Lili Qiu

Huiqiang Jiang, Qianhui Wu, Xufang Luo et al. · microsoft-research

In long context scenarios, large language models (LLMs) face three main challenges: higher computational cost, performance reduction, and position bias. Research indicates that LLM performance hinges on the density and position of key information in the input prompt. Inspired by these findings, we propose LongLLMLingua for prompt compression towards improving LLMs' perception of the key information to simultaneously address the three challenges. Our extensive evaluation across various long context scenarios demonstrates that LongLLMLingua not only enhances performance but also significantly reduces costs and latency. For instance, in the NaturalQuestions benchmark, LongLLMLingua boosts performance by up to 21.4% with around 4x fewer tokens in GPT-3.5-Turbo, leading to substantial cost savings. It achieves a 94.0% cost reduction in the LooGLE benchmark. Moreover, when compressing prompts of about 10k tokens at ratios of 2x-6x, LongLLMLingua can accelerate end-to-end latency by 1.4x-2.6x. Our code is available at https://aka.ms/LongLLMLingua.

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.

Huiqiang Jiang, Qianhui Wu, Chin-Yew Lin et al. · microsoft-research

Large language models (LLMs) have been applied in various applications due to their astonishing capabilities. With advancements in technologies such as chain-of-thought (CoT) prompting and in-context learning (ICL), the prompts fed to LLMs are becoming increasingly lengthy, even exceeding tens of thousands of tokens. To accelerate model inference and reduce cost, this paper presents LLMLingua, a coarse-to-fine prompt compression method that involves a budget controller to maintain semantic integrity under high compression ratios, a token-level iterative compression algorithm to better model the interdependence between compressed contents, and an instruction tuning based method for distribution alignment between language models. We conduct experiments and analysis over four datasets from different scenarios, i.e., GSM8K, BBH, ShareGPT, and Arxiv-March23; showing that the proposed approach yields state-of-the-art performance and allows for up to 20x compression with little performance loss. Our code is available at https://aka.ms/LLMLingua.

Di Liu, Meng Chen, Baotong Lu et al. · microsoft-research

Transformer-based Large Language Models (LLMs) have become increasingly important. However, due to the quadratic time complexity of attention computation, scaling LLMs to longer contexts incurs extremely slow inference speed and high GPU memory consumption for caching key-value (KV) vectors. This paper proposes RetrievalAttention, a training-free approach to both accelerate attention computation and reduce GPU memory consumption. By leveraging the dynamic sparsity of attention mechanism, RetrievalAttention proposes to build approximate nearest neighbor search (ANNS) indexes for KV vectors in CPU memory and retrieve the most relevant ones through vector search during generation. Unfortunately, we observe that the off-the-shelf ANNS indexes are often ineffective for such retrieval tasks due to the out-of-distribution (OOD) between query vectors and key vectors in the attention mechanism. RetrievalAttention addresses the OOD challenge by designing an attention-aware vector search algorithm that can adapt to the distribution of query vectors. Our evaluation demonstrates that RetrievalAttention achieves near full attention accuracy while only requiring access to 1--3% of the data. This leads to a significant reduction in the inference cost of long-context LLMs, with a much lower GPU memory footprint. In particular, RetrievalAttention only needs a single NVIDIA RTX4090 (24GB) to serve 128K tokens for LLMs with 8B parameters, which is capable of generating one token in 0.188 seconds.

Yifan Yang, Weiquan Huang, Yixuan Wei et al. · microsoft-research

Image token removal is an efficient augmentation strategy for reducing the cost of computing image features. However, this efficient augmentation strategy has been found to adversely affect the accuracy of CLIP-based training. We hypothesize that removing a large portion of image tokens may improperly discard the semantic content associated with a given text description, thus constituting an incorrect pairing target in CLIP training. To address this issue, we propose an attentive token removal approach for CLIP training, which retains tokens with a high semantic correlation to the text description. The correlation scores are computed in an online fashion using the EMA version of the visual encoder. Our experiments show that the proposed attentive masking approach performs better than the previous method of random token removal for CLIP training. The approach also makes it efficient to apply multiple augmentation views to the image, as well as introducing instance contrastive learning tasks between these views into the CLIP framework. Compared to other CLIP improvements that combine different pre-training targets such as SLIP and MaskCLIP, our method is not only more effective, but also much more efficient. Specifically, using ViT-B and YFCC-15M dataset, our approach achieves $43.9\%$ top-1 accuracy on ImageNet-1K zero-shot classification, as well as $62.7/42.1$ and $38.0/23.2$ I2T/T2I retrieval accuracy on Flickr30K and MS COCO, which are $+1.1\%$, $+5.5/+0.9$, and $+4.4/+1.3$ higher than the SLIP method, while being $2.30\times$ faster. An efficient version of our approach running $1.16\times$ faster than the plain CLIP model achieves significant gains of $+5.3\%$, $+11.3/+8.0$, and $+9.5/+4.9$ on these benchmarks.

Ningxin Zheng, Huiqiang Jiang, Quanlu Zhang et al. · microsoft-research

Dynamic sparsity, where the sparsity patterns are unknown until runtime, poses a significant challenge to deep learning. The state-of-the-art sparsity-aware deep learning solutions are restricted to pre-defined, static sparsity patterns due to significant overheads associated with preprocessing. Efficient execution of dynamic sparse computation often faces the misalignment between the GPU-friendly tile configuration for efficient execution and the sparsity-aware tile shape that minimizes coverage wastes (non-zero values in tensor). In this paper, we propose PIT, a deep-learning compiler for dynamic sparsity. PIT proposes a novel tiling mechanism that leverages Permutation Invariant Transformation (PIT), a mathematically proven property, to transform multiple sparsely located micro-tiles into a GPU-efficient dense tile without changing the computation results, thus achieving both high GPU utilization and low coverage waste. Given a model, PIT first finds feasible PIT rules for all its operators and generates efficient GPU kernels accordingly. At runtime, with the novel SRead and SWrite primitives, PIT rules can be executed extremely fast to support dynamic sparsity in an online manner. Extensive evaluation on diverse models shows that PIT can accelerate dynamic sparsity computation by up to 5.9x (average 2.43x) over state-of-the-art compilers.

Yukang Liang, Kaitao Song, Shaoguang Mao et al. · microsoft-research

Pronunciation assessment is a major challenge in the computer-aided pronunciation training system, especially at the word (phoneme)-level. To obtain word (phoneme)-level scores, current methods usually rely on aligning components to obtain acoustic features of each word (phoneme), which limits the performance of assessment to the accuracy of alignments. Therefore, to address this problem, we propose a simple yet effective method, namely \underline{M}asked pre-training for \underline{P}ronunciation \underline{A}ssessment (MPA). Specifically, by incorporating a mask-predict strategy, our MPA supports end-to-end training without leveraging any aligning components and can solve misalignment issues to a large extent during prediction. Furthermore, we design two evaluation strategies to enable our model to conduct assessments in both unsupervised and supervised settings. Experimental results on SpeechOcean762 dataset demonstrate that MPA could achieve better performance than previous methods, without any explicit alignment. In spite of this, MPA still has some limitations, such as requiring more inference time and reference text. They expect to be addressed in future work.

Ziyue Li, Yuchen Fang, You Li et al. · cmu, tsinghua

A timely detection of seizures for newborn infants with electroencephalogram (EEG) has been a common yet life-saving practice in the Neonatal Intensive Care Unit (NICU). However, it requires great human efforts for real-time monitoring, which calls for automated solutions to neonatal seizure detection. Moreover, the current automated methods focusing on adult epilepsy monitoring often fail due to (i) dynamic seizure onset location in human brains; (ii) different montages on neonates and (iii) huge distribution shift among different subjects. In this paper, we propose a deep learning framework, namely STATENet, to address the exclusive challenges with exquisite designs at the temporal, spatial and model levels. The experiments over the real-world large-scale neonatal EEG dataset illustrate that our framework achieves significantly better seizure detection performance.

Yasheng Sun, Yifan Yang, Houwen Peng et al.

While language-guided image manipulation has made remarkable progress, the challenge of how to instruct the manipulation process faithfully reflecting human intentions persists. An accurate and comprehensive description of a manipulation task using natural language is laborious and sometimes even impossible, primarily due to the inherent uncertainty and ambiguity present in linguistic expressions. Is it feasible to accomplish image manipulation without resorting to external cross-modal language information? If this possibility exists, the inherent modality gap would be effortlessly eliminated. In this paper, we propose a novel manipulation methodology, dubbed ImageBrush, that learns visual instructions for more accurate image editing. Our key idea is to employ a pair of transformation images as visual instructions, which not only precisely captures human intention but also facilitates accessibility in real-world scenarios. Capturing visual instructions is particularly challenging because it involves extracting the underlying intentions solely from visual demonstrations and then applying this operation to a new image. To address this challenge, we formulate visual instruction learning as a diffusion-based inpainting problem, where the contextual information is fully exploited through an iterative process of generation. A visual prompting encoder is carefully devised to enhance the model's capacity in uncovering human intent behind the visual instructions. Extensive experiments show that our method generates engaging manipulation results conforming to the transformations entailed in demonstrations. Moreover, our model exhibits robust generalization capabilities on various downstream tasks such as pose transfer, image translation and video inpainting.

Jian He, Chenxi Yang, Zhaoyuan He et al.

Deep neural networks (DNNs) have been widely used in various video analytic tasks. These tasks demand real-time responses. Due to the limited processing power on mobile devices, a common way to support such real-time analytics is to offload the processing to an edge server. This paper examines how to speed up the edge server DNN processing for multiple clients. In particular, we observe batching multiple DNN requests significantly speeds up the processing time. Based on this observation, we first design a novel scheduling algorithm to exploit the batching benefits of all requests that run the same DNN. This is compelling since there are only a handful of DNNs and many requests tend to use the same DNN. Our algorithms are general and can support different objectives, such as minimizing the completion time or maximizing the on-time ratio. We then extend our algorithm to handle requests that use different DNNs with or without shared layers. Finally, we develop a collaborative approach to further improve performance by adaptively processing some of the requests or portions of the requests locally at the clients. This is especially useful when the network and/or server is congested. Our implementation shows the effectiveness of our approach under different request distributions (e.g., Poisson, Pareto, and Constant inter-arrivals).

Guanghao Yin, Zefan Qu, Xinyang Jiang et al.

Online video streaming has fundamental limitations on the transmission bandwidth and computational capacity and super-resolution is a promising potential solution. However, applying existing video super-resolution methods to online streaming is non-trivial. Existing video codecs and streaming protocols (\eg, WebRTC) dynamically change the video quality both spatially and temporally, which leads to diverse and dynamic degradations. Furthermore, online streaming has a strict requirement for latency that most existing methods are less applicable. As a result, this paper focuses on the rarely exploited problem setting of online streaming video super resolution. To facilitate the research on this problem, a new benchmark dataset named LDV-WebRTC is constructed based on a real-world online streaming system. Leveraging the new benchmark dataset, we proposed a novel method specifically for online video streaming, which contains a convolution and Look-Up Table (LUT) hybrid model to achieve better performance-latency trade-off. To tackle the changing degradations, we propose a mixture-of-expert-LUT module, where a set of LUT specialized in different degradations are built and adaptively combined to handle different degradations. Experiments show our method achieves 720P video SR around 100 FPS, while significantly outperforms existing LUT-based methods and offers competitive performance compared to efficient CNN-based methods.

Jiaqi Gao, Xinyang Jiang, Yuqing Yang et al.

Monitoring and analyzing stereotypical behaviours is important for early intervention and care taking in Autism Spectrum Disorder (ASD). This paper focuses on automatically detecting stereotypical behaviours with computer vision techniques. Off-the-shelf methods tackle this task by supervised classification and activity recognition techniques. However, the unbounded types of stereotypical behaviours and the difficulty in collecting video recordings of ASD patients largely limit the feasibility of the existing supervised detection methods. As a result, we tackle these challenges from a new perspective, i.e. unsupervised video anomaly detection for stereotypical behaviours detection. The models can be trained among unlabeled videos containing only normal behaviours and unknown types of abnormal behaviours can be detected during inference. Correspondingly, we propose a Dual Stream deep model for Stereotypical Behaviours Detection, DS-SBD, based on the temporal trajectory of human poses and the repetition patterns of human actions. Extensive experiments are conducted to verify the effectiveness of our proposed method and suggest that it serves as a potential benchmark for future research.

Xiaoxuan He, Yifan Yang, Xinyang Jiang et al.

Vision-Language Pre-training (VLP) has shown the merits of analysing medical images, by leveraging the semantic congruence between medical images and their corresponding reports. It efficiently learns visual representations, which in turn facilitates enhanced analysis and interpretation of intricate imaging data. However, such observation is predominantly justified on single-modality data (mostly 2D images like X-rays), adapting VLP to learning unified representations for medical images in real scenario remains an open challenge. This arises from medical images often encompass a variety of modalities, especially modalities with different various number of dimensions (e.g., 3D images like Computed Tomography). To overcome the aforementioned challenges, we propose an Unified Medical Image Pre-training framework, namely UniMedI, which utilizes diagnostic reports as common semantic space to create unified representations for diverse modalities of medical images (especially for 2D and 3D images). Under the text's guidance, we effectively uncover visual modality information, identifying the affected areas in 2D X-rays and slices containing lesion in sophisticated 3D CT scans, ultimately enhancing the consistency across various medical imaging modalities. To demonstrate the effectiveness and versatility of UniMedI, we evaluate its performance on both 2D and 3D images across 10 different datasets, covering a wide range of medical image tasks such as classification, segmentation, and retrieval. UniMedI has demonstrated superior performance in downstream tasks, showcasing its effectiveness in establishing a universal medical visual representation.

Fucheng Jia, Shiqi Jiang, Ting Cao et al.

Web is increasingly becoming the primary platform to deliver AI services onto edge devices, making in-browser deep learning (DL) inference more prominent. Nevertheless, the heterogeneity of edge devices, combined with the underdeveloped state of Web hardware acceleration practices, hinders current in-browser inference from achieving its full performance potential on target devices. To address this issue, this paper presents the pioneering inbrowser inference system, nnJIT, which enables just-in-time (JIT) auto-generation of optimized computing kernels for edge devices. nnJIT is built upon two novel techniques that significantly reduce kernel search and compilation overhead while improving performance firmly: Tensor-Web Compiling Co-Design lowers compiling costs by around 100X through eliminating redundant and ineffective compiling passes; Web-Specific Lite Kernel Optimization Space reduces kernel tuning costs by focusing on Web programming requirements and efficient device resource utilization, pruning the optimization space from millions to only dozens. nnJIT is evaluated for modern models, e.g., BART, T5, and Llama 2, on a range of edge devices including laptops and smartphones using different browsers and hardware from ARM, Intel, AMD and Nvidia. The results show that nnJIT can achieve up to 8.2X faster within 30 seconds compared to the existing baselines.

Zhaoyuan He, Yifan Yang, Lili Qiu et al.

As mobile devices become increasingly popular for video streaming, it's crucial to optimize the streaming experience for these devices. Although deep learning-based video enhancement techniques are gaining attention, most of them cannot support real-time enhancement on mobile devices. Additionally, many of these techniques are focused solely on super-resolution and cannot handle partial or complete loss or corruption of video frames, which is common on the Internet and wireless networks. To overcome these challenges, we present a novel approach in this paper. Our approach consists of (i) a novel video frame recovery scheme, (ii) a new super-resolution algorithm, and (iii) a receiver enhancement-aware video bit rate adaptation algorithm. We have implemented our approach on an iPhone 12, and it can support 30 frames per second (FPS). We have evaluated our approach in various networks such as WiFi, 3G, 4G, and 5G networks. Our evaluation shows that our approach enables real-time enhancement and results in a significant increase in video QoE (Quality of Experience) of 24\% - 82\% in our video streaming system.

Zhaoyuan He, Yifan Yang, Shuozhe Li et al.

Cloud gaming is a multi-billion dollar industry. A client in cloud gaming sends its movement to the game server on the Internet, which renders and transmits the resulting video back. In order to provide a good gaming experience, a latency below 80 ms is required. This means that video rendering, encoding, transmission, decoding, and display have to finish within that time frame, which is especially challenging to achieve due to server overload, network congestion, and losses. In this paper, we propose a new method for recovering lost or corrupted video frames in cloud gaming. Unlike traditional video frame recovery, our approach uses game states to significantly enhance recovery accuracy and utilizes partially decoded frames to recover lost portions. We develop a holistic system that consists of (i) efficiently extracting game states, (ii) modifying H.264 video decoder to generate a mask to indicate which portions of video frames need recovery, and (iii) designing a novel neural network to recover either complete or partial video frames. Our approach is extensively evaluated using iPhone 12 and laptop implementations, and we demonstrate the utility of game states in the game video recovery and the effectiveness of our overall design.

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.

Zilong Wang, Xufang Luo, Xinyang Jiang et al.

Evaluating generated radiology reports is crucial for the development of radiology AI, but existing metrics fail to reflect the task's clinical requirements. This study proposes a novel evaluation framework using large language models (LLMs) to compare radiology reports for assessment. We compare the performance of various LLMs and demonstrate that, when using GPT-4, our proposed metric achieves evaluation consistency close to that of radiologists. Furthermore, to reduce costs and improve accessibility, making this method practical, we construct a dataset using LLM evaluation results and perform knowledge distillation to train a smaller model. The distilled model achieves evaluation capabilities comparable to GPT-4. Our framework and distilled model offer an accessible and efficient evaluation method for radiology report generation, facilitating the development of more clinically relevant models. The model will be further open-sourced and accessible.

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.

Yike Zhang, Zhiyuan He, Huiqiang Jiang et al. · microsoft-research

Large language models (LLMs) enable long-context tasks but face efficiency challenges due to the growing key-value (KV) cache. We propose LeanK, a learning-based method that prunes unimportant key (K) cache channels by leveraging static channel sparsity. With a novel two-stage training process, LeanK learns channel-wise static mask that could satisfy specific sparsity ratio and hardware alignment requirement. LeanK reduces GPU memory and accelerates decoding without sacrificing accuracy. Experiments demonstrate up to 70% K cache and 16%-18% V cache memory reduction. Custom decoding kernel enables 1.3x speedup for attention computation. We also provide insights into model channels and attention heads during long-context inference by analyzing the learned importance distribution. Our code is available at https://aka.ms/LeanK.

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.

Yijiong Yu, Huiqiang Jiang, Xufang Luo et al.

Large Language Models (LLMs) are increasingly applied in various real-world scenarios due to their excellent generalization capabilities and robust generative abilities. However, they exhibit position bias, also known as "lost in the middle", a phenomenon that is especially pronounced in long-context scenarios, which indicates the placement of the key information in different positions of a prompt can significantly affect accuracy. This paper first explores the micro-level manifestations of position bias, concluding that attention weights are a micro-level expression of position bias. It further identifies that, in addition to position embeddings, causal attention mask also contributes to position bias by creating position-specific hidden states. Based on these insights, we propose a method to mitigate position bias by scaling this positional hidden states. Experiments on the NaturalQuestions Multi-document QA, KV retrieval, LongBench and timeline reorder tasks, using various models including RoPE models, context windowextended models, and Alibi models, demonstrate the effectiveness and generalizability of our approach. Our method can improve performance by up to 15.2% by modifying just one dimension of hidden states. Our code is available at https://aka.ms/PositionalHidden.

Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang et al.

This paper focuses on task-agnostic prompt compression for better generalizability and efficiency. Considering the redundancy in natural language, existing approaches compress prompts by removing tokens or lexical units according to their information entropy obtained from a causal language model such as LLaMa-7B. The challenge is that information entropy may be a suboptimal compression metric: (i) it only leverages unidirectional context and may fail to capture all essential information needed for prompt compression; (ii) it is not aligned with the prompt compression objective. To address these issues, we propose a data distillation procedure to derive knowledge from an LLM to compress prompts without losing crucial information, and meantime, introduce an extractive text compression dataset. We formulate prompt compression as a token classification problem to guarantee the faithfulness of the compressed prompt to the original one, and use a Transformer encoder as the base architecture to capture all essential information for prompt compression from the full bidirectional context. Our approach leads to lower latency by explicitly learning the compression objective with smaller models such as XLM-RoBERTa-large and mBERT. We evaluate our method on both in-domain and out-of-domain datasets, including MeetingBank, LongBench, ZeroScrolls, GSM8K, and BBH. Despite its small size, our model shows significant performance gains over strong baselines and demonstrates robust generalization ability across different LLMs. Additionally, our model is 3x-6x faster than existing prompt compression methods, while accelerating the end-to-end latency by 1.6x-2.9x with compression ratios of 2x-5x. Our code is available at https://aka.ms/LLMLingua-2.

Yuqi Chen, Kan Ren, Kaitao Song et al. · cmu, tsinghua

Self-supervised learning has emerged as a highly effective approach in the fields of natural language processing and computer vision. It is also applicable to brain signals such as electroencephalography (EEG) data, given the abundance of available unlabeled data that exist in a wide spectrum of real-world medical applications ranging from seizure detection to wave analysis. The existing works leveraging self-supervised learning on EEG modeling mainly focus on pretraining upon each individual dataset corresponding to a single downstream task, which cannot leverage the power of abundant data, and they may derive sub-optimal solutions with a lack of generalization. Moreover, these methods rely on end-to-end model learning which is not easy for humans to understand. In this paper, we present a novel EEG foundation model, namely EEGFormer, pretrained on large-scale compound EEG data. The pretrained model cannot only learn universal representations on EEG signals with adaptable performance on various downstream tasks but also provide interpretable outcomes of the useful patterns within the data. To validate the effectiveness of our model, we extensively evaluate it on various downstream tasks and assess the performance under different transfer settings. Furthermore, we demonstrate how the learned model exhibits transferable anomaly detection performance and provides valuable interpretability of the acquired patterns via self-supervised learning.

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.

Weiquan Huang, Aoqi Wu, Yifan Yang et al.

CLIP is a foundational multimodal model that aligns image and text features into a shared representation space via contrastive learning on large-scale image-text pairs. Its effectiveness primarily stems from the use of natural language as rich supervision. Motivated by the remarkable advancements in large language models (LLMs), this work explores how LLMs' superior text understanding and extensive open-world knowledge can enhance CLIP's capability, especially for processing longer and more complex image captions. We propose an efficient post-training strategy that integrates LLMs into pretrained CLIP. To address the challenge posed by the autoregressive nature of LLMs, we introduce a caption-to-caption contrastive fine-tuning framework, significantly enhancing the discriminative quality of LLM outputs. Extensive experiments demonstrate that our approach outperforms LoRA-based methods, achieving nearly fourfold faster training with superior performance. Furthermore, we validate substantial improvements over state-of-the-art models such as CLIP, EVA02, and SigLip2 across various zero-shot multimodal retrieval tasks, cross-lingual retrieval tasks, and multimodal language model pretraining.

Zhuoshi Pan, Qianhui Wu, Huiqiang Jiang et al. · microsoft-research

To deliver coherent and personalized experiences in long-term conversations, existing approaches typically perform retrieval augmented response generation by constructing memory banks from conversation history at either the turn-level, session-level, or through summarization techniques.In this paper, we present two key findings: (1) The granularity of memory unit matters: turn-level, session-level, and summarization-based methods each exhibit limitations in both memory retrieval accuracy and the semantic quality of the retrieved content. (2) Prompt compression methods, such as LLMLingua-2, can effectively serve as a denoising mechanism, enhancing memory retrieval accuracy across different granularities. Building on these insights, we propose SeCom, a method that constructs the memory bank at segment level by introducing a conversation segmentation model that partitions long-term conversations into topically coherent segments, while applying compression based denoising on memory units to enhance memory retrieval. Experimental results show that SeCom exhibits a significant performance advantage over baselines on long-term conversation benchmarks LOCOMO and Long-MT-Bench+. Additionally, the proposed conversation segmentation method demonstrates superior performance on dialogue segmentation datasets such as DialSeg711, TIAGE, and SuperDialSeg.

Zhiyuan He, Huiqiang Jiang, Zilong Wang et al. · microsoft-research

The performance of large language models (LLMs) is significantly influenced by the quality of the prompts provided. In response, researchers have developed enormous prompt engineering strategies aimed at modifying the prompt text to enhance task performance. In this paper, we introduce a novel technique termed position engineering, which offers a more efficient way to guide large language models. Unlike prompt engineering, which requires substantial effort to modify the text provided to LLMs, position engineering merely involves altering the positional information in the prompt without modifying the text itself. We have evaluated position engineering in two widely-used LLM scenarios: retrieval-augmented generation (RAG) and in-context learning (ICL). Our findings show that position engineering substantially improves upon the baseline in both cases. Position engineering thus represents a promising new strategy for exploiting the capabilities of large language models.

Ziqin Zhou, Yifan Yang, Yuqing Yang et al.

Text-to-video generation poses significant challenges due to the inherent complexity of video data, which spans both temporal and spatial dimensions. It introduces additional redundancy, abrupt variations, and a domain gap between language and vision tokens while generation. Addressing these challenges requires an effective video tokenizer that can efficiently encode video data while preserving essential semantic and spatiotemporal information, serving as a critical bridge between text and vision. Inspired by the observation in VQ-VAE-2 and workflows of traditional animation, we propose HiTVideo for text-to-video generation with hierarchical tokenizers. It utilizes a 3D causal VAE with a multi-layer discrete token framework, encoding video content into hierarchically structured codebooks. Higher layers capture semantic information with higher compression, while lower layers focus on fine-grained spatiotemporal details, striking a balance between compression efficiency and reconstruction quality. Our approach efficiently encodes longer video sequences (e.g., 8 seconds, 64 frames), reducing bits per pixel (bpp) by approximately 70\% compared to baseline tokenizers, while maintaining competitive reconstruction quality. We explore the trade-offs between compression and reconstruction, while emphasizing the advantages of high-compressed semantic tokens in text-to-video tasks. HiTVideo aims to address the potential limitations of existing video tokenizers in text-to-video generation tasks, striving for higher compression ratios and simplify LLMs modeling under language guidance, offering a scalable and promising framework for advancing text to video generation. Demo page: https://ziqinzhou66.github.io/project/HiTVideo.

Chendong Wang, Donglin Bai, Yifan Yang et al.

We present \emph{Video-in-the-Loop} (ViTL), a two-stage long-video QA framework that preserves a fixed token budget by first \emph{localizing} question-relevant interval(s) with a low-fps skim and then \emph{answering} via span-aware reallocation of visual tokens at higher effective frame rate, emitting an interleaved output with both spans and the final option for direct attribution. We also introduce \dataname{}, which converts description based event graphs into \emph{span-grounded} multiple-choice QA by pairing each question with \emph{ground-truth} time span(s) and related reasoning. ViTL is trained end-to-end with an interleaved group-relative objective that couples temporal IoU for localization with answer correctness, allowing credit to flow from answers back to spans without increasing compute. Under fixed token budgets, ViTL attains up to 8.6% with 50% less frame input on long-video QA and temporal grounding (e.g., Charades-STA, ActivityNet-Captions) and ablations show that span-aware token reallocation consistently surpasses uniform sampling. Together, \dataname{} and ViTL provide an interpretable, compute-efficient recipe for scalable long-video QA.

Kyoungjun Park, Yifan Yang, Juheon Yi et al.

With the rapid advancement of AI-generated videos, there is an urgent need for effective detection tools to mitigate societal risks such as misinformation and reputational harm. In addition to accurate classification, it is essential that detection models provide interpretable explanations to ensure transparency for regulators and end users. To address these challenges, we introduce VidGuard-R1, the first video authenticity detector that fine-tunes a multi-modal large language model (MLLM) using group relative policy optimization (GRPO). Our model delivers both highly accurate judgments and insightful reasoning. We curate a challenging dataset of 140k real and AI-generated videos produced by state-of-the-art generation models, carefully designing the generation process to maximize discrimination difficulty. We then fine-tune Qwen-VL using GRPO with two specialized reward models that target temporal artifacts and generation complexity. Extensive experiments demonstrate that VidGuard-R1 achieves state-of-the-art zero-shot performance on existing benchmarks, with additional training pushing accuracy above 95%. Case studies further show that VidGuard-R1 produces precise and interpretable rationales behind its predictions. The code is publicly available at https://VidGuard-R1.github.io.

Jiashuo Wang, Kaitao Song, Chunpu Xu et al.

Enhancing user engagement through interactions plays an essential role in socially-driven dialogues. While prior works have optimized models to reason over relevant knowledge or plan a dialogue act flow, the relationship between user engagement and knowledge or dialogue acts is subtle and does not guarantee user engagement in socially-driven dialogues. To this end, we enable interactive LLMs to learn user engagement by leveraging signals from the future development of conversations. Specifically, we adopt a more direct and relevant indicator of user engagement, i.e., the user's reaction related to dialogue intention after the interaction, as a reward to align interactive LLMs. To achieve this, we develop a user simulator to interact with target interactive LLMs and explore interactions between the user and the interactive LLM system via \textit{i$\times$MCTS} (\textit{M}onte \textit{C}arlo \textit{T}ree \textit{S}earch for \textit{i}nteraction). In this way, we collect a dataset containing pairs of higher and lower-quality experiences using \textit{i$\times$MCTS}, and align interactive LLMs for high-level user engagement by direct preference optimization (DPO) accordingly. Experiments conducted on two socially-driven dialogue scenarios (emotional support conversations and persuasion for good) demonstrate that our method effectively enhances user engagement in interactive LLMs.

Huiqiang Jiang, Li Lyna Zhang, Yuang Li et al.

Automatic Speech Recognition (ASR) has seen remarkable advancements with deep neural networks, such as Transformer and Conformer. However, these models typically have large model sizes and high inference costs, posing a challenge to deploy on resource-limited devices. In this paper, we propose a novel compression strategy that leverages structured pruning and knowledge distillation to reduce the model size and inference cost of the Conformer model while preserving high recognition performance. Our approach utilizes a set of binary masks to indicate whether to retain or prune each Conformer module, and employs L0 regularization to learn the optimal mask values. To further enhance pruning performance, we use a layerwise distillation strategy to transfer knowledge from unpruned to pruned models. Our method outperforms all pruning baselines on the widely used LibriSpeech benchmark, achieving a 50% reduction in model size and a 28% reduction in inference cost with minimal performance loss.