalize the answer selection as a text classification for DeBERTaV3. The strong natural language inference ability of DeBERTaV3 helps its single and ensemble model set the new (w/o external knowledge) state-of-the-art on CommonsenseQA.

Letian Peng, Zuchao Li, Hai Zhao

Commonsense reasoning is an appealing topic in natural language processing (NLP) as it plays a fundamental role in supporting the human-like actions of NLP systems. With large-scale language models as the backbone, unsupervised pre-training on numerous corpora shows the potential to capture commonsense knowledge. Current pre-trained language model (PLM)-based reasoning follows the traditional practice using perplexity metric. However, commonsense reasoning is more than existing probability evaluation, which is biased by word frequency. This paper reconsiders the nature of commonsense reasoning and proposes a novel commonsense reasoning metric, Non-Replacement Confidence (NRC). In detail, it works on PLMs according to the Replaced Token Detection (RTD) pre-training objective in ELECTRA, in which the corruption detection objective reflects the confidence on contextual integrity that is more relevant to commonsense reasoning than existing probability. Our proposed novel method boosts zero-shot performance on two commonsense reasoning benchmark datasets and further seven commonsense question-answering datasets. Our analysis shows that pre-endowed commonsense knowledge, especially for RTD-based PLMs, is essential in downstream reasoning.

Xiangyu Ma, Teng Xiao, Zuchao Li et al.

Diffusion models promise efficient parallel text generation but rely on bidirectional attention, creating a structural mismatch with pre-trained Autoregressive (AR) models. This incompatibility precludes reusing robust AR priors, necessitating prohibitive pre-training from scratch. To bridge this gap, we propose FLUID, a framework that efficiently adapts AR backbones to the diffusion paradigm. By enforcing Strictly Causal Alignment, FLUID enables seamless initialization from standard GPT-style checkpoints, circumventing the need for massive pre-training. Furthermore, we introduce Elastic Horizons, an entropy-driven mechanism that dynamically modulates denoising strides based on local information density rather than fixed schedules. Experiments demonstrate that FLUID achieves state-of-the-art performance while reducing training costs by orders of magnitude, effectively reconciling established AR foundations with efficient parallel generation. Our code is available at https://github.com/Oli-lab-nun/FLUID/tree/main.

Yifei Yang, Zuchao Li, Hai Zhao

As a fundamental natural language processing task and one of core knowledge extraction techniques, named entity recognition (NER) is widely used to extract information from texts for downstream tasks. Nested NER is a branch of NER in which the named entities (NEs) are nested with each other. However, most of the previous studies on nested NER usually apply linear structure to model the nested NEs which are actually accommodated in a hierarchical structure. Thus in order to address this mismatch, this work models the full nested NEs in a sentence as a holistic structure, then we propose a holistic structure parsing algorithm to disclose the entire NEs once for all. Besides, there is no research on applying corpus-level information to NER currently. To make up for the loss of this information, we introduce Point-wise Mutual Information (PMI) and other frequency features from corpus-aware statistics for even better performance by holistic modeling from sentence-level to corpus-level. Experiments show that our model yields promising results on widely-used benchmarks which approach or even achieve state-of-the-art. Further empirical studies show that our proposed corpus-aware features can substantially improve NER domain adaptation, which demonstrates the surprising advantage of our proposed corpus-level holistic structure modeling.

Yao Yao, Zuchao Li, Hai Zhao

As Large Language Models (LLMs) become increasingly prevalent in various domains, their ability to process inputs of any length and maintain a degree of memory becomes essential. However, the one-off input of overly long texts is limited, as studies have shown that when input lengths exceed the LLMs' pre-trained text length, there is a dramatic decline in text generation capabilities. Moreover, simply extending the length of pre-training texts is impractical due to the difficulty in obtaining long text data and the substantial memory consumption costs this would entail for LLMs. Recent efforts have employed streaming inputs to alleviate the pressure of excessively long text inputs, but this approach can significantly impair the model's long-term memory capabilities. Motivated by this challenge, we introduce Streaming Infinite Retentive LLM (SirLLM), which allows LLMs to maintain longer memory during infinite-length dialogues without the need for fine-tuning. SirLLM utilizes the Token Entropy metric and a memory decay mechanism to filter key phrases, endowing LLMs with both long-lasting and flexible memory. We designed three distinct tasks and constructed three datasets to measure the effectiveness of SirLLM from various angles: (1) DailyDialog; (2) Grocery Shopping; (3) Rock-Paper-Scissors. Our experimental results robustly demonstrate that SirLLM can achieve stable and significant improvements across different LLMs and tasks, compellingly proving its effectiveness. When having a coversation, "A sir could forget himself," but SirLLM never does! Our code is publicly available at https://github.com/Zoeyyao27/SirLLM

Weixi Song, Zuchao Li, Lefei Zhang et al.

With the prevalence of pre-training-fine-tuning paradigm, how to efficiently adapt the pre-trained model to the downstream tasks has been an intriguing issue. Parameter-Efficient Fine-Tuning (PEFT) methods have been proposed for low-cost adaptation. Although PEFT has demonstrated effectiveness and been widely applied, the underlying principles are still unclear. In this paper, we adopt the PAC-Bayesian generalization error bound, viewing pre-training as a shift of prior distribution which leads to a tighter bound for generalization error. We validate this shift from the perspectives of oscillations in the loss landscape and the quasi-sparsity in gradient distribution. Based on this, we propose a gradient-based sparse fine-tuning algorithm, named Sparse Increment Fine-Tuning (SIFT), and validate its effectiveness on a range of tasks including the GLUE Benchmark and Instruction-tuning. The code is accessible at https://github.com/song-wx/SIFT/.

Jiaqi Chen, Xiaoye Zhu, Tianyang Liu et al.

Large Language Models (LLMs) have revolutionized text generation, making detecting machine-generated text increasingly challenging. Although past methods have achieved good performance on detecting pure machine-generated text, those detectors have poor performance on distinguishing machine-revised text (rewriting, expansion, and polishing), which can have only minor changes from its original human prompt. As the content of text may originate from human prompts, detecting machine-revised text often involves identifying distinctive machine styles, e.g., worded favored by LLMs. However, existing methods struggle to detect machine-style phrasing hidden within the content contributed by humans. We propose the "Imitate Before Detect" (ImBD) approach, which first imitates the machine-style token distribution, and then compares the distribution of the text to be tested with the machine-style distribution to determine whether the text has been machine-revised. To this end, we introduce style preference optimization (SPO), which aligns a scoring LLM model to the preference of text styles generated by machines. The aligned scoring model is then used to calculate the style-conditional probability curvature (Style-CPC), quantifying the log probability difference between the original and conditionally sampled texts for effective detection. We conduct extensive comparisons across various scenarios, encompassing text revisions by six LLMs, four distinct text domains, and three machine revision types. Compared to existing state-of-the-art methods, our method yields a 13% increase in AUC for detecting text revised by open-source LLMs, and improves performance by 5% and 19% for detecting GPT-3.5 and GPT-4o revised text, respectively. Notably, our method surpasses the commercially trained GPT-Zero with just $1,000$ samples and five minutes of SPO, demonstrating its efficiency and effectiveness.

Suqing Wang, Ziyang Ma, Xinyi Li et al.

Large Language Models (LLMs) have rapidly advanced and are widely adopted across diverse fields. Due to the substantial computational cost and data requirements of training from scratch, many developers choose to fine-tune or modify existing open-source models. While most adhere to open-source licenses, some falsely claim original training despite clear derivation from public models. This raises pressing concerns about intellectual property protection and highlights the need for reliable methods to verify model provenance. In this paper, we propose GhostSpec, a lightweight yet effective method for verifying LLM lineage without access to training data or modification of model behavior. Our approach constructs compact and robust fingerprints by applying singular value decomposition (SVD) to invariant products of internal attention weight matrices, effectively capturing the structural identity of a model. Unlike watermarking or output-based methods, GhostSpec is fully data-free, non-invasive, and computationally efficient. It demonstrates strong robustness to sequential fine-tuning, pruning, block expansion, and even adversarial transformations. Extensive experiments show that GhostSpec can reliably trace the lineage of transformed models with minimal overhead. By offering a practical solution for model verification and reuse tracking, our method contributes to the protection of intellectual property and fosters a transparent, trustworthy ecosystem for large-scale language models.

Jinhao Tian, Zuchao Li, Jiajia Li et al.

The first step to apply deep learning techniques for symbolic music understanding is to transform musical pieces (mainly in MIDI format) into sequences of predefined tokens like note pitch, note velocity, and chords. Subsequently, the sequences are fed into a neural sequence model to accomplish specific tasks. Music sequences exhibit strong correlations between adjacent elements, making them prime candidates for N-gram techniques from Natural Language Processing (NLP). Consider classical piano music: specific melodies might recur throughout a piece, with subtle variations each time. In this paper, we propose a novel method, NG-Midiformer, for understanding symbolic music sequences that leverages the N-gram approach. Our method involves first processing music pieces into word-like sequences with our proposed unsupervised compoundation, followed by using our N-gram Transformer encoder, which can effectively incorporate N-gram information to enhance the primary encoder part for better understanding of music sequences. The pre-training process on large-scale music datasets enables the model to thoroughly learn the N-gram information contained within music sequences, and subsequently apply this information for making inferences during the fine-tuning stage. Experiment on various datasets demonstrate the effectiveness of our method and achieved state-of-the-art performance on a series of music understanding downstream tasks. The code and model weights will be released at https://github.com/CinqueOrigin/NG-Midiformer.

Luohe Shi, Zuchao Li, Lefei Zhang et al.

Large language models (LLMs) based on Transformer Decoders have become the preferred choice for conversational generative AI. Despite the overall superiority of the Decoder architecture, the gradually increasing Key-Value (KV) cache during inference has emerged as a primary efficiency bottleneck, both in aspects of memory consumption and data transfer bandwidth limitations. To address these challenges, we propose a paradigm called KV-Latent. By down-sampling the Key-Value vector dimensions into a latent space, we can significantly reduce the KV Cache footprint and improve inference speed, only with a small amount of extra training, less than 1\% of pre-training takes. Besides, we enhanced the stability of Rotary Positional Embedding applied on lower-dimensional vectors by modifying its frequency sampling mechanism, avoiding noise introduced by higher frequencies while retaining position attenuation. Our experiments, including both models with Grouped Query Attention and those without, have yielded satisfactory results. Finally, we conducted comparative experiments to study the impact of separately reducing Key and Value components on model's performance. Our approach allows for the construction of more efficient language model systems, and opens the new possibility on KV Cache saving and efficient LLMs. Our code is available at https://github.com/ShiLuohe/KV-Latent.

Xiangke Zeng, Zuchao Li, Lefei Zhang et al.

Chinese Spelling Correction (CSC) stands as a foundational Natural Language Processing (NLP) task, which primarily focuses on the correction of erroneous characters in Chinese texts. Certain existing methodologies opt to disentangle the error correction process, employing an additional error detector to pinpoint error positions. However, owing to the inherent performance limitations of error detector, precision and recall are like two sides of the coin which can not be both facing up simultaneously. Furthermore, it is also worth investigating how the error position information can be judiciously applied to assist the error correction. In this paper, we introduce a novel approach based on error detector-corrector framework. Our detector is designed to yield two error detection results, each characterized by high precision and recall. Given that the occurrence of errors is context-dependent and detection outcomes may be less precise, we incorporate the error detection results into the CSC task using an innovative feature fusion strategy and a selective masking strategy. Empirical experiments conducted on mainstream CSC datasets substantiate the efficacy of our proposed method.

Jie Wu, Haoling Li, Xin Zhang et al.

Preference learning extends the performance of Code LLMs beyond traditional supervised fine-tuning by leveraging relative quality comparisons. In existing approaches, a set of n candidate solutions is evaluated based on test case success rates, with the candidate demonstrating a higher pass rate being labeled as positive and its counterpart with a lower pass rate as negative. However, because this approach aligns entire failing code blocks rather than pinpointing specific errors, it lacks the granularity necessary to capture meaningful error-correction relationships. As a result, the model is unable to learn more informative error-correction patterns. To address these issues, we propose Target-DPO, a new preference alignment framework that mimics human iterative debugging to refine Code LLMs. Target-DPO explicitly locates error regions and aligns the corresponding tokens via a tailored DPO algorithm. To facilitate it, we introduce the CodeFlow dataset, where samples are iteratively refined until passing tests, with modifications capturing error corrections. Extensive experiments show that a diverse suite of Code LLMs equipped with Target-DPO achieves significant performance gains in code generation and improves on challenging tasks like BigCodeBench. In-depth analysis reveals that Target-DPO yields fewer errors. Code, model and datasets are in: https://github.com/JieWu02/Target-DPO.

Shitou Zhang, Zuchao Li, Xingshen Liu et al.

In light of the rapidly evolving capabilities of large language models (LLMs), it becomes imperative to develop rigorous domain-specific evaluation benchmarks to accurately assess their capabilities. In response to this need, this paper introduces ArcMMLU, a specialized benchmark tailored for the Library & Information Science (LIS) domain in Chinese. This benchmark aims to measure the knowledge and reasoning capability of LLMs within four key sub-domains: Archival Science, Data Science, Library Science, and Information Science. Following the format of MMLU/CMMLU, we collected over 6,000 high-quality questions for the compilation of ArcMMLU. This extensive compilation can reflect the diverse nature of the LIS domain and offer a robust foundation for LLM evaluation. Our comprehensive evaluation reveals that while most mainstream LLMs achieve an average accuracy rate above 50% on ArcMMLU, there remains a notable performance gap, suggesting substantial headroom for refinement in LLM capabilities within the LIS domain. Further analysis explores the effectiveness of few-shot examples on model performance and highlights challenging questions where models consistently underperform, providing valuable insights for targeted improvements. ArcMMLU fills a critical gap in LLM evaluations within the Chinese LIS domain and paves the way for future development of LLMs tailored to this specialized area.

Lu Yang, Jiajia Li, En Ci et al.

Universal Information Extraction (UIE) has garnered significant attention due to its ability to address model explosion problems effectively. Extractive UIE can achieve strong performance using a relatively small model, making it widely adopted. Extractive UIEs generally rely on task instructions for different tasks, including single-target instructions and multiple-target instructions. Single-target instruction UIE enables the extraction of only one type of relation at a time, limiting its ability to model correlations between relations and thus restricting its capability to extract complex relations. While multiple-target instruction UIE allows for the extraction of multiple relations simultaneously, the inclusion of irrelevant relations introduces decision complexity and impacts extraction accuracy. Therefore, for multi-relation extraction, we propose LDNet, which incorporates multi-aspect relation modeling and a label drop mechanism. By assigning different relations to different levels for understanding and decision-making, we reduce decision confusion. Additionally, the label drop mechanism effectively mitigates the impact of irrelevant relations. Experiments show that LDNet outperforms or achieves competitive performance with state-of-the-art systems on 9 tasks, 33 datasets, in both single-modal and multi-modal, few-shot and zero-shot settings.\footnote{https://github.com/Lu-Yang666/LDNet}

Mingni Tang, Jiajia Li, Lu Yang et al.

Symbolic music is represented in two distinct forms: two-dimensional, visually intuitive score images, and one-dimensional, standardized text annotation sequences. While large language models have shown extraordinary potential in music, current research has primarily focused on unimodal symbol sequence text. Existing general-domain visual language models still lack the ability of music notation understanding. Recognizing this gap, we propose NOTA, the first large-scale comprehensive multimodal music notation dataset. It consists of 1,019,237 records, from 3 regions of the world, and contains 3 tasks. Based on the dataset, we trained NotaGPT, a music notation visual large language model. Specifically, we involve a pre-alignment training phase for cross-modal alignment between the musical notes depicted in music score images and their textual representation in ABC notation. Subsequent training phases focus on foundational music information extraction, followed by training on music notation analysis. Experimental results demonstrate that our NotaGPT-7B achieves significant improvement on music understanding, showcasing the effectiveness of NOTA and the training pipeline. Our datasets are open-sourced at https://huggingface.co/datasets/MYTH-Lab/NOTA-dataset.

Xinhui Chen, Zuchao Li, Mengqi Gao et al.

Deciphering the function of unseen protein sequences is a fundamental challenge with broad scientific impact, yet most existing methods depend on task-specific adapters or large-scale supervised fine-tuning. We introduce the "Protein-as-Second-Language" framework, which reformulates amino-acid sequences as sentences in a novel symbolic language that large language models can interpret through contextual exemplars. Our approach adaptively constructs sequence-question-answer triples that reveal functional cues in a zero-shot setting, without any further training. To support this process, we curate a bilingual corpus of 79,926 protein-QA instances spanning attribute prediction, descriptive understanding, and extended reasoning. Empirically, our method delivers consistent gains across diverse open-source LLMs and GPT-4, achieving up to 17.2% ROUGE-L improvement (average +7%) and even surpassing fine-tuned protein-specific language models. These results highlight that generic LLMs, when guided with protein-as-language cues, can outperform domain-specialized models, offering a scalable pathway for protein understanding in foundation models.

Teng Xiao, Zuchao Li, Lefei Zhang

Recent advances in multimodal large language models (LLMs) have led to significant progress in understanding, generation, and retrieval tasks. However, current solutions often treat these tasks in isolation or require training LLMs from scratch, resulting in high computational costs and limited generalization across modalities. In this work, we present OmniBridge, a unified and modular multimodal framework that supports vision-language understanding, generation, and retrieval within a unified architecture. OmniBridge adopts a language-centric design that reuses pretrained LLMs and introduces a lightweight bidirectional latent alignment module. To address the challenge of task interference, we propose a two-stage decoupled training strategy: supervised fine-tuning and latent space alignment for aligning LLM behavior with multimodal reasoning, and semantic-guided diffusion training to align cross-modal latent spaces via learnable query embeddings. Extensive experiments across a wide range of benchmarks demonstrate that OmniBridge achieves competitive or state-of-the-art performance in all three tasks. Moreover, our results highlight the effectiveness of latent space alignment for unifying multimodal modeling under a shared representation space. Code and models are released at https://github.com/xiao-xt/OmniBridge.

Suqing Wang, Zuchao Li, Luohe Shi et al.

Large language models (LLMs) have achieved remarkable success across various domains, driving significant technological advancements and innovations. Despite the rapid growth in model scale and capability, systematic, data-driven research on how structural configurations affect performance remains scarce. To address this gap, we present a large-scale dataset encompassing diverse open-source LLM structures and their performance across multiple benchmarks. Leveraging this dataset, we conduct a systematic, data mining-driven analysis to validate and quantify the relationship between structural configurations and performance. Our study begins with a review of the historical development of LLMs and an exploration of potential future trends. We then analyze how various structural choices impact performance across benchmarks and further corroborate our findings using mechanistic interpretability techniques. By providing data-driven insights into LLM optimization, our work aims to guide the targeted development and application of future models. We will release our dataset at https://huggingface.co/datasets/DX0369/LLM-Structure-Performance-Dataset

Jiliang Hu, Zuchao Li, Baoyuan Qi et al.

Significant progress has been made in spoken question answering (SQA) in recent years. However, many existing methods, including large audio language models, struggle with processing long audio. Follow the success of retrieval augmented generation, a speech-related retriever shows promising in help preprocessing long-form speech. But the performance of existing speech-related retrievers is lacking. To address this challenge, we propose CLSR, an end-to-end contrastive language-speech retriever that efficiently extracts question-relevant segments from long audio recordings for downstream SQA task. Unlike conventional speech-text contrastive models, CLSR incorporates an intermediate step that converts acoustic features into text-like representations prior to alignment, thereby more effectively bridging the gap between modalities. Experimental results across four cross-modal retrieval datasets demonstrate that CLSR surpasses both end-to-end speech related retrievers and pipeline approaches combining speech recognition with text retrieval, providing a robust foundation for advancing practical long-form SQA applications.

Zihong Zhang, Liqi He, Zuchao Li et al.

Word segmentation stands as a cornerstone of Natural Language Processing (NLP). Based on the concept of "comprehend first, segment later", we propose a new framework to explore the limit of unsupervised word segmentation with Large Language Models (LLMs) and evaluate the semantic understanding capabilities of LLMs based on word segmentation. We employ current mainstream LLMs to perform word segmentation across multiple languages to assess LLMs' "comprehension". Our findings reveal that LLMs are capable of following simple prompts to segment raw text into words. There is a trend suggesting that models with more parameters tend to perform better on multiple languages. Additionally, we introduce a novel unsupervised method, termed LLACA ($\textbf{L}$arge $\textbf{L}$anguage Model-Inspired $\textbf{A}$ho-$\textbf{C}$orasick $\textbf{A}$utomaton). Leveraging the advanced pattern recognition capabilities of Aho-Corasick automata, LLACA innovatively combines these with the deep insights of well-pretrained LLMs. This approach not only enables the construction of a dynamic $n$-gram model that adjusts based on contextual information but also integrates the nuanced understanding of LLMs, offering significant improvements over traditional methods. Our source code is available at https://github.com/hkr04/LLACA

Jiajia Li, Lu Yang, Mingni Tang et al.

Benchmark plays a pivotal role in assessing the advancements of large language models (LLMs). While numerous benchmarks have been proposed to evaluate LLMs' capabilities, there is a notable absence of a dedicated benchmark for assessing their musical abilities. To address this gap, we present ZIQI-Eval, a comprehensive and large-scale music benchmark specifically designed to evaluate the music-related capabilities of LLMs. ZIQI-Eval encompasses a wide range of questions, covering 10 major categories and 56 subcategories, resulting in over 14,000 meticulously curated data entries. By leveraging ZIQI-Eval, we conduct a comprehensive evaluation over 16 LLMs to evaluate and analyze LLMs' performance in the domain of music. Results indicate that all LLMs perform poorly on the ZIQI-Eval benchmark, suggesting significant room for improvement in their musical capabilities. With ZIQI-Eval, we aim to provide a standardized and robust evaluation framework that facilitates a comprehensive assessment of LLMs' music-related abilities. The dataset is available at GitHub\footnote{https://github.com/zcli-charlie/ZIQI-Eval} and HuggingFace\footnote{https://huggingface.co/datasets/MYTH-Lab/ZIQI-Eval}.

Cong Yang, Zuchao Li, Lefei Zhang

Recently, large multimodal models have built a bridge from visual to textual information, but they tend to underperform in remote sensing scenarios. This underperformance is due to the complex distribution of objects and the significant scale differences among targets in remote sensing images, leading to visual ambiguities and insufficient descriptions by these multimodal models. Moreover, the lack of multimodal fine-tuning data specific to the remote sensing field makes it challenging for the model's behavior to align with user queries. To address these issues, this paper proposes an attribute-guided \textbf{Multi-Granularity Instruction Multimodal Model (MGIMM)} for remote sensing image detailed description. MGIMM guides the multimodal model to learn the consistency between visual regions and corresponding text attributes (such as object names, colors, and shapes) through region-level instruction tuning. Then, with the multimodal model aligned on region-attribute, guided by multi-grain visual features, MGIMM fully perceives both region-level and global image information, utilizing large language models for comprehensive descriptions of remote sensing images. Due to the lack of a standard benchmark for generating detailed descriptions of remote sensing images, we construct a dataset featuring 38,320 region-attribute pairs and 23,463 image-detailed description pairs. Compared with various advanced methods on this dataset, the results demonstrate the effectiveness of MGIMM's region-attribute guided learning approach. Code can be available at https://github.com/yangcong356/MGIMM.git

Liqi He, Zuchao Li, Xiantao Cai et al.

Chain-of-thought (CoT) reasoning has exhibited impressive performance in language models for solving complex tasks and answering questions. However, many real-world questions require multi-modal information, such as text and images. Previous research on multi-modal CoT has primarily focused on extracting fixed image features from off-the-shelf vision models and then fusing them with text using attention mechanisms. This approach has limitations because these vision models were not designed for complex reasoning tasks and do not align well with language thoughts. To overcome this limitation, we introduce a novel approach for multi-modal CoT reasoning that utilizes latent space learning via diffusion processes to generate effective image features that align with language thoughts. Our method fuses image features and text representations at a deep level and improves the complex reasoning ability of multi-modal CoT. We demonstrate the efficacy of our proposed method on multi-modal ScienceQA and machine translation benchmarks, achieving state-of-the-art performance on ScienceQA. Overall, our approach offers a more robust and effective solution for multi-modal reasoning in language models, enhancing their ability to tackle complex real-world problems.

Tianshuo Peng, Zuchao Li, Ping Wang et al.

Multi-modal aspect-based sentiment analysis (MABSA) has recently attracted increasing attention. The span-based extraction methods, such as FSUIE, demonstrate strong performance in sentiment analysis due to their joint modeling of input sequences and target labels. However, previous methods still have certain limitations: (i) They ignore the difference in the focus of visual information between different analysis targets (aspect or sentiment). (ii) Combining features from uni-modal encoders directly may not be sufficient to eliminate the modal gap and can cause difficulties in capturing the image-text pairwise relevance. (iii) Existing span-based methods for MABSA ignore the pairwise relevance of target span boundaries. To tackle these limitations, we propose a novel framework called DQPSA for multi-modal sentiment analysis. Specifically, our model contains a Prompt as Dual Query (PDQ) module that uses the prompt as both a visual query and a language query to extract prompt-aware visual information and strengthen the pairwise relevance between visual information and the analysis target. Additionally, we introduce an Energy-based Pairwise Expert (EPE) module that models the boundaries pairing of the analysis target from the perspective of an Energy-based Model. This expert predicts aspect or sentiment span based on pairwise stability. Experiments on three widely used benchmarks demonstrate that DQPSA outperforms previous approaches and achieves a new state-of-the-art performance.

Juncheng Yang, Zuchao Li, Shuai Xie et al.

The chain-of-thought technique has been received well in multi-modal tasks. It is a step-by-step linear reasoning process that adjusts the length of the chain to improve the performance of generated prompts. However, human thought processes are predominantly non-linear, as they encompass multiple aspects simultaneously and employ dynamic adjustment and updating mechanisms. Therefore, we propose a novel Aggregation-Graph-of-Thought (AGoT) mechanism for soft-prompt tuning in multi-modal representation learning. The proposed AGoT models the human thought process not only as a chain but also models each step as a reasoning aggregation graph to cope with the overlooked multiple aspects of thinking in single-step reasoning. This turns the entire reasoning process into prompt aggregation and prompt flow operations. Experiments show that our multi-modal model enhanced with AGoT soft-prompting achieves good results in several tasks such as text-image retrieval, visual question answering, and image recognition. In addition, we demonstrate that it has good domain generalization performance due to better reasoning.

Zicong Tang, Shi Luohe, Zuchao Li et al.

Large Language Models (LLMs) have achieved impressive accomplishments in recent years. However, the increasing memory consumption of KV cache has possessed a significant challenge to the inference system. Eviction methods have revealed the inherent redundancy within the KV cache, demonstrating its potential for reduction, particularly in deeper layers. However, KV cache reduction for shallower layers has been found to be insufficient. Based on our observation that, the KV cache exhibits a high degree of similarity. Based on this observation, we proposed a novel KV cache reduction method, SpindleKV, which balances both shallow and deep layers. For deep layers, we employ an attention weight based eviction method, while for shallow layers, we apply a codebook based replacement approach which is learnt by similarity and merging policy. Moreover, SpindleKV addressed the Grouped-Query Attention (GQA) dilemma faced by other attention based eviction methods. Experiments on two common benchmarks with three different LLMs shown that SpindleKV obtained better KV cache reduction effect compared to baseline methods, while preserving similar or even better model performance.

Tianshuo Peng, Zuchao Li, Lefei Zhang et al.

Large Language Models (LLMs), benefiting from the auto-regressive modelling approach performed on massive unannotated texts corpora, demonstrates powerful perceptual and reasoning capabilities. However, as for extending auto-regressive modelling to multi-modal scenarios to build Large Multi-modal Models (LMMs), there lies a great difficulty that the image information is processed in the LMM as continuous visual embeddings, which cannot obtain discrete supervised labels for classification.In this paper, we successfully perform multi-modal auto-regressive modeling with a unified objective for the first time.Specifically, we propose the concept of visual tokens, which maps the visual features to probability distributions over LLM's vocabulary, providing supervision information for visual modelling.We further explore the distribution of visual features in the semantic space within LMM and the possibility of using text embeddings to represent visual information.Experimental results and ablation studies on 5 VQA tasks and 4 benchmark toolkits validate the powerful performance of our proposed approach.

Yi Zhao, Zuchao Li, Hai Zhao

LLMs encounter significant challenges in resource consumption nowadays, especially with long contexts. Despite extensive efforts dedicate to enhancing inference efficiency, these methods primarily exploit internal sparsity within the models, without leveraging external information for optimization. We identify the high similarity of attention matrices across different-scale LLMs, which offers a novel perspective for optimization. We first conduct a comprehensive analysis of how to measure similarity, how to select mapping Layers and whether mapping is consistency. Based on these insights, we introduce the IAM framework, which achieves dual benefits of accelerated attention computation and reduced KV cache usage by performing attention mapping between small and large LLMs. Our experimental results demonstrate that IAM can accelerate prefill by 15% and reduce KV cache usage by 22.1% without appreciably sacrificing performance. Experiments on different series of models show the generalizability of IAM. Importantly, it is also orthogonal to many existing KV cache optimization methods, making it a versatile addition to the current toolkit for enhancing LLM efficiency.

Juncheng Yang, Zuchao Li, Shuai Xie et al.

Adapter-based parameter-efficient transfer learning has achieved exciting results in vision-language models. Traditional adapter methods often require training or fine-tuning, facing challenges such as insufficient samples or resource limitations. While some methods overcome the need for training by leveraging image modality cache and retrieval, they overlook the text modality's importance and cross-modal cues for the efficient adaptation of parameters in visual-language models. This work introduces a cross-modal parameter-efficient approach named XMAdapter. XMAdapter establishes cache models for both text and image modalities. It then leverages retrieval through visual-language bimodal information to gather clues for inference. By dynamically adjusting the affinity ratio, it achieves cross-modal fusion, decoupling different modal similarities to assess their respective contributions. Additionally, it explores hard samples based on differences in cross-modal affinity and enhances model performance through adaptive adjustment of sample learning intensity. Extensive experimental results on benchmark datasets demonstrate that XMAdapter outperforms previous adapter-based methods significantly regarding accuracy, generalization, and efficiency.

Haoqi Yang, Luohe Shi, Qiwei Li et al.

Sparse Mixture of Experts (MoE) large language models (LLMs) are gradually becoming the mainstream approach for ultra-large-scale models. Existing optimization efforts for MoE models have focused primarily on coarse-grained MoE architectures. With the emergence of DeepSeek Models, fine-grained MoE models are gaining popularity, yet research on them remains limited. Therefore, we want to discuss the efficiency dynamic under different service loads. Additionally, fine-grained models allow deployers to reduce the number of routed experts, both activated counts and total counts, raising the question of how this reduction affects the trade-off between MoE efficiency and performance. Our findings indicate that while deploying MoE models presents greater challenges, it also offers significant optimization opportunities. Reducing the number of activated experts can lead to substantial efficiency improvements in certain scenarios, with only minor performance degradation. Reducing the total number of experts provides limited efficiency gains but results in severe performance degradation. Our method can increase throughput by at least 10\% without any performance degradation. Overall, we conclude that MoE inference optimization remains an area with substantial potential for exploration and improvement.

Yi Zhao, Zuchao Li, Hai Zhao et al.

Task-agnostic prompt compression leverages the redundancy in natural language to reduce computational overhead and enhance information density within prompts, especially in long-context scenarios. Existing methods predominantly rely on information entropy as the metric to compress lexical units, aiming to achieve minimal information loss. However, these approaches overlook two critical aspects: (i) the importance of attention-critical tokens at the algorithmic level, and (ii) shifts in information entropy during the compression process. Motivated by these challenges, we propose a dynamic attention-aware approach for task-agnostic prompt compression (DAC). This approach effectively integrates entropy and attention information, dynamically sensing entropy shifts during compression to achieve fine-grained prompt compression. Extensive experiments across various domains, including LongBench, GSM8K, and BBH, show that DAC consistently yields robust and substantial improvements across a diverse range of tasks and LLMs, offering compelling evidence of its efficacy.

Jiliang Hu, Wenfu Wang, Zuchao Li et al.

Recent advances in large audio language models (LALMs) have greatly enhanced multimodal conversational systems. However, existing benchmarks remain limited -- they are mainly English-centric, rely on synthetic speech, and lack comprehensive, discriminative evaluation across multiple dimensions. To address these gaps, we present Voice Chat Bot Bench (VCB Bench) -- a high-quality Chinese benchmark built entirely on real human speech. VCB Bench evaluates LALMs from three complementary perspectives: instruction following (including speech-level control beyond text commands), knowledge understanding (general knowledge, reasoning, and daily dialogue), and robustness (stability under perturbations in content, environment, and speaker traits). Experiments on representative LALMs reveal notable performance gaps and highlight future directions for improvement. VCB Bench provides a reproducible and fine-grained evaluation framework, offering standardized methodology and practical insights for advancing Chinese voice conversational models.

Yi Zhao, Yajuan Peng, Cam-Tu Nguyen et al.

KV cache eviction has emerged as an effective solution to alleviate resource constraints faced by LLMs in long-context scenarios. However, existing token-level eviction methods often overlook two critical aspects: (1) their irreversible eviction strategy fails to adapt to dynamic attention patterns during decoding (the saliency shift problem), and (2) they treat both marginally important tokens and truly unimportant tokens equally, despite the collective significance of marginal tokens to model performance (the marginal information over-compression problem). To address these issues, we design two compensation mechanisms based on the high similarity of attention matrices between LLMs of different scales. We propose SmallKV, a small model assisted compensation method for KV cache compression. SmallKV can maintain attention matching between different-scale LLMs to: 1) assist the larger model in perceiving globally important information of attention; and 2) use the smaller model's attention scores to approximate those of marginal tokens in the larger model. Extensive experiments on benchmarks including GSM8K, BBH, MT-Bench, and LongBench demonstrate the effectiveness of SmallKV. Moreover, efficiency evaluations show that SmallKV achieves 1.75 - 2.56 times higher throughput than baseline methods, highlighting its potential for efficient and performant LLM inference in resource constrained environments.

Ziyang Ma, Zuchao Li, Lefei Zhang et al.

Large language models (LLMs) demonstrate strong performance across natural language processing tasks, yet undergo significant performance degradation when modified for deployment through quantization, pruning, or decoding strategy adjustments. We define this phenomenon as model hemorrhage - performance decline caused by parameter alterations and architectural changes. Through systematic analysis of various LLM frameworks, we identify key vulnerability patterns: layer expansion frequently disrupts attention mechanisms, compression techniques induce information loss cascades, and decoding adjustments amplify prediction divergences. Our investigation reveals transformer architectures exhibit inherent robustness thresholds that determine hemorrhage severity across modification types. We propose three mitigation strategies: gradient-aware pruning preserves critical weight pathways, dynamic quantization scaling maintains activation integrity, and decoding calibration aligns generation trajectories with original model distributions. This work establishes foundational metrics for evaluating model stability during adaptation, providing practical guidelines for maintaining performance while enabling efficient LLM deployment. Our findings advance understanding of neural network resilience under architectural transformations, particularly for large-scale language models.