Xue Li

Yuxing Xiang, Xue Li, Kun Qian et al.

With the widespread adoption of Large Language Models (LLMs), serving LLM inference requests has become an increasingly important task, attracting active research advancements. Practical workloads play an essential role in this process: they are critical for motivating and benchmarking serving techniques and systems. However, the existing understanding of real-world LLM serving workloads is limited due to the lack of a comprehensive workload characterization. Prior analyses remain insufficient in scale and scope, thus failing to fully capture intricate workload characteristics. In this paper, we fill the gap with an in-depth characterization of LLM serving workloads collected from our worldwide cloud inference serving service, covering not only language models but also emerging multimodal and reasoning models, and unveiling important new findings in each case. Moreover, based on our findings, we propose ServeGen, a principled framework for generating realistic LLM serving workloads by composing them on a per-client basis. A practical use case in production validates that ServeGen avoids 50% under-provisioning compared to naive workload generation, demonstrating ServeGen's advantage in performance benchmarking. ServeGen is available at https://github.com/alibaba/ServeGen.

Manjiang Yu, Hongji Li, Junwei Chen et al.

Representation intervention has emerged as a promising paradigm for aligning large language models toward desired behaviors without modifying model weights. Existing methods typically apply a fixed intervention uniformly across all inputs. However, we find that the appropriate intervention direction and strength vary substantially across samples, and such indiscriminate intervention leads to degradation of general capabilities on benign inputs. To address these challenges, we propose Multi-Adapter Representation Interventions via Energy Calibration (MARI). Specifically, we introduce a competitive multi-adapter mechanism in which specialized experts capture non-linear correction patterns and adaptively determine the appropriate intervention direction and strength for different samples. Furthermore, we design an energy-based gating module that leverages internal propagation dynamics to distinguish inputs that are applicable for intervention. Extensive experiments across diverse model families and parameter scales demonstrate that MARI achieves state-of-the-art alignment performance. Our method significantly improves performance on TruthfulQA, BBQ, and safety benchmarks, while maintaining and even improving general capabilities on tasks such as MMLU and ARC. Our code is available at https://github.com/V1centNevwake/MARI.

Hong Sun, Xue Li, Yinchuan Xu et al. · microsoft-research

This paper presents AutoHint, a novel framework for automatic prompt engineering and optimization for Large Language Models (LLM). While LLMs have demonstrated remarkable ability in achieving high-quality annotation in various tasks, the key to applying this ability to specific tasks lies in developing high-quality prompts. Thus we propose a framework to inherit the merits of both in-context learning and zero-shot learning by incorporating enriched instructions derived from input-output demonstrations to optimize original prompt. We refer to the enrichment as the hint and propose a framework to automatically generate the hint from labeled data. More concretely, starting from an initial prompt, our method first instructs a LLM to deduce new hints for selected samples from incorrect predictions, and then summarizes from per-sample hints and adds the results back to the initial prompt to form a new, enriched instruction. The proposed method is evaluated on the BIG-Bench Instruction Induction dataset for both zero-shot and few-short prompts, where experiments demonstrate our method is able to significantly boost accuracy for multiple tasks.

Chaowei Fang, Dingwen Zhang, Wen Zheng et al.

Recently, long-tailed image classification harvests lots of research attention, since the data distribution is long-tailed in many real-world situations. Piles of algorithms are devised to address the data imbalance problem by biasing the training process towards less frequent classes. However, they usually evaluate the performance on a balanced testing set or multiple independent testing sets having distinct distributions with the training data. Considering the testing data may have arbitrary distributions, existing evaluation strategies are unable to reflect the actual classification performance objectively. We set up novel evaluation benchmarks based on a series of testing sets with evolving distributions. A corpus of metrics are designed for measuring the accuracy, robustness, and bounds of algorithms for learning with long-tailed distribution. Based on our benchmarks, we re-evaluate the performance of existing methods on CIFAR10 and CIFAR100 datasets, which is valuable for guiding the selection of data rebalancing techniques. We also revisit existing methods and categorize them into four types including data balancing, feature balancing, loss balancing, and prediction balancing, according the focused procedure during the training pipeline.

Xue Li, Wei Shen, Denis Charles · microsoft-research

In this paper, we propose TEDL, a two-stage learning approach to quantify uncertainty for deep learning models in classification tasks, inspired by our findings in experimenting with Evidential Deep Learning (EDL) method, a recently proposed uncertainty quantification approach based on the Dempster-Shafer theory. More specifically, we observe that EDL tends to yield inferior AUC compared with models learnt by cross-entropy loss and is highly sensitive in training. Such sensitivity is likely to cause unreliable uncertainty estimation, making it risky for practical applications. To mitigate both limitations, we propose a simple yet effective two-stage learning approach based on our analysis on the likely reasons causing such sensitivity, with the first stage learning from cross-entropy loss, followed by a second stage learning from EDL loss. We also re-formulate the EDL loss by replacing ReLU with ELU to avoid the Dying ReLU issue. Extensive experiments are carried out on varied sized training corpus collected from a large-scale commercial search engine, demonstrating that the proposed two-stage learning framework can increase AUC significantly and greatly improve training robustness.

Fuliang Liu, Xue Li, Ketai Zhao et al.

Speculative decoding is an effective and lossless approach for accelerating LLM inference. However, existing widely adopted model-based draft designs, such as EAGLE3, improve accuracy at the cost of multi-step autoregressive inference, resulting in high drafting latency and ultimately rendering the drafting stage itself a performance bottleneck. Inspired by diffusion-based large language models (dLLMs), we propose DART, which leverages parallel generation to reduce drafting latency. DART predicts logits for multiple future masked positions in parallel within a single forward pass based on hidden states of the target model, thereby eliminating autoregressive rollouts in the draft model while preserving a lightweight design. Based on these parallel logit predictions, we further introduce an efficient tree pruning algorithm that constructs high-quality draft token trees with N-gram-enforced semantic continuity. DART substantially reduces draft-stage overhead while preserving high draft accuracy, leading to significantly improved end-to-end decoding speed. Experimental results demonstrate that DART achieves a 2.03x--3.44x wall-clock time speedup across multiple datasets, surpassing EAGLE3 by 30% on average and offering a practical speculative decoding framework. Code is released at https://github.com/fvliang/DART.

Haoyu Ren, Xue Li, Darko Anicic et al.

The field of Tiny Machine Learning (TinyML) has made substantial advancements in democratizing machine learning on low-footprint devices, such as microcontrollers. The prevalence of these miniature devices raises the question of whether aggregating their knowledge can benefit TinyML applications. Federated meta-learning is a promising answer to this question, as it addresses the scarcity of labeled data and heterogeneous data distribution across devices in the real world. However, deploying TinyML hardware faces unique resource constraints, making existing methods impractical due to energy, privacy, and communication limitations. We introduce TinyMetaFed, a model-agnostic meta-learning framework suitable for TinyML. TinyMetaFed facilitates collaborative training of a neural network initialization that can be quickly fine-tuned on new devices. It offers communication savings and privacy protection through partial local reconstruction and Top-P% selective communication, computational efficiency via online learning, and robustness to client heterogeneity through few-shot learning. The evaluations on three TinyML use cases demonstrate that TinyMetaFed can significantly reduce energy consumption and communication overhead, accelerate convergence, and stabilize the training process.

Bingqing Zhang, Sen Wang, Yifan Liu et al.

Current video object detection (VOD) models often encounter issues with over-aggregation due to redundant aggregation strategies, which perform feature aggregation on every frame. This results in suboptimal performance and increased computational complexity. In this work, we propose an image-level Object Detection Difficulty (ODD) metric to quantify the difficulty of detecting objects in a given image. The derived ODD scores can be used in the VOD process to mitigate over-aggregation. Specifically, we train an ODD predictor as an auxiliary head of a still-image object detector to compute the ODD score for each image based on the discrepancies between detection results and ground-truth bounding boxes. The ODD score enhances the VOD system in two ways: 1) it enables the VOD system to select superior global reference frames, thereby improving overall accuracy; and 2) it serves as an indicator in the newly designed ODD Scheduler to eliminate the aggregation of frames that are easy to detect, thus accelerating the VOD process. Comprehensive experiments demonstrate that, when utilized for selecting global reference frames, ODD-VOD consistently enhances the accuracy of Global-frame-based VOD models. When employed for acceleration, ODD-VOD consistently improves the frames per second (FPS) by an average of 73.3% across 8 different VOD models without sacrificing accuracy. When combined, ODD-VOD attains state-of-the-art performance when competing with many VOD methods in both accuracy and speed. Our work represents a significant advancement towards making VOD more practical for real-world applications.

Bingqing Zhang, Zhuo Cao, Heming Du et al.

Text-Video Retrieval (TVR) methods typically match query-candidate pairs by aligning text and video features in coarse-grained, fine-grained, or combined (coarse-to-fine) manners. However, these frameworks predominantly employ a one(query)-to-one(candidate) alignment paradigm, which struggles to discern nuanced differences among candidates, leading to frequent mismatches. Inspired by Comparative Judgement in human cognitive science, where decisions are made by directly comparing items rather than evaluating them independently, we propose TokenBinder. This innovative two-stage TVR framework introduces a novel one-to-many coarse-to-fine alignment paradigm, imitating the human cognitive process of identifying specific items within a large collection. Our method employs a Focused-view Fusion Network with a sophisticated cross-attention mechanism, dynamically aligning and comparing features across multiple videos to capture finer nuances and contextual variations. Extensive experiments on six benchmark datasets confirm that TokenBinder substantially outperforms existing state-of-the-art methods. These results demonstrate its robustness and the effectiveness of its fine-grained alignment in bridging intra- and inter-modality information gaps in TVR tasks.

Yiwen Chen, Xue Li, Sheng Guo et al.

Reinforcement learning has shown a wide usage in robotics tasks, such as insertion and grasping. However, without a practical sim2real strategy, the policy trained in simulation could fail on the real task. There are also wide researches in the sim2real strategies, but most of those methods rely on heavy image rendering, domain randomization training, or tuning. In this work, we solve the insertion task using a pure visual reinforcement learning solution with minimum infrastructure requirement. We also propose a novel sim2real strategy, Real2Sim, which provides a novel and easier solution in policy adaptation. We discuss the advantage of Real2Sim compared with Sim2Real.

Jiahui Peng, He Yao, Jingwen Li et al.

Contrastive Language-Image Pre-training (CLIP) has demonstrated outstanding performance in global image understanding and zero-shot transfer through large-scale text-image alignment. However, the core of medical image analysis often lies in the fine-grained understanding of specific anatomical structures or lesion regions. Therefore, precisely comprehending region-of-interest (RoI) information provided by medical professionals or perception models becomes crucial. To address this need, we propose MedP-CLIP, a region-aware medical vision-language model (VLM). MedP-CLIP innovatively integrates medical prior knowledge and designs a feature-level region prompt integration mechanism, enabling it to flexibly respond to various prompt forms (e.g., points, bounding boxes, masks) while maintaining global contextual awareness when focusing on local regions. We pre-train the model on a meticulously constructed large-scale dataset (containing over 6.4 million medical images and 97.3 million region-level annotations), equipping it with cross-disease and cross-modality fine-grained spatial semantic understanding capabilities. Experiments demonstrate that MedP-CLIP significantly outperforms baseline methods in various medical tasks, including zero-shot recognition, interactive segmentation, and empowering multimodal large language models. This model provides a scalable, plug-and-play visual backbone for medical AI, combining holistic image understanding with precise regional analysis.

Zhuo Cao, Bingqing Zhang, Heming Du et al.

Text-guided Video Temporal Grounding (VTG) aims to localize relevant segments in untrimmed videos based on textual descriptions, encompassing two subtasks: Moment Retrieval (MR) and Highlight Detection (HD). Although previous typical methods have achieved commendable results, it is still challenging to retrieve short video moments. This is primarily due to the reliance on sparse and limited decoder queries, which significantly constrain the accuracy of predictions. Furthermore, suboptimal outcomes often arise because previous methods rank predictions based on isolated predictions, neglecting the broader video context. To tackle these issues, we introduce FlashVTG, a framework featuring a Temporal Feature Layering (TFL) module and an Adaptive Score Refinement (ASR) module. The TFL module replaces the traditional decoder structure to capture nuanced video content variations across multiple temporal scales, while the ASR module improves prediction ranking by integrating context from adjacent moments and multi-temporal-scale features. Extensive experiments demonstrate that FlashVTG achieves state-of-the-art performance on four widely adopted datasets in both MR and HD. Specifically, on the QVHighlights dataset, it boosts mAP by 5.8% for MR and 3.3% for HD. For short-moment retrieval, FlashVTG increases mAP to 125% of previous SOTA performance. All these improvements are made without adding training burdens, underscoring its effectiveness. Our code is available at https://github.com/Zhuo-Cao/FlashVTG.

Juexiang Ye, Xue Li, Xinyu Yang et al.

Large language model-based agents operating in long-horizon interactions require memory systems that support temporal consistency, multi-hop reasoning, and evidence-grounded reuse across sessions. Existing approaches largely rely on unstructured retrieval or coarse abstractions, which often lead to temporal conflicts, brittle reasoning, and limited traceability. We propose MemWeaver, a unified memory framework that consolidates long-term agent experiences into three interconnected components: a temporally grounded graph memory for structured relational reasoning, an experience memory that abstracts recurring interaction patterns from repeated observations, and a passage memory that preserves original textual evidence. MemWeaver employs a dual-channel retrieval strategy that jointly retrieves structured knowledge and supporting evidence to construct compact yet information-dense contexts for reasoning. Experiments on the LoCoMo benchmark demonstrate that MemWeaver substantially improves multi-hop and temporal reasoning accuracy while reducing input context length by over 95\% compared to long-context baselines.

Xuehui Yu, Jingchi Jiang, Xinmiao Yu et al.

Complex systems are ubiquitous in the real world and tend to have complicated and poorly understood dynamics. For their control issues, the challenge is to guarantee accuracy, robustness, and generalization in such bloated and troubled environments. Fortunately, a complex system can be divided into multiple modular structures that human cognition appears to exploit. Inspired by this cognition, a novel control method, Causal Coupled Mechanisms (CCMs), is proposed that explores the cooperation in division and competition in combination. Our method employs the theory of hierarchical reinforcement learning (HRL), in which 1) the high-level policy with competitive awareness divides the whole complex system into multiple functional mechanisms, and 2) the low-level policy finishes the control task of each mechanism. Specifically for cooperation, a cascade control module helps the series operation of CCMs, and a forward coupled reasoning module is used to recover the coupling information lost in the division process. On both synthetic systems and a real-world biological regulatory system, the CCM method achieves robust and state-of-the-art control results even with unpredictable random noise. Moreover, generalization results show that reusing prepared specialized CCMs helps to perform well in environments with different confounders and dynamics.

Junhua Ma, Jiajun Li, Yuxuan Liu et al.

Recent progress on parse tree encoder for sentence representation learning is notable. However, these works mainly encode tree structures recursively, which is not conducive to parallelization. On the other hand, these works rarely take into account the labels of arcs in dependency trees. To address both issues, we propose Dependency-Transformer, which applies a relation-attention mechanism that works in concert with the self-attention mechanism. This mechanism aims to encode the dependency and the spatial positional relations between nodes in the dependency tree of sentences. By a score-based method, we successfully inject the syntax information without affecting Transformer's parallelizability. Our model outperforms or is comparable to the state-of-the-art methods on four tasks for sentence representation and has obvious advantages in computational efficiency.

Xue Li, Yuanzhi Cheng

In designing and applying graph neural networks, we often fall into some optimization pitfalls, the most deceptive of which is that we can only build a deep model by solving over-smoothing. The fundamental reason is that we do not understand how graph neural networks work. Stress graph drawing can offer a unique viewpoint to message iteration in the graph, such as the root of the over-smoothing problem lies in the inability of graph models to maintain an ideal distance between nodes. We further elucidate the trigger conditions of over-smoothing and propose Stress Graph Neural Networks. By introducing the attractive and repulsive message passing from stress iteration, we show how to build a deep model without preventing over-smoothing, how to use repulsive information, and how to optimize the current message-passing scheme to approximate the full stress message propagation. By performing different tasks on 23 datasets, we verified the effectiveness of our attractive and repulsive models and the derived relationship between stress iteration and graph neural networks. We believe that stress graph drawing will be a popular resource for understanding and designing graph neural networks.

Ben Cao, Tiantian He, Xue Li et al.

In this paper, we present Reed-Solomon coded single-stranded representation learning (RSRL), a novel end-to-end model for learning representations for multi-modal lossless DNA storage. In contrast to existing learning-based methods, the proposed RSRL is inspired by both error-correction codec and structural biology. Specifically, RSRL first learns the representations for the subsequent storage from the binary data transformed by the Reed-Solomon codec. Then, the representations are masked by an RS-code-informed mask to focus on correcting the burst errors occurring in the learning process. With the decoded representations with error corrections, a novel biologically stabilized loss is formulated to regularize the data representations to possess stable single-stranded structures. By incorporating these novel strategies, the proposed RSRL can learn highly durable, dense, and lossless representations for the subsequent storage tasks into DNA sequences. The proposed RSRL has been compared with a number of strong baselines in real-world tasks of multi-modal data storage. The experimental results obtained demonstrate that RSRL can store diverse types of data with much higher information density and durability but much lower error rates.

Jiayun Li, Wen Hua, Shiqi Fan et al.

Temporal Entity Alignment (TEA), which aims to identify equivalent entities across Temporal Knowledge Graphs (TKGs), is crucial for integrating knowledge facts from multiple sources. However, existing TEA models often fail to capture the orthogonal yet complementary effects between structural and temporal features, and typically overlook the importance of information richness, a key factor for effective message passing in neural feature encoders. To address these limitations, we propose the RCTEA framework, which jointly models both structural and temporal aspects of TKGs for entity alignment. Specifically, we design a richness-guided attention mechanism along with an adaptive weighting strategy to facilitate effective feature fusion. To ensure robust alignment despite noisy entity contexts, we introduce a dual-view neighborhood consensus algorithm that jointly refines the feature encoders to enforce local structural consistency of the predicted alignments. Extensive experiments demonstrate the superiority of RCTEA, achieving state-of-the-art performance on public TEA benchmarks.

Haodong Yuan, Yudong Zhang, Shengyin Fan et al.

The integration of a SLAM algorithm with place recognition technology empowers it with the ability to mitigate accumulated errors and to relocalize itself. However, existing methods for point cloud-based place recognition predominantly rely on the matching of descriptors, which are mostly lidar-centric. These methods suffer from two major drawbacks: first, they cannot perform place recognition when the distance between two point clouds is significant, and second, they can only calculate the rotation angle without considering the offset in the X and Y directions. To overcome these limitations, we propose a novel local descriptor that is constructed around the Main Object. By using a geometric method, we can accurately calculate the relative pose. We have provided a theoretical analysis to demonstrate that this method can overcome the aforementioned limitations. Furthermore, we conducted extensive experiments on KITTI Odometry and KITTI360, which indicate that our proposed method has significant advantages over state-of-the-art methods.

Shuai Wu, Xue Li, Yanna Feng et al.

Large Language Models (LLMs), particularly those employing Mixture-of-Experts (MoE) architectures, have achieved remarkable capabilities across diverse natural language processing tasks. However, these models frequently suffer from hallucinations -- generating plausible but factually incorrect content -- and exhibit systematic biases that are amplified by uneven expert activation during inference. In this paper, we propose the Council Mode, a novel multi-agent consensus framework that addresses these limitations by dispatching queries to multiple heterogeneous frontier LLMs in parallel and synthesizing their outputs through a dedicated consensus model. The Council pipeline operates in three phases: (1) an intelligent triage classifier that routes queries based on complexity, (2) parallel expert generation across architecturally diverse models, and (3) a structured consensus synthesis that explicitly identifies agreement, disagreement, and unique findings before producing the final response. We implement and evaluate this architecture within an open-source AI workspace. Our comprehensive evaluation across multiple benchmarks demonstrates that the Council Mode achieves a 35.9% relative reduction in hallucination rates on the HaluEval benchmark and a 7.8-point improvement on TruthfulQA compared to the best-performing individual model, while maintaining significantly lower bias variance across domains. We provide the mathematical formulation of the consensus mechanism, detail the system architecture, and present extensive empirical results with ablation studies.

Xiangyu Meng, Xue Li, Qing Yang et al.

Benefiting from the advancements in deep learning, various genomic analytical techniques, such as survival analysis, classification of tumors and their subtypes, and exploration of specific pathways, have significantly enhanced our understanding of the biological mechanisms driving cancer. However, the overfitting issue, arising from the limited number of patient samples, poses a challenge in improving the accuracy of genome analysis by deepening the neural network. Furthermore, it remains uncertain whether novel approaches such as the sparsely gated mixture of expert (MOE) and self-attention mechanisms can improve the accuracy of genomic analysis. In this paper, we introduce a novel sparsely gated RNA-seq analysis framework called Gene-MOE. This framework exploits the potential of the MOE layers and the proposed mixture of attention expert (MOAE) layers to enhance the analysis accuracy. Additionally, it addresses overfitting challenges by integrating pan-cancer information from 33 distinct cancer types through pre-training.We pre-trained Gene-MOE on TCGA pan-cancer RNA-seq dataset with 33 cancer types. Subsequently, we conducted experiments involving cancer classification and survival analysis based on the pre-trained Gene-MOE. According to the survival analysis results on 14 cancer types, Gene-MOE outperformed state-of-the-art models on 12 cancer types. Through detailed feature analysis, we found that the Gene-MOE model could learn rich feature representations of high-dimensional genes. According to the classification results, the total accuracy of the classification model for 33 cancer classifications reached 95.8%, representing the best performance compared to state-of-the-art models. These results indicate that Gene-MOE holds strong potential for use in cancer classification and survival analysis.

Manjiang Yu, Hongji Li, Priyanka Singh et al.

Reliable behavior control is central to deploying large language models (LLMs) on the web. Activation steering offers a tuning-free route to align attributes (e.g., truthfulness) that ensure trustworthy generation. Prevailing approaches rely on coarse heuristics and lack a principled account of where to steer and how strongly to intervene. To this end, we propose Position-wise Injection with eXact Estimated Levels (PIXEL), a position-wise activation steering framework that, in contrast to prior work, learns a property-aligned subspace from dual views (tail-averaged and end-token) and selects intervention strength via a constrained geometric objective with a closed-form solution, thereby adapting to token-level sensitivity without global hyperparameter tuning. PIXEL further performs sample-level orthogonal residual calibration to refine the global attribute direction and employs a lightweight position-scanning routine to identify receptive injection sites. We additionally provide representation-level guarantees for the minimal-intervention rule, supporting reliable alignment. Across diverse models and evaluation paradigms, PIXEL consistently improves attribute alignment while preserving model general capabilities, offering a practical and principled method for LLMs' controllable generation. Our code is available at https://github.com/V1centNevwake/PIXEL-Adaptive-Steering

Haoyu Chen, Xue Li, Kun Qian et al.

In Large Language Model (LLM) inference services, it is challenging to make a parallelism strategy configuration, to efficiently process the requests of variance context lengths. Requests of long context require high degree of parallelism to provide more memory for Key-Value (KV) Cache, while requests of short context prefer low degree of parallelism to increase concurrency, thus improving throughput. To maintain high throughput while supporting large context lengths on demand, we propose Amoeba, a runtime Tensor Parallel (TP) transformation for online LLM inference services, which adaptively adjusts the TP of running instances to align with the dynamics of incoming requests. Evaluations using real-world traces show that Amoeba improves throughput by 1.75x-6.57x compared to state-of-the-art solutions.

Shibo Sun, Xue Li, Donglin Di et al.

While large language models (LLMs) have advanced procedural planning for embodied AI systems through strong reasoning abilities, the integration of multimodal inputs and counterfactual reasoning remains underexplored. To tackle these challenges, we introduce LLaPa, a vision-language model framework designed for multimodal procedural planning. LLaPa generates executable action sequences from textual task descriptions and visual environmental images using vision-language models (VLMs). Furthermore, we enhance LLaPa with two auxiliary modules to improve procedural planning. The first module, the Task-Environment Reranker (TER), leverages task-oriented segmentation to create a task-sensitive feature space, aligning textual descriptions with visual environments and emphasizing critical regions for procedural execution. The second module, the Counterfactual Activities Retriever (CAR), identifies and emphasizes potential counterfactual conditions, enhancing the model's reasoning capability in counterfactual scenarios. Extensive experiments on ActPlan-1K and ALFRED benchmarks demonstrate that LLaPa generates higher-quality plans with superior LCS and correctness, outperforming advanced models. The code and models are available https://github.com/sunshibo1234/LLaPa.

Bingqing Zhang, Yi Zhang, Zhuo Cao et al.

Composed video retrieval (CoVR) searches for target videos using a reference video and a modification text, but existing methods are restricted to a single interaction round and cannot support the progressive nature of real-world visual search. To bridge this gap, we first formalize interactive composed video retrieval, a multi-turn extension of CoVR, where users progressively refine their search intent through natural-language feedback across turns. Adapting existing interactive retrieval methods to this setting reveals two structural weaknesses: reliance on a single retrieval channel and an open-loop retrieval design that consumes user feedback but does not diagnose whether its own retrieval trajectory is drifting or stagnating. To address these limitations, we propose ReCoVR (Reflexive Composed Video Retrieval), a dual-pathway architecture built on reflexive perception, where the system treats its retrieval history as diagnostic evidence alongside user feedback. Specifically, an Intent Pathway routes heterogeneous feedback to complementary retrieval channels, while a Reflection Pathway performs trajectory-level reflection to monitor result evolution and correct retrieval errors across turns. Experiments on multiple benchmarks show that ReCoVR consistently outperforms interactive baselines, notably achieving 74.30% R@1 after just one interactive round on the WebVid-CoVR-Test dataset.

Zhuo Cao, Heming Du, Bingqing Zhang et al.

Existing Moment retrieval (MR) methods focus on Single-Moment Retrieval (SMR). However, one query can correspond to multiple relevant moments in real-world applications. This makes the existing datasets and methods insufficient for video temporal grounding. By revisiting the gap between current MR tasks and real-world applications, we introduce a high-quality datasets called QVHighlights Multi-Moment Dataset (QV-M$^2$), along with new evaluation metrics tailored for multi-moment retrieval (MMR). QV-M$^2$ consists of 2,212 annotations covering 6,384 video segments. Building on existing efforts in MMR, we propose a framework called FlashMMR. Specifically, we propose a Multi-moment Post-verification module to refine the moment boundaries. We introduce constrained temporal adjustment and subsequently leverage a verification module to re-evaluate the candidate segments. Through this sophisticated filtering pipeline, low-confidence proposals are pruned, and robust multi-moment alignment is achieved. We retrain and evaluate 6 existing MR methods on QV-M$^2$ and QVHighlights under both SMR and MMR settings. Results show that QV-M$^2$ serves as an effective benchmark for training and evaluating MMR models, while FlashMMR provides a strong baseline. Specifically, on QV-M$^2$, it achieves improvements over prior SOTA method by 3.00% on G-mAP, 2.70% on mAP@3+tgt, and 2.56% on mR@3. The proposed benchmark and method establish a foundation for advancing research in more realistic and challenging video temporal grounding scenarios. Code is released at https://github.com/Zhuo-Cao/QV-M2.

Gagandeep Singh, Samudi Amarsinghe, Priyanka Singh et al.

The rapid advances in generative models have significantly lowered the barrier to producing convincing multimodal disinformation. Fabricated images and manipulated captions increasingly co-occur to create persuasive false narratives. While the Detecting and Grounding Multi-Modal Media Manipulation (DGM4) dataset established a foundation for research in this area, it is restricted to local manipulations such as face swaps, attribute edits, and caption changes. This leaves a critical gap: global inconsistencies, such as mismatched foregrounds and backgrounds, which are now prevalent in real-world forgeries. To address this, we extend DGM4 with 5,000 high-quality samples that introduce Foreground-Background (FG-BG) mismatches and their hybrids with text manipulations. Using OpenAI's gpt-image-1 and carefully designed prompts, we generate human-centric news-style images where authentic figures are placed into absurd or impossible backdrops (e.g., a teacher calmly addressing students on the surface of Mars). Captions are produced under three conditions: literal, text attribute, and text split, yielding three new manipulation categories: FG-BG, FG-BG+TA, and FG-BG+TS. Quality control pipelines enforce one-to-three visible faces, perceptual hash deduplication, OCR-based text scrubbing, and realistic headline length. By introducing global manipulations, our extension complements existing datasets, creating a benchmark DGM4+ that tests detectors on both local and global reasoning. This resource is intended to strengthen evaluation of multimodal models such as HAMMER, which currently struggle with FG-BG inconsistencies. We release our DGM4+ dataset and generation script at https://github.com/Gaganx0/DGM4plus

Gagandeep Singh, Samudi Amarsinghe, Urawee Thani et al.

We extend HAMMER, a state-of-the-art model for multimodal manipulation detection, to handle global scene inconsistencies such as foreground-background (FG-BG) mismatch. While HAMMER achieves strong performance on the DGM4 dataset, it consistently fails when the main subject is contextually misplaced into an implausible background. We diagnose this limitation as a combination of label-space bias, local attention focus, and spurious text-foreground alignment. To remedy this without retraining, we propose a lightweight segmentation-guided scoring (SGS) pipeline. SGS uses person/face segmentation masks to separate foreground and background regions, extracts embeddings with a joint vision-language model, and computes region-aware coherence scores. These scores are fused with HAMMER's original prediction to improve binary detection, grounding, and token-level explanations. SGS is inference-only, incurs negligible computational overhead, and significantly enhances robustness to global manipulations. This work demonstrates the importance of region-aware reasoning in multimodal disinformation detection. We release scripts for segmentation and scoring at https://github.com/Gaganx0/HAMMER-sgs

Guoxin Zang, Xue Li, Donglin Di et al.

While Vision-Language Models (VLMs) have shown promising progress in general multimodal tasks, they often struggle in industrial anomaly detection and reasoning, particularly in delivering interpretable explanations and generalizing to unseen categories. This limitation stems from the inherently domain-specific nature of anomaly detection, which hinders the applicability of existing VLMs in industrial scenarios that require precise, structured, and context-aware analysis. To address these challenges, we propose SAGE, a VLM-based framework that enhances anomaly reasoning through Self-Guided Fact Enhancement (SFE) and Entropy-aware Direct Preference Optimization (E-DPO). SFE integrates domain-specific knowledge into visual reasoning via fact extraction and fusion, while E-DPO aligns model outputs with expert preferences using entropy-aware optimization. Additionally, we introduce AD-PL, a preference-optimized dataset tailored for industrial anomaly reasoning, consisting of 28,415 question-answering instances with expert-ranked responses. To evaluate anomaly reasoning models, we develop Multiscale Logical Evaluation (MLE), a quantitative framework analyzing model logic and consistency. SAGE demonstrates superior performance on industrial anomaly datasets under zero-shot and one-shot settings. The code, model and dataset are available at https://github.com/amoreZgx1n/SAGE.

Dingyi Yang, Junqi Zhao, Xue Li et al.

Cognitive anthropology suggests that the distinction of human intelligence lies in the ability to infer other individuals' knowledge states and understand their intentions. In comparison, our closest animal relative, chimpanzees, lack the capacity to do so. With this paper, we aim to evaluate LLM performance in estimating other individuals' knowledge states and their potential actions. We design two tasks to test (1) if LLMs can predict story characters' next actions based on their own knowledge vs. improperly using information unavailable from their perspective, and (2) if LLMs can detect when story characters, through their actions, demonstrate knowledge they should not possess. Results reveal that most current state-of-the-art LLMs achieve near-random performance on both tasks, and are substantially inferior to humans. We argue future LLM research should place more weight on the abilities of knowledge estimation and intention understanding.

Achmad Ginanjar, Xue Li, Wen Hua et al.

Learning from vertical partitioned data silos is challenging due to the segmented nature of data, sample misalignment, and strict privacy concerns. Federated learning has been proposed as a solution. However, sample misalignment across silos often hinders optimal model performance and suggests data sharing within the model, which breaks privacy. Our proposed solution is Contrastive Federated Learning with Tabular Data Silos (CFL), which offers a solution for data silos with sample misalignment without the need for sharing original or representative data to maintain privacy. CFL begins with local acquisition of contrastive representations of the data within each silo and aggregates knowledge from other silos through the federated learning algorithm. Our experiments demonstrate that CFL solves the limitations of existing algorithms for data silos and outperforms existing tabular contrastive learning. CFL provides performance improvements without loosening privacy.

Tao He, Xue Li, Zhibin Wang et al.

Training large-scale language models is increasingly critical in various domains, but it is hindered by frequent failures, leading to significant time and economic costs. Current failure recovery methods in cloud-based settings inadequately address the diverse and complex scenarios that arise, focusing narrowly on erasing downtime for individual tasks without considering the overall cost impact on a cluster. We introduce Unicron, a workload manager designed for efficient self-healing in large-scale language model training. Unicron optimizes the training process by minimizing failure-related costs across multiple concurrent tasks within a cluster. Its key features include in-band error detection for real-time error identification without extra overhead, a dynamic cost-aware plan generation mechanism for optimal reconfiguration, and an efficient transition strategy to reduce downtime during state changes. Deployed on a 128-GPU distributed cluster, Unicron demonstrates up to a 1.9x improvement in training efficiency over state-of-the-art methods, significantly reducing failure recovery costs and enhancing the reliability of large-scale language model training.

Shan Yu, Junyi Shu, Yuanjiang Ni et al.

As LLM applications grow more complex, developers are increasingly adopting multi-agent architectures to decompose workflows into specialized, collaborative components, introducing structure that constrains agent behavior and exposes useful semantic predictability. Unlike traditional LLM serving, which operates under highly dynamic and uncertain conditions, this structured topology enables opportunities to reduce runtime uncertainty -- yet existing systems fail to exploit it, treating agentic workloads as generic traffic and incurring significant inefficiencies. Our analysis of production traces from an agent-serving platform and an internal coding assistant reveals key bottlenecks, including low prefix cache hit rates, severe resource contention from long-context requests, and substantial queuing delays due to suboptimal scaling. To address these challenges, we propose Pythia, a multi-agent serving system that captures workflow semantics through a simple interface at the serving layer, unlocking new optimization opportunities and substantially improving throughput and job completion time over state-of-the-art baselines.

Shuai Wu, Xue Li, Yanna Feng et al.

Frontier image generation has moved from artistic synthesis toward synthetic visual evidence. Systems such as GPT Image 2, Nano Banana Pro, Nano Banana 2, Grok Imagine, Qwen Image 2.0 Pro, and Seedream 5.0 Lite combine photorealistic rendering, readable typography, reference consistency, editing control, and in several cases reasoning or search-grounded image construction. These capabilities create large benefits for design, education, accessibility, and communication, yet they also weaken one of society's most common trust shortcuts: the belief that a plausible picture is a reliable record. This paper provides a source-grounded technical and policy analysis of synthetic visual risk. We first summarize the public capabilities of recent image models, then analyze public incidents involving fake crisis images, celebrity and public-figure imagery, medical scans, forged-looking documents, synthetic screenshots, phishing assets, and market-moving rumors. We introduce a capability-weighted risk framework that links model affordances to real-world harm in finance, medicine, news, law, emergency response, identity verification, and civic discourse. Our findings show that risk is driven less by photorealism alone than by the convergence of realism, legible text, identity persistence, fast iteration, and distribution context. We argue for layered control: model-side restrictions, cryptographic provenance, visible labeling, platform friction, sector-grade verification, and incident response. The paper closes with practical recommendations for model providers, platforms, newsrooms, financial institutions, healthcare systems, legal organizations, regulators, and ordinary users.

Zhibin Wang, Shipeng Li, Yuhang Zhou et al.

User experience is a critical factor Large Language Model (LLM) serving systems must consider, where service level objectives (SLOs) considering the experience of individual requests and system level metrics (SLMs) considering the overall system performance are two key performance measures. However, we observe two notable issues in existing metrics: 1) manually delaying the delivery of some tokens can improve SLOs, and 2) actively abandoning requests that do not meet SLOs can improve SLMs, both of which are counterintuitive. In this paper, we revisit SLOs and SLMs in LLM serving, and propose a new SLO that aligns with user experience. Based on the SLO, we propose a comprehensive metric framework called smooth goodput, which integrates SLOs and SLMs to reflect the nature of user experience in LLM serving. Through this unified framework, we reassess the performance of different LLM serving systems under multiple workloads. Evaluation results show that our metric framework provides a more comprehensive view of token delivery and request processing, and effectively captures the optimal point of user experience and system performance with different serving strategies.

Haoyu Ren, Xue Li, Darko Anicic et al.

Recent advances in Tiny Machine Learning (TinyML) empower low-footprint embedded devices for real-time on-device Machine Learning. While many acknowledge the potential benefits of TinyML, its practical implementation presents unique challenges. This study aims to bridge the gap between prototyping single TinyML models and developing reliable TinyML systems in production: (1) Embedded devices operate in dynamically changing conditions. Existing TinyML solutions primarily focus on inference, with models trained offline on powerful machines and deployed as static objects. However, static models may underperform in the real world due to evolving input data distributions. We propose online learning to enable training on constrained devices, adapting local models towards the latest field conditions. (2) Nevertheless, current on-device learning methods struggle with heterogeneous deployment conditions and the scarcity of labeled data when applied across numerous devices. We introduce federated meta-learning incorporating online learning to enhance model generalization, facilitating rapid learning. This approach ensures optimal performance among distributed devices by knowledge sharing. (3) Moreover, TinyML's pivotal advantage is widespread adoption. Embedded devices and TinyML models prioritize extreme efficiency, leading to diverse characteristics ranging from memory and sensors to model architectures. Given their diversity and non-standardized representations, managing these resources becomes challenging as TinyML systems scale up. We present semantic management for the joint management of models and devices at scale. We demonstrate our methods through a basic regression example and then assess them in three real-world TinyML applications: handwritten character image classification, keyword audio classification, and smart building presence detection, confirming our approaches' effectiveness.

Shuai Wu, Xue Li, Yanna Feng et al.

As Large Language Models (LLMs) continue to evolve through alignment techniques such as Reinforcement Learning from Human Feedback (RLHF) and Constitutional AI, a growing and increasingly conspicuous phenomenon has emerged: the proliferation of verbal tics -- repetitive, formulaic linguistic patterns that pervade model outputs. These range from sycophantic openers ("That's a great question!", "Awesome!") to pseudo-empathetic affirmations ("I completely understand your concern", "I'm right here to catch you") and overused vocabulary ("delve", "tapestry", "nuanced"). In this paper, we present a systematic analysis of the verbal tic phenomenon across eight state-of-the-art LLMs: GPT-5.4, Claude Opus 4.7, Gemini 3.1 Pro, Grok 4.2, Doubao-Seed-2.0-pro, Kimi K2.5, DeepSeek V3.2, and MiMo-V2-Pro. Utilizing a custom evaluation framework for standardized API-based evaluation, we assess 10,000 prompts across 10 task categories in both English and Chinese, yielding 160,000 model responses. We introduce the Verbal Tic Index (VTI), a composite metric quantifying tic prevalence, and analyze its correlation with sycophancy, lexical diversity, and human-perceived naturalness. Our findings reveal significant inter-model variation: Gemini 3.1 Pro exhibits the highest VTI (0.590), while DeepSeek V3.2 achieves the lowest (0.295). We further demonstrate that verbal tics accumulate over multi-turn conversations, are amplified in subjective tasks, and show distinct cross-lingual patterns. Human evaluation (N = 120) confirms a strong inverse relationship between sycophancy and perceived naturalness (r = -0.87, p < 0.001). These results underscore the "alignment tax" of current training paradigms and highlight the urgent need for more authentic human-AI interaction frameworks.

Zhibin Wang, Rui Ning, Chao Fang et al.

Prefix-sharing among multiple prompts presents opportunities to combine the operations of the shared prefix, while attention computation in the decode stage, which becomes a critical bottleneck with increasing context lengths, is a memory-intensive process requiring heavy memory access on the key-value (KV) cache of the prefixes. Therefore, in this paper, we explore the potential of prefix-sharing in the attention computation of the decode stage. However, the tree structure of the prefix-sharing mechanism presents significant challenges for attention computation in efficiently processing shared KV cache access patterns while managing complex dependencies and balancing irregular workloads. To address the above challenges, we propose a dedicated attention kernel to combine the memory access of shared prefixes in the decoding stage, namely FlashForge. FlashForge delivers two key innovations: a novel shared-prefix attention kernel that optimizes memory hierarchy and exploits both intra-block and inter-block parallelism, and a comprehensive workload balancing mechanism that efficiently estimates cost, divides tasks, and schedules execution. Experimental results show that FlashForge achieves an average 1.9x speedup and 120.9x memory access reduction compared to the state-of-the-art FlashDecoding kernel regarding attention computation in the decode stage and 3.8x end-to-end time per output token compared to the vLLM.

Zhibin Wang, Shipeng Li, Xue Li et al.

Large language models have been widely deployed in various applications, encompassing both interactive online tasks and batched offline tasks. Given the burstiness and latency sensitivity of online tasks, over-provisioning resources is common practice. This allows for the integration of latency-insensitive offline tasks during periods of low online load, enhancing resource utilization. However, strategically serving online and offline tasks through a preemption mechanism fails to fully leverage the flexibility of offline tasks and suffers from KV cache recomputation and irregular workloads. In this paper, we introduce Echo, a collaborative online-offline task serving system, including a scheduler, a KV cache manager, and estimation toolkits. The scheduler and KV cache manager work tightly to maximize the throughput of offline tasks, while the estimator further predicts execution time to ensure online task SLOs. The scheduler leverages the batch information of last iteration to reduce the search space for finding the optimal schedule. The KV cache manager sets the priority of the KV cache based on the type of tasks and the opportunity of prefix sharing to reduce the recomputation. Finally, the estimation toolkits predict the execution time, future memory consumption, and the throughput of offline tasks to guide the scheduler, KV cache manager, and the system deployer. Evaluation based on real-world workloads demonstrates that Echo can increase offline task throughput by up to $3.3\times$, while satisfying online task SLOs.

Bingqing Zhang, Zhuo Cao, Heming Du et al.

Despite recent advances, Text-to-video retrieval (TVR) is still hindered by multiple inherent uncertainties, such as ambiguous textual queries, indistinct text-video mappings, and low-quality video frames. Although interactive systems have emerged to address these challenges by refining user intent through clarifying questions, current methods typically rely on heuristic or ad-hoc strategies without explicitly quantifying these uncertainties, limiting their effectiveness. Motivated by this gap, we propose UMIVR, an Uncertainty-Minimizing Interactive Text-to-Video Retrieval framework that explicitly quantifies three critical uncertainties-text ambiguity, mapping uncertainty, and frame uncertainty-via principled, training-free metrics: semantic entropy-based Text Ambiguity Score (TAS), Jensen-Shannon divergence-based Mapping Uncertainty Score (MUS), and a Temporal Quality-based Frame Sampler (TQFS). By adaptively generating targeted clarifying questions guided by these uncertainty measures, UMIVR iteratively refines user queries, significantly reducing retrieval ambiguity. Extensive experiments on multiple benchmarks validate UMIVR's effectiveness, achieving notable gains in Recall@1 (69.2\% after 10 interactive rounds) on the MSR-VTT-1k dataset, thereby establishing an uncertainty-minimizing foundation for interactive TVR.

Binxiong Li, Yuefei Wang, Binyu Zhao et al.

This study introduces the Multi-Scale Weight-Based Pairwise Coarsening and Contrastive Learning (MPCCL) model, a novel approach for attributed graph clustering that effectively bridges critical gaps in existing methods, including long-range dependency, feature collapse, and information loss. Traditional methods often struggle to capture high-order graph features due to their reliance on low-order attribute information, while contrastive learning techniques face limitations in feature diversity by overemphasizing local neighborhood structures. Similarly, conventional graph coarsening methods, though reducing graph scale, frequently lose fine-grained structural details. MPCCL addresses these challenges through an innovative multi-scale coarsening strategy, which progressively condenses the graph while prioritizing the merging of key edges based on global node similarity to preserve essential structural information. It further introduces a one-to-many contrastive learning paradigm, integrating node embeddings with augmented graph views and cluster centroids to enhance feature diversity, while mitigating feature masking issues caused by the accumulation of high-frequency node weights during multi-scale coarsening. By incorporating a graph reconstruction loss and KL divergence into its self-supervised learning framework, MPCCL ensures cross-scale consistency of node representations. Experimental evaluations reveal that MPCCL achieves a significant improvement in clustering performance, including a remarkable 15.24% increase in NMI on the ACM dataset and notable robust gains on smaller-scale datasets such as Citeseer, Cora and DBLP.

Guanlin Liu, Xiaomei Yu, Zihao Liu et al.

Medication Recommendation (MR) is a promising research topic which booms diverse applications in the healthcare and clinical domains. However, existing methods mainly rely on sequential modeling and static graphs for representation learning, which ignore the dynamic correlations in diverse medical events of a patient's temporal visits, leading to insufficient global structural exploration on nodes. Additionally, mitigating drug-drug interactions (DDIs) is another issue determining the utility of the MR systems. To address the challenges mentioned above, this paper proposes a novel MR method with the integration of dynamic networks and multi-view drug representations (DNMDR). Specifically, weighted snapshot sequences for dynamic heterogeneous networks are constructed based on discrete visits in temporal EHRs, and all the dynamic networks are jointly trained to gain both structural correlations in diverse medical events and temporal dependency in historical health conditions, for achieving comprehensive patient representations with both semantic features and structural relationships. Moreover, combining the drug co-occurrences and adverse drug-drug interactions (DDIs) in internal view of drug molecule structure and interactive view of drug pairs, the safe drug representations are available to obtain high-quality medication combination recommendation. Finally, extensive experiments on real world datasets are conducted for performance evaluation, and the experimental results demonstrate that the proposed DNMDR method outperforms the state-of-the-art baseline models with a large margin on various metrics such as PRAUC, Jaccard, DDI rates and so on.

Xue Li, Xiaonan Song, Henry Hu

Real-world deployment of Vision-Language Models (VLMs) is hindered by high computational demands, as existing architectures inefficiently process all tokens uniformly. We introduce Adaptive Token Pruning (ATP), a dynamic inference mechanism that retains only the most informative tokens based on contextual relevance. ATP operates at the vision-language interface, assigning a hybrid importance score combining ViT CLS attention (intra-modal saliency) and CLIP text-image similarity (inter-modal relevance) to keep top-K tokens for the LLM. Unlike static compression, ATP adapts to each input without modifying the backbone. Proposed as a lightweight gating module, ATP is compatible with popular backbones like BLIP-2, LLaVA, and Flamingo. Preliminary evaluations across VQAv2, GQA, and COCO indicate that ATP reduces inference FLOPs by around 40% and achieves roughly 1.5x speedups in end-to-end latency with negligible accuracy loss (less than 1%). Qualitative analyses suggest ATP preserves visual grounding and enhances interpretability. Beyond efficiency, we investigate robustness under corruptions; observations suggest adaptive pruning suppresses spurious correlations, improving stability. These findings imply that resource-constrained inference and model reliability are not competing objectives. Finally, we discuss ATP's role in efficient multimodal edge computing pipelines.

Bingqing Zhang, Zhuo Cao, Heming Du et al.

Modern video-text retrieval (VTR) models excel on in-distribution benchmarks but are highly vulnerable to real-world query shifts, where the distribution of query data deviates from the training domain, leading to a sharp performance drop. Existing image-focused robustness solutions are inadequate to handle this vulnerability in video, as they fail to address the complex spatio-temporal dynamics inherent in these shifts. To systematically evaluate this vulnerability, we first introduce a comprehensive benchmark featuring 12 distinct types of video perturbations across five severity degrees. Analysis on this benchmark reveals that query shifts amplify the hubness phenomenon, where a few gallery items become dominant "hubs" that attract a disproportionate number of queries. To mitigate this, we then propose HAT-VTR (Hubness Alleviation for Test-time Video-Text Retrieval), as our baseline test-time adaptation framework designed to directly counteract hubness in VTR. It leverages two key components: a Hubness Suppression Memory to refine similarity scores, and multi-granular losses to enforce temporal feature consistency. Extensive experiments demonstrate that HAT-VTR substantially improves robustness, consistently outperforming prior methods across diverse query shift scenarios, and enhancing model reliability for real-world applications.

Hongji Li, Junchi yao, Manjiang Yu et al.

Machine unlearning aims to erase requested data from trained models without full retraining. For Reasoning Multimodal Large Language Models (RMLLMs), this is uniquely challenging: intermediate chain-of-thought steps can still leak sensitive information even when final answers are forgotten, and overly aggressive interventions easily damage general reasoning ability. Yet no benchmark jointly evaluates how well unlearning methods suppress reasoning-level leakage while preserving reasoning competence. We address this gap with RMLLMU-Bench, the first benchmark for RMLLM unlearning that extends standard forgetting metrics with dedicated measures of reasoning leakage and reasoning retention. A systematic evaluation on RMLLMU-Bench reveals that existing unlearning methods for MLLMs and Large (Language) Reasoning Models (LRMs) either leave substantial leakage in the reasoning process or severely degrade reasoning performance. To address these gaps, we propose R-MUSE (Reasoning-preserving MLLM Unlearning via Subspace guidance and Adaptive Steering), a training-free and inference-time intervention framework that steers internal representations to forget both answers and reasoning traces while explicitly preserving general reasoning. Experiments on RMLLMU-Bench demonstrate that R-MUSE achieves a substantially better balance between effective forgetting and reasoning retention.

Zhibin Wang, Zetao Hong, Xue Li et al.

Large Language Model (LLM) inference has emerged as a fundamental paradigm. In real-world scenarios, variations in output length cause severe workload imbalance in the decode phase, particularly for long-output reasoning tasks. Existing systems, such as PD disaggregation architectures, rely on static prefill-to-decode scheduling, which often results in SLO violations and OOM failures under evolving decode workloads. In this paper, we propose ARES, an adaptive decoding rescheduling system powered by length prediction to anticipate future workloads. Our core contributions include: (1) A lightweight and continuous LLM-native prediction method that leverages LLM hidden state to model remaining generation length with high precision (reducing MAE by 49.42%) and low overhead (cutting predictor parameters by 93.28%); (2) A rescheduling solution in decode phase with : A dynamic balancing mechanism that integrates current and predicted workloads, reducing P99 TPOT by 74.77% and achieving up to 2.24 times higher goodput.

Manjiang Yu, Priyanka Singh, Xue Li et al.

Large language models (LLMs) frequently memorize sensitive or personal information, raising significant privacy concerns. Existing variants of differential privacy stochastic gradient descent (DPSGD) inject uniform noise into every gradient step, significantly extending training time and reducing model accuracy. We propose that concentrating noise primarily on gradients associated with sensitive tokens can substantially decrease DP training time, strengthen the protection of sensitive information, and simultaneously preserve the model's performance on non-sensitive data. We operationalize this insight through Adaptive Token-Weighted Differential Privacy (ATDP), a modification of vanilla DP-SGD that adaptively assigns different gradient weights to sensitive and non-sensitive tokens. By employing a larger noise scale at the early stage of training, ATDP rapidly disrupts memorization of sensitive content. As a result, ATDP only requires a few additional epochs of lightweight post-processing following standard fine-tuning, injecting targeted noise primarily on parameters corresponding to sensitive tokens, thus minimally affecting the model's general capabilities. ATDP can be seamlessly integrated into any existing DP-based fine-tuning pipeline or directly applied to non-private models as a fast privacy-enhancing measure. Additionally, combined with an initial redacted fine-tuning phase, ATDP forms a streamlined DP pipeline that achieves comparable canary protection to state-of-the-art DP-SGD methods, significantly reduces the computational overhead of DP fine-tuning, shortening training time by approximately 90 percent, while achieving comparable or superior privacy protection and minimal accuracy degradation.

Binxiong Li, Xu Xiang, Xue Li et al.

Attributed graph clustering holds significant importance in modern data analysis. However, due to the complexity of graph data and the heterogeneity of node attributes, leveraging graph information for clustering remains challenging. To address this, we propose a novel deep graph clustering model, GCL-GCN, specifically designed to address the limitations of existing models in capturing local dependencies and complex structures when dealing with sparse and heterogeneous graph data. GCL-GCN introduces an innovative Graphormer module that combines centrality encoding and spatial relationships, effectively capturing both global and local information between nodes, thereby enhancing the quality of node representations. Additionally, we propose a novel contrastive learning module that significantly enhances the discriminative power of feature representations. In the pre-training phase, this module increases feature distinction through contrastive learning on the original feature matrix, ensuring more identifiable initial representations for subsequent graph convolution and clustering tasks. Extensive experimental results on six datasets demonstrate that GCL-GCN outperforms 14 advanced methods in terms of clustering quality and robustness. Specifically, on the Cora dataset, it improves ACC, NMI, and ARI by 4.94%, 13.01%, and 10.97%, respectively, compared to the primary comparison method MBN.

Binxiong Li, Xu Xiang, Xue Li et al.

In recent years, models based on Graph Convolutional Networks (GCN) have made significant strides in the field of graph data analysis. However, challenges such as over-smoothing and over-compression remain when handling large-scale and complex graph datasets, leading to a decline in clustering quality. Although the Graph Transformer architecture has mitigated some of these issues, its performance is still limited when processing heterogeneous graph data. To address these challenges, this study proposes a novel deep clustering framework that comprising GCN, Autoencoder (AE), and Graph Transformer, termed the Tri-Learn Graph Fusion Network (Tri-GFN). This framework enhances the differentiation and consistency of global and local information through a unique tri-learning mechanism and feature fusion enhancement strategy. The framework integrates GCN, AE, and Graph Transformer modules. These components are meticulously fused by a triple-channel enhancement module, which maximizes the use of both node attributes and topological structures, ensuring robust clustering representation. The tri-learning mechanism allows mutual learning among these modules, while the feature fusion strategy enables the model to capture complex relationships, yielding highly discriminative representations for graph clustering. It surpasses many state-of-the-art methods, achieving an accuracy improvement of approximately 0.87% on the ACM dataset, 14.14 % on the Reuters dataset, and 7.58 % on the USPS dataset. Due to its outstanding performance on the Reuters dataset, Tri-GFN can be applied to automatic news classification, topic retrieval, and related fields.

Fenghua Cheng, Jinxiang Wang, Sen Wang et al.

Multimodal reasoning is a process of understanding, integrating and inferring information across different data modalities. It has recently attracted surging academic attention as a benchmark for Artificial Intelligence (AI). Although there are various tasks for evaluating multimodal reasoning ability, they still have limitations. Lack of reasoning on hierarchical visual clues at different levels of granularity, e.g., local details and global context, is of little discussion, despite its frequent involvement in real scenarios. To bridge the gap, we introduce a novel and challenging task for multimodal reasoning, namely GeoGuess. Given a street view image, the task is to identify its location and provide a detailed explanation. A system that succeeds in GeoGuess should be able to detect tiny visual clues, perceive the broader landscape, and associate with vast geographic knowledge. Therefore, GeoGuess would require the ability to reason between hierarchical visual information and geographic knowledge. In this work, we establish a benchmark for GeoGuess by introducing a specially curated dataset GeoExplain which consists of panoramas-geocoordinates-explanation tuples. Additionally, we present a multimodal and multilevel reasoning method, namely SightSense which can make prediction and generate comprehensive explanation based on hierarchy of visual information and external knowledge. Our analysis and experiments demonstrate their outstanding performance in GeoGuess.