Yucheng Li

Huiqiang Jiang, Yucheng Li, Chengruidong Zhang et al. · microsoft-research

The computational challenges of Large Language Model (LLM) inference remain a significant barrier to their widespread deployment, especially as prompt lengths continue to increase. Due to the quadratic complexity of the attention computation, it takes 30 minutes for an 8B LLM to process a prompt of 1M tokens (i.e., the pre-filling stage) on a single A100 GPU. Existing methods for speeding up prefilling often fail to maintain acceptable accuracy or efficiency when applied to long-context LLMs. To address this gap, we introduce MInference (Milliontokens Inference), a sparse calculation method designed to accelerate pre-filling of long-sequence processing. Specifically, we identify three unique patterns in long-context attention matrices-the A-shape, Vertical-Slash, and Block-Sparsethat can be leveraged for efficient sparse computation on GPUs. We determine the optimal pattern for each attention head offline and dynamically build sparse indices based on the assigned pattern during inference. With the pattern and sparse indices, we perform efficient sparse attention calculations via our optimized GPU kernels to significantly reduce the latency in the pre-filling stage of long-context LLMs. Our proposed technique can be directly applied to existing LLMs without any modifications to the pre-training setup or additional fine-tuning. By evaluating on a wide range of downstream tasks, including InfiniteBench, RULER, PG-19, and Needle In A Haystack, and models including LLaMA-3-1M, GLM4-1M, Yi-200K, Phi-3-128K, and Qwen2-128K, we demonstrate that MInference effectively reduces inference latency by up to 10x for pre-filling on an A100, while maintaining accuracy. Our code is available at https://aka.ms/MInference.

Shun Wang, Yucheng Li, Chenghua Lin et al. · meta-ai

We propose a novel RoBERTa-based model, RoPPT, which introduces a target-oriented parse tree structure in metaphor detection. Compared to existing models, RoPPT focuses on semantically relevant information and achieves the state-of-the-art on several main metaphor datasets. We also compare our approach against several popular denoising and pruning methods, demonstrating the effectiveness of our approach in context denoising. Our code and dataset can be found at https://github.com/MajiBear000/RoPPT

Yucheng Li, Shun Wang, Chenghua Lin et al. · meta-ai

In this paper, we propose FrameBERT, a RoBERTa-based model that can explicitly learn and incorporate FrameNet Embeddings for concept-level metaphor detection. FrameBERT not only achieves better or comparable performance to the state-of-the-art, but also is more explainable and interpretable compared to existing models, attributing to its ability of accounting for external knowledge of FrameNet.

Oscar Sainz, Iker García-Ferrero, Alon Jacovi et al. · ibm-research

The 1st Workshop on Data Contamination (CONDA 2024) focuses on all relevant aspects of data contamination in natural language processing, where data contamination is understood as situations where evaluation data is included in pre-training corpora used to train large scale models, compromising evaluation results. The workshop fostered a shared task to collect evidence on data contamination in current available datasets and models. The goal of the shared task and associated database is to assist the community in understanding the extent of the problem and to assist researchers in avoiding reporting evaluation results on known contaminated resources. The shared task provides a structured, centralized public database for the collection of contamination evidence, open to contributions from the community via GitHub pool requests. This first compilation paper is based on 566 reported entries over 91 contaminated sources from a total of 23 contributors. The details of the individual contamination events are available in the platform. The platform continues to be online, open to contributions from the community.

Yucheng Li, Frank Guerin, Chenghua Lin

Data contamination in model evaluation has become increasingly prevalent with the growing popularity of large language models. It allows models to "cheat" via memorisation instead of displaying true capabilities. Therefore, contamination analysis has become an crucial part of reliable model evaluation to validate results. However, existing contamination analysis is usually conducted internally by large language model developers and often lacks transparency and completeness. This paper presents an extensive data contamination report for over 15 popular large language models across six popular multiple-choice QA benchmarks. We also introduce an open-source pipeline that enables the community to perform contamination analysis on customised data and models. Our experiments reveal varying contamination levels ranging from 1\% to 45\% across benchmarks, with the contamination degree increasing rapidly over time. Performance analysis of large language models indicates that data contamination does not necessarily lead to increased model metrics: while significant accuracy boosts of up to 14\% and 7\% are observed on contaminated C-Eval and Hellaswag benchmarks, only a minimal increase is noted on contaminated MMLU. We also find larger models seem able to gain more advantages than smaller models on contaminated test sets.

Yucheng Li, Bo Dong, Chenghua Lin et al.

Large language models (LLMs) achieved remarkable performance across various tasks. However, they face challenges in managing long documents and extended conversations, due to significantly increased computational requirements, both in memory and inference time, and potential context truncation when the input exceeds the LLM's fixed context length. This paper proposes a method called Selective Context that enhances the inference efficiency of LLMs by identifying and pruning redundancy in the input context to make the input more compact. We test our approach using common data sources requiring long context processing: arXiv papers, news articles, and long conversations, on tasks of summarisation, question answering, and response generation. Experimental results show that Selective Context significantly reduces memory cost and decreases generation latency while maintaining comparable performance compared to that achieved when full context is used. Specifically, we achieve a 50\% reduction in context cost, resulting in a 36\% reduction in inference memory usage and a 32\% reduction in inference time, while observing only a minor drop of .023 in BERTscore and .038 in faithfulness on four downstream applications, indicating that our method strikes a good balance between efficiency and performance.

Yucheng Li

Large language models (LLMs) have received significant attention by achieving remarkable performance across various tasks. However, their fixed context length poses challenges when processing long documents or maintaining extended conversations. This paper proposes a method called \textit{Selective Context} that employs self-information to filter out less informative content, thereby enhancing the efficiency of the fixed context length. We demonstrate the effectiveness of our approach on tasks of summarisation and question answering across different data sources, including academic papers, news articles, and conversation transcripts.

Yucheng Li

Data contamination in model evaluation is getting increasingly prevalent as the massive training corpora of large language models often unintentionally include benchmark samples. Therefore, contamination analysis has became an inevitable part of reliable model evaluation. However, existing method of contamination analysis requires the access of the entire training data which is often confidential for recent models. This prevent the community to rigorously audit these models and conduct accurate assessment of their capability. In this paper, we propose a novel method to quantify contamination without the access of the full training set, that measure the extent of contamination with perplexity. Our analysis provides evidence of significant memorisation of recent foundation models in popular reading comprehension, summarisation benchmarks, while multiple choice appears less contaminated.

Yucheng Li, Chenghua Lin, Frank Geurin

Metaphor generation is a challenging task which can impact many downstream tasks such as improving user satisfaction with dialogue systems and story generation. This paper tackles the problem of Chinese nominal metaphor generation by introducing a multitask metaphor generation framework with self-training and metaphor identification mechanisms. Self-training addresses the data scarcity issue of metaphor datasets. That is, instead of solely relying on labelled metaphor datasets which are usually small in size, self-training helps identify potential metaphors from a large-scale unlabelled corpus for metaphor generation. The metaphor weighting mechanism enables our model to focus on the metaphor-related parts of the input (e.g., the comparison of the metaphor and comparator) during model learning and thus improves the metaphoricity of the generated metaphors. Our model is trained on an annotated corpus consisting of 6.3k sentences that contain diverse metaphorical expressions. Experimental results show that our model is able to generate metaphors with better readability and creativity compared to the baseline models, even in the situation where training data is insufficient.

Yucheng Li, Frank Guerin, Chenghua Lin

Metaphors are proven to have stronger emotional impact than literal expressions. Although this conclusion is shown to be promising in benefiting various NLP applications, the reasons behind this phenomenon are not well studied. This paper conducts the first study in exploring how metaphors convey stronger emotion than their literal counterparts. We find that metaphors are generally more specific than literal expressions. The more specific property of metaphor can be one of the reasons for metaphors' superiority in emotion expression. When we compare metaphors with literal expressions with the same specificity level, the gap of emotion expressing ability between both reduces significantly. In addition, we observe specificity is crucial in literal language as well, as literal language can express stronger emotion by making it more specific.

Yucheng Li, Frank Guerin, Chenghua Lin

Data contamination in evaluation is getting increasingly prevalent with the emergence of language models pre-trained on super large, automatically crawled corpora. This problem leads to significant challenges in the accurate assessment of model capabilities and generalisations. In this paper, we propose LatestEval, an automatic method that leverages the most recent texts to create uncontaminated reading comprehension evaluations. LatestEval avoids data contamination by only using texts published within a recent time window, ensuring no overlap with the training corpora of pre-trained language models. We develop the LatestEval automated pipeline to 1) gather the latest texts; 2) identify key information, and 3) construct questions targeting the information while removing the existing answers from the context. This encourages models to infer the answers themselves based on the remaining context, rather than just copy-paste. Our experiments demonstrate that language models exhibit negligible memorisation behaviours on LatestEval as opposed to previous benchmarks, suggesting a significantly reduced risk of data contamination and leading to a more robust evaluation. Data and code are publicly available at: https://github.com/liyucheng09/LatestEval.

Tyler Loakman, Yucheng Li, Chenghua Lin

Recently, Large Language Models (LLMs) and Vision Language Models (VLMs) have demonstrated aptitude as potential substitutes for human participants in experiments testing psycholinguistic phenomena. However, an understudied question is to what extent models that only have access to vision and text modalities are able to implicitly understand sound-based phenomena via abstract reasoning from orthography and imagery alone. To investigate this, we analyse the ability of VLMs and LLMs to demonstrate sound symbolism (i.e., to recognise a non-arbitrary link between sounds and concepts) as well as their ability to "hear" via the interplay of the language and vision modules of open and closed-source multimodal models. We perform multiple experiments, including replicating the classic Kiki-Bouba and Mil-Mal shape and magnitude symbolism tasks, and comparing human judgements of linguistic iconicity with that of LLMs. Our results show that VLMs demonstrate varying levels of agreement with human labels, and more task information may be required for VLMs versus their human counterparts for in silico experimentation. We additionally see through higher maximum agreement levels that Magnitude Symbolism is an easier pattern for VLMs to identify than Shape Symbolism, and that an understanding of linguistic iconicity is highly dependent on model size.

Yucheng Li, Huiqiang Jiang, Chengruidong Zhang et al. · microsoft-research

The integration of long-context capabilities with visual understanding unlocks unprecedented potential for Vision Language Models (VLMs). However, the quadratic attention complexity during the pre-filling phase remains a significant obstacle to real-world deployment. To overcome this limitation, we introduce MMInference (Multimodality Million tokens Inference), a dynamic sparse attention method that accelerates the prefilling stage for long-context multi-modal inputs. First, our analysis reveals that the temporal and spatial locality of video input leads to a unique sparse pattern, the Grid pattern. Simultaneously, VLMs exhibit markedly different sparse distributions across different modalities. We introduce a permutation-based method to leverage the unique Grid pattern and handle modality boundary issues. By offline search the optimal sparse patterns for each head, MMInference constructs the sparse distribution dynamically based on the input. We also provide optimized GPU kernels for efficient sparse computations. Notably, MMInference integrates seamlessly into existing VLM pipelines without any model modifications or fine-tuning. Experiments on multi-modal benchmarks-including Video QA, Captioning, VisionNIAH, and Mixed-Modality NIAH-with state-of-the-art long-context VLMs (LongVila, LlavaVideo, VideoChat-Flash, Qwen2.5-VL) show that MMInference accelerates the pre-filling stage by up to 8.3x at 1M tokens while maintaining accuracy. Our code is available at https://aka.ms/MMInference.

Shaobo Wang, Yujie Chen, Yafeng Sun et al.

Modern reasoning agents are increasingly evaluated on their ability to generate multiple valid solution paths, plans, or tool-use traces for a given input. Standard reward-maximizing RL tends to collapse onto the most easily reinforced high-reward mode, whereas distribution-matching RL aims to allocate probability mass across the entire reward-shaped solution set. Achieving this objective requires computing a prompt-dependent partition function over the trajectory space. Because existing distribution-matching methods learn this partition function online alongside the policy, calibration errors in the partition function directly distort policy updates and remain impossible to diagnose independently. We introduce DISA, short for Decoupled Importance-Sampled Anchoring, which moves this calibration problem outside the RL loop. DISA draws proposal trajectories offline, estimates the partition function via importance sampling, and freezes the resulting partition-function estimate before policy optimization begins. This decoupling preserves the distribution-matching objective while strictly separating partition-function estimation from policy learning in data, gradients, loss, and diagnostics. Empirically, on two open-weight backbones across six math and three code benchmarks, DISA matches or exceeds the online-coupled distribution-matching baseline FlowRL, outperforms rewardmaximization baselines GRPO and GSPO on math averages, and exceeds LoRASFT distillation by up to 13.8 Mean@8 points on the same offline trajectories. An LLM-as-judge evaluation further shows that DISA retains substantially more strategy-level diversity than reward-maximization baselines, and sensitivity studies on the proposal strength and inverse temperature follow the bias-variance pattern predicted by the analysis.

Siwei Wu, Yizhi Li, Xingwei Qu et al.

Large Language Models (LLMs) have achieved remarkable success in various natural language processing tasks, yet their ability to generate long-form content remains poorly understood and evaluated. Our analysis reveals that current LLMs struggle with length requirements and information density in long-text generation, with performance deteriorating as text length increases. To quantitively locate such a performance degradation and provide further insights on model development, we present LongEval, a benchmark that evaluates long-text generation through both direct and plan-based generation paradigms, inspired by cognitive and linguistic writing models. The comprehensive experiments in this work reveal interesting findings such as that while model size correlates with generation ability, the small-scale model (e.g., LongWriter), well-trained on long texts, has comparable performance. All code and datasets are released in https://github.com/Wusiwei0410/LongEval.

Yucheng Li, Surin Ahn, Huiqiang Jiang et al. · microsoft-research

Large language models (LLMs) have achieved widespread adoption across numerous applications. However, many LLMs are vulnerable to malicious attacks even after safety alignment. These attacks typically bypass LLMs' safety guardrails by wrapping the original malicious instructions inside adversarial jailbreaks prompts. Previous research has proposed methods such as adversarial training and prompt rephrasing to mitigate these safety vulnerabilities, but these methods often reduce the utility of LLMs or lead to significant computational overhead and online latency. In this paper, we propose SecurityLingua, an effective and efficient approach to defend LLMs against jailbreak attacks via security-oriented prompt compression. Specifically, we train a prompt compressor designed to discern the "true intention" of the input prompt, with a particular focus on detecting the malicious intentions of adversarial prompts. Then, in addition to the original prompt, the intention is passed via the system prompt to the target LLM to help it identify the true intention of the request. SecurityLingua ensures a consistent user experience by leaving the original input prompt intact while revealing the user's potentially malicious intention and stimulating the built-in safety guardrails of the LLM. Moreover, thanks to prompt compression, SecurityLingua incurs only a negligible overhead and extra token cost compared to all existing defense methods, making it an especially practical solution for LLM defense. Experimental results demonstrate that SecurityLingua can effectively defend against malicious attacks and maintain utility of the LLM with negligible compute and latency overhead. Our code is available at https://aka.ms/SecurityLingua.

Wenxuan Li, Chengruidong Zhang, Huiqiang Jiang et al. · microsoft-research

The adoption of long context windows has become a standard feature in Large Language Models (LLMs), as extended contexts significantly enhance their capacity for complex reasoning and broaden their applicability across diverse scenarios. Dynamic sparse attention is a promising approach for reducing the computational cost of long-context. However, efficiently training LLMs with dynamic sparse attention on ultra-long contexts-especially in distributed settings-remains a significant challenge, due in large part to worker- and step-level imbalance. This paper introduces MTraining, a novel distributed methodology leveraging dynamic sparse attention to enable efficient training for LLMs with ultra-long contexts. Specifically, MTraining integrates three key components: a dynamic sparse training pattern, balanced sparse ring attention, and hierarchical sparse ring attention. These components are designed to synergistically address the computational imbalance and communication overheads inherent in dynamic sparse attention mechanisms during the training of models with extensive context lengths. We demonstrate the efficacy of MTraining by training Qwen2.5-3B, successfully expanding its context window from 32K to 512K tokens on a cluster of 32 A100 GPUs. Our evaluations on a comprehensive suite of downstream tasks, including RULER, PG-19, InfiniteBench, and Needle In A Haystack, reveal that MTraining achieves up to a 6x higher training throughput while preserving model accuracy. Our code is available at https://github.com/microsoft/MInference/tree/main/MTraining.

Yucheng Li, Xiaofan Wang, Junyi Wang et al.

Whole-brain parcellation from MRI is a critical yet challenging task due to the complexity of subdividing the brain into numerous small, irregular shaped regions. Traditionally, template-registration methods were used, but recent advances have shifted to deep learning for faster workflows. While large models like the Segment Anything Model (SAM) offer transferable feature representations, they are not tailored for the high precision required in brain parcellation. To address this, we propose BrainSegNet, a novel framework that adapts SAM for accurate whole-brain parcellation into 95 regions. We enhance SAM by integrating U-Net skip connections and specialized modules into its encoder and decoder, enabling fine-grained anatomical precision. Key components include a hybrid encoder combining U-Net skip connections with SAM's transformer blocks, a multi-scale attention decoder with pyramid pooling for varying-sized structures, and a boundary refinement module to sharpen edges. Experimental results on the Human Connectome Project (HCP) dataset demonstrate that BrainSegNet outperforms several state-of-the-art methods, achieving higher accuracy and robustness in complex, multi-label parcellation.

Yucheng Li, Huiqiang Jiang, Qianhui Wu et al. · microsoft-research

Long-context LLMs have enabled numerous downstream applications but also introduced significant challenges related to computational and memory efficiency. To address these challenges, optimizations for long-context inference have been developed, centered around the KV cache. However, existing benchmarks often evaluate in single-request, neglecting the full lifecycle of the KV cache in real-world use. This oversight is particularly critical, as KV cache reuse has become widely adopted in LLMs inference frameworks, such as vLLM and SGLang, as well as by LLM providers, including OpenAI, Microsoft, Google, and Anthropic. To address this gap, we introduce SCBench(SharedContextBench), a comprehensive benchmark for evaluating long-context methods from a KV cachecentric perspective: 1) KV cache generation, 2) KV cache compression, 3) KV cache retrieval, 4) KV cache loading. Specifically, SCBench uses test examples with shared context, ranging 12 tasks with two shared context modes, covering four categories of long-context capabilities: string retrieval, semantic retrieval, global information, and multi-task. With it, we provide an extensive KV cache-centric analysis of eight categories long-context solutions, including Gated Linear RNNs, Mamba-Attention hybrids, and efficient methods such as sparse attention, KV cache dropping, quantization, retrieval, loading, and prompt compression. The evaluation is conducted on 8 long-context LLMs. Our findings show that sub-O(n) memory methods suffer in multi-turn scenarios, while sparse encoding with O(n) memory and sub-O(n^2) pre-filling computation perform robustly. Dynamic sparsity yields more expressive KV caches than static patterns, and layer-level sparsity in hybrid architectures reduces memory usage with strong performance. Additionally, we identify attention distribution shift issues in long-generation scenarios. https://aka.ms/SCBench.

Siyu Teng, Xuan Li, Yucheng Li et al.

In recent years, open-pit mining has seen significant advancement, the cooperative operation of various specialized machinery substantially enhancing the efficiency of mineral extraction. However, the harsh environment and complex conditions in open-pit mines present substantial challenges for the implementation of autonomous transportation systems. This research introduces a novel paradigm that integrates Scenario Engineering (SE) with autonomous transportation systems to significantly improve the trustworthiness, robustness, and efficiency in open-pit mines by incorporating the four key components of SE, including Scenario Feature Extractor, Intelligence and Index (I&I), Calibration and Certification (C&C), and Verification and Validation (V&V). This paradigm has been validated in two famous open-pit mines, the experiment results demonstrate marked improvements in robustness, trustworthiness, and efficiency. By enhancing the capacity, scalability, and diversity of autonomous transportation, this paradigm fosters the integration of SE and parallel driving and finally propels the achievement of the '6S' objectives.

Yucheng Li, Yunhao Guo, Frank Guerin et al.

Existing methods for evaluating large language models face challenges such as data contamination, sensitivity to prompts, and the high cost of benchmark creation. To address this, we propose a lossless data compression based evaluation approach that tests how models' predictive abilities generalize after their training cutoff. Specifically, we collect comprehensive test data spanning 83 months from 2017 to 2023 and split the data into training and testing periods according to models' training data cutoff. We measure: 1) the compression performance on the testing period as a measure of generalization on unseen data; and 2) the performance gap between the training and testing period as a measure of robustness. Our experiments test 14 representative large language models with various sizes on sources including Wikipedia, news articles, code, arXiv papers, and multi-modal data. We find that the compression rate of many models reduces significantly after their cutoff date, but models such as Mistral and Llama-2 demonstrate a good balance between performance and robustness. Results also suggest that models struggle to generalize on news and code data, but work especially well on arXiv papers. We also find the context size and tokenization implementation have a big impact of on the overall compression performance.

Zefan Cai, Wen Xiao, Hanshi Sun et al. · microsoft-research

Reasoning models have demonstrated impressive performance in self-reflection and chain-of-thought reasoning. However, they often produce excessively long outputs, leading to prohibitively large key-value (KV) caches during inference. While chain-of-thought inference significantly improves performance on complex reasoning tasks, it can also lead to reasoning failures when deployed with existing KV cache compression approaches. To address this, we propose Redundancy-aware KV Cache Compression for Reasoning models (R-KV), a novel method specifically targeting redundant tokens in reasoning models. Our method preserves nearly 100% of the full KV cache performance using only 10% of the KV cache, substantially outperforming existing KV cache baselines, which reach only 60% of the performance. Remarkably, R-KV even achieves 105% of full KV cache performance with 16% of the KV cache. This KV-cache reduction also leads to a 90% memory saving and a 6.6X throughput over standard chain-of-thought reasoning inference. Experimental results show that R-KV consistently outperforms existing KV cache compression baselines across two mathematical reasoning datasets.

Yucheng Li, Frank Guerin, Chenghua Lin

Metaphors are considered to pose challenges for a wide spectrum of NLP tasks. This gives rise to the area of computational metaphor processing. However, it remains unclear what types of metaphors challenge current state-of-the-art models. In this paper, we test various NLP models on the VUA metaphor dataset and quantify to what extent metaphors affect models' performance on various downstream tasks. Analysis reveals that VUA includes a large number of metaphors that pose little difficulty to downstream tasks. We would like to shift the attention of researchers away from these metaphors to instead focus on challenging metaphors. To identify hard metaphors, we propose an automatic pipeline that identifies metaphors that challenge a particular model. Our analysis demonstrates that our detected hard metaphors contrast significantly with VUA and reduce the accuracy of machine translation by 16\%, QA performance by 4\%, NLI by 7\%, and metaphor identification recall by over 14\% for various popular NLP systems.

Joseph James, Chenghao Xiao, Yucheng Li et al.

Scientific rigour tends to be sidelined in favour of bold statements, leading authors to overstate claims beyond what their results support. We present RIGOURATE, a two-stage multimodal framework that retrieves supporting evidence from a paper's body and assigns each claim an overstatement score. The framework consists of a dataset of over 10K claim-evidence sets from ICLR and NeurIPS papers, annotated using eight LLMs, with overstatement scores calibrated using peer-review comments and validated through human evaluation. It employes a fine-tuned reranker for evidence retrieval and a fine-tuned model to predict overstatement scores with justification. Compared to strong baselines, RIGOURATE enables improved evidence retrieval and overstatement detection. Overall, our work operationalises evidential proportionality and supports clearer, more transparent scientific communication.

Yiwei Liu, Yucheng Li, Xiao Li et al.

Joint logical-numerical reasoning remains a major challenge for language models, yet existing datasets rely on fixed rule sets and offer limited control over task complexity, constraining their generalizability for evaluation and training. We present LogiNumSynth, a flexible natural language problem synthesizer that synthesizes tasks requiring proficiency in joint logical reasoning (e.g., rule-based reasoning) and numerical reasoning (e.g., arithmetic computation). LogiNumSynth supports fine-grained control over reasoning world richness, logical reasoning depth, and the complexity of numerical computations, enabling flexible data synthesis across difficulty levels. We demonstrate three key contributions: (1) Synthesizer -- synthesizing fully controllable joint reasoning tasks over natural language; (2) Evaluation & Process Analysis -- evaluating both process accuracy and answer accuracy; (3) Targeted Training -- using synthesized data to enhance LLMs' reasoning performance. Experiments with multiple LLMs highlight persistent weaknesses in logical-numerical reasoning, showing that LogiNumSynth can serve as both a diagnostic tool and a source of targeted supervision for advancing integrated reasoning skills.

Zefan Cai, Yichi Zhang, Bofei Gao et al.

In this study, we investigate whether attention-based information flow inside large language models (LLMs) is aggregated through noticeable patterns for long context processing. Our observations reveal that LLMs aggregate information through Pyramidal Information Funneling where attention is scattering widely in lower layers, progressively consolidating within specific contexts, and ultimately focusing on critical tokens (a.k.a massive activation or attention sink) in higher layers. Motivated by these insights, we developed PyramidKV, a novel and effective KV cache compression method. This approach dynamically adjusts the KV cache size across different layers, allocating more cache in lower layers and less in higher ones, diverging from traditional methods that maintain a uniform KV cache size. Our experimental evaluations, utilizing the LongBench benchmark, show that PyramidKV matches the performance of models with a full KV cache while retaining only 12% of the KV cache, thus significantly reducing memory usage. In scenarios emphasizing memory efficiency, where only 0.7% of the KV cache is maintained, PyramidKV surpasses other KV cache compression techniques, achieving up to a 20.5 absolute accuracy improvement on TREC dataset. In the Needle-in-a-Haystack experiment, PyramidKV outperforms competing methods in maintaining long-context comprehension in LLMs; notably, retaining just 128 KV cache entries enables the LLAMA-3-70B model to achieve 100.0 Acc. performance.

Yucheng Li, Shun Wang, Chenghua Lin et al.

One noticeable trend in metaphor detection is the embrace of linguistic theories such as the metaphor identification procedure (MIP) for model architecture design. While MIP clearly defines that the metaphoricity of a lexical unit is determined based on the contrast between its \textit{contextual meaning} and its \textit{basic meaning}, existing work does not strictly follow this principle, typically using the \textit{aggregated meaning} to approximate the basic meaning of target words. In this paper, we propose a novel metaphor detection method, which models the basic meaning of the word based on literal annotation from the training set, and then compares this with the contextual meaning in a target sentence to identify metaphors. Empirical results show that our method outperforms the state-of-the-art method significantly by 1.0\% in F1 score. Moreover, our performance even reaches the theoretical upper bound on the VUA18 benchmark for targets with basic annotations, which demonstrates the importance of modelling basic meanings for metaphor detection.

Wenying Ma, Liangliang Cao, Lei Yu et al.

Fisher vector has been widely used in many multimedia retrieval and visual recognition applications with good performance. However, the computation complexity prevents its usage in real-time video monitoring. In this work, we proposed and implemented GPU-FV, a fast Fisher vector extraction method with the help of modern GPUs. The challenge of implementing Fisher vector on GPUs lies in the data dependency in feature extraction and expensive memory access in Fisher vector computing. To handle these challenges, we carefully designed GPU-FV in a way that utilizes the computing power of GPU as much as possible, and applied optimizations such as loop tiling to boost the performance. GPU-FV is about 12 times faster than the CPU version, and 50\% faster than a non-optimized GPU implementation. For standard video input (320*240), GPU-FV can process each frame within 34ms on a model GPU. Our experiments show that GPU-FV obtains a similar recognition accuracy as traditional FV on VOC 2007 and Caltech 256 image sets. We also applied GPU-FV for realtime video monitoring tasks and found that GPU-FV outperforms a number of previous works. Especially, when the number of training examples are small, GPU-FV outperforms the recent popular deep CNN features borrowed from ImageNet. The code can be downloaded from the following link https://bitbucket.org/mawenjing/gpu-fv.