arXiv subjects

Le Xu

Publications and source records attributed to Le Xu.

At least 19 recordsLinked to original sources

arXiv2026

Read Less, Solve More: Token-Efficient Sparse Reading for AI Agents

Long-horizon agents increasingly rely on repeated access to external artifacts, yet current reading interfaces often expose entire objects even when only sparse evidence is needed. This over-reading increases token and latency costs and can dilute task-relevant evidence, while existing context-reduction methods mainly intervene after broad content has already entered the trajectory. We present SparseRead, a training-free, model-transparent reading layer that controls content admission before unnecessary evidence reaches the model context. SparseRead combines a regime-aware Read Gate, extensible Reader Backends, and a stateful protocol for bounded, source-anchored evidence acquisition with explicit refinement, verification, stopping, and fallback. Across six frontier models, including Claude Opus 5, and five workload scenarios, SparseRead reduces token volume by up to 92.9% and wall time by up to 89.0%, while preserving or improving task quality. Its consistent gains across three agent frameworks further demonstrate broad portability.

cs.AI↗

arXiv2026

LLMVisor: A Real-Time Latency Attribution Model for Multi-Tenant LLM Serving

As LLM inference shifts to multi-tenant GPU clusters, co-batching improves throughput but obscures per-tenant usage and limits control. Enabling fractional sharing of the inference engine requires a real-time, per-request attribution primitive that is accurate and light enough to run inside the scheduling loop. We present LLMVisor, a roofline-guided latency attribution model that captures the memory-bound and compute-bound phases via a concise piecewise-linear form over features proportional to FLOPs and memory I/O traffic. LLMVisor decomposes batch latency into additive, per-request shares and runs efficiently at microsecond scale. We evaluate LLMVisor across Llama 3.1-8B and Qwen 2.5-14B/32B on A100/H100 GPUs under varying tensor parallelism and workload mixes. Compared to a token-count baseline, LLMVisor attains near-perfect R-squared and reduces relative error by up to 2.5x and 3.3x at p90 and p99, respectively, for prefill, and by up to 3.5x and 4.4x for decode, despite batching variability and sequence divergence.

cs.AI↗

arXiv2026

WorldSample: Closed-loop Real-robot RL with World Modelling

Reinforcement learning (RL) can overcome the demonstration-coverage limitation of imitation learning (IL) by allowing robots to improve through trial-and-error interaction beyond the states observed in demonstrations. However, deploying RL on real robots remains constrained by high interaction costs, since each physical rollout is costly and reflects only one realized action-outcome path. To address this challenge, we propose WorldSample, a physically grounded data augmentation framework for real-robot RL that closes a real-synthetic loop between physical rollouts, world-model generation, and policy improvement. Grounded on real rollouts, WorldSample generates high-fidelity synthetic transitions through a post-trained world model, which greatly lowers the visual hallucination. Specifically, rather than simply using these transitions as real-world experience, WorldSample introduces Policy-Paced Learning (PPL) to regulate the training process through sample selection and scheduling, balancing useful augmentation against value overestimation and mitigating the hallucination-induced noise. Experiments on robot manipulation tasks involving contact-rich and precise tasks show that WorldSample improves policy success rate by 28% while reducing training steps by 59% compared with baselines. Furthermore, WorldSample improves world model visual fidelity by 19.4dB in PSNR and 0.47 in SSIM over demonstration-only post-training, validating the effectiveness of the real-synthetic loop for both policy and world model performance.

cs.RO↗

arXiv2026

SwarmX: Agentic Scheduling for Low-Latency Agentic Systems

Agentic AI applications compose multiple model calls and tool executions, creating new scheduling challenges for GPU-CPU clusters. Their inference time and model-call structure often depend on prompt semantics, making conventional scheduling approaches ineffective for low-latency serving. This paper presents SwarmX, a system that implements agentic scheduling for low-latency agentic applications. SwarmX uses scheduling-specific neural predictors to capture prompt, device, runtime, and target-model features; exposes distributional predictions to routers and scalers for tail-aware decisions; and provides mechanisms for predictor training and online adaptation. These predictors and mechanisms are integrated into a scheduler-agent framework that provides a common substrate for integration with existing scheduling and model-serving infrastructure. We evaluate SwarmX using production deployment (nearly one thousand GPUs and one million CPU cores) and controlled experiments on a 128-GPU testbed. Across multi-agent code generation, deep research, and multimodal agentic workflows, SwarmX reduces tail latency by up to 61.5% compared to state-of-the-art schedulers and sustains up to 2x the throughput of production schedulers under the same SLO.

cs.DC↗

arXiv2026

BatchGen: An Architecture for Scalable and Efficient Batch Inference

Batch inference has become a central mode of AI computation, yet existing inference engines still rely on execution models designed for interactive serving. When scaled to millions of sequences, batch workloads reveal two fundamental requirements: the ability to handle extreme inter- and intra-sequence load variation that emerges only at runtime, and the ability to sustain high utilization across large fleets of GPUs. Existing systems fail to meet these requirements, losing substantial fractions of achievable throughput. We introduce a new architectural foundation for batch inference: the sequence coroutine compute model, which represents each sequence as a fine-grained, event-driven coroutine. This model exposes expressive primitives that allow the runtime to reorganize work dynamically, enabling larger expert-level batches, mitigating stragglers, reallocating work across devices, and maintaining utilization even on cost-effective or memory-constrained GPUs. Building on this abstraction, we implement BatchGen, a production-ready system that uses the coroutine model at cluster scale. On a 128-GPU cluster, BatchGen reduces batch completion time by up to $2.3\times$, and on memory-constrained accelerators it outperforms the strongest offloading baseline by up to $9.6\times$. We will open-source BatchGen at https://github.com/batchgen-project/batchgen

cs.DC↗

arXiv2026

Lodestar: An Online-Learning LLM Inference Router

Efficiently serving large language model (LLM) inference tasks is crucial both for user-perceived latency such as time-to-first-token (TTFT) and for GPU utilization. However, LLM request routing, that is, assigning each inference request to a GPU instance, is particularly challenging: execution is highly input-dependent; batching and KV-cache reuse create strong cross-request coupling; and latency responds nonlinearly to context length, model/engine settings, and heterogeneous accelerators. As a result, simple traditional load balancing algorithms, and even heuristics tailored for LLM inference, fail to achieve good performance. We present Lodestar, a novel learning-based request routing system for distributed GPU clusters. Lodestar continuously collects a snapshot of the cluster at per-request level, including real-time instance state, request characteristics, and observed performance, and trains an online reward predictor that it uses to route inference requests to the instance that will maximize given reward (e.g., minimizing TTFT). Lodestar is cloud-native and works seamlessly with existing serving stacks (vLLM). With continuous online adaptation to changing workloads and infrastructure conditions, Lodestar achieves 1.41x lower average TTFT and 1.47x lower P99 TTFT on average (up to 2.15x/1.86x on homogeneous and 4.38x/4.42x on heterogeneous clusters) compared to a state-of-the-art prefix cache and load-aware heuristic, and learns these efficient routing strategies within about 5 minutes, based on experiments in a public cloud GPU cluster.

cs.DC↗

arXiv2026

A Triple-Modal Contrastive Learning Framework with Sequence, Graph, and 3D Features for Drug-Target Interaction Prediction

Accurate prediction of drug-target interactions (DTI) is critical for drug discovery. Existing methods often rely on single-modal representations (e.g., sequences or graphs) or combine only two modalities, overlooking 3D structural features. To address this challenge, we propose TriMod-DTI, a triple-modal contrastive learning framework that incorporates 1D sequences, 2D graphs, and 3D structures of drugs and proteins, obtaining the universal and complementary feature representations for DTI prediction. We design a Feature Extractor to capture drug and target features across the three modalities, thereby enriching their representations. We further propose a triple-modal contrastive learning strategy to align different modal representations of the same drug or protein in the latent space. By constructing cross-modal positive and negative sample pairs, this approach enhances the model's discriminative ability. Experiments on three benchmark datasets demonstrate that TriMod-DTI outperforms state-of-the-art methods. The ablation studies validate the contributions of each modality. Moreover, case studies highlight its practical potential for DTI prediction and drug discovery.

cs.LG↗

arXiv2026

SlideSparse: Fast and Flexible (2N-2):2N Structured Sparsity

NVIDIA's 2:4 Sparse Tensor Cores deliver 2x throughput but demand strict 50% pruning -- a ratio that collapses LLM reasoning accuracy (Qwen3: 54% to 15%). Milder $(2N-2):2N$ patterns (e.g., 6:8, 25% pruning) preserve accuracy yet receive no hardware support, falling back to dense execution without any benefit from sparsity. We present SlideSparse, the first system to unlock Sparse Tensor Core acceleration for the $(2N-2):2N$ model family on commodity GPUs. Our Sliding Window Decomposition reconstructs any $(2N-2):2N$ weight block into $N-1$ overlapping 2:4-compliant windows without any accuracy loss; Activation Lifting fuses the corresponding activation rearrangement into per-token quantization at near-zero cost. Integrated into vLLM, SlideSparse is evaluated across various GPUs (A100, H100, B200, RTX 4090, RTX 5080, DGX-spark), precisions (FP4, INT8, FP8, BF16, FP16), and model families (Llama, Qwen, BitNet). On compute-bound workloads, the measured speedup ratio (1.33x) approaches the theoretical upper-bound $N/(N-1)=4/3$ at 6:8 weight sparsity in Qwen2.5-7B, establishing $(2N-2):2N$ as a practical path to accuracy-preserving LLM acceleration. Code available at https://github.com/bcacdwk/vllmbench.

cs.LG↗

arXiv2026

Covo-Audio Technical Report

In this work, we present Covo-Audio, a 7B-parameter end-to-end LALM that directly processes continuous audio inputs and generates audio outputs within a single unified architecture. Through large-scale curated pretraining and targeted post-training, Covo-Audio achieves state-of-the-art or competitive performance among models of comparable scale across a broad spectrum of tasks, including speech-text modeling, spoken dialogue, speech understanding, audio understanding, and full-duplex voice interaction. Extensive evaluations demonstrate that the pretrained foundation model exhibits strong speech-text comprehension and semantic reasoning capabilities on multiple benchmarks, outperforming representative open-source models of comparable scale. Furthermore, Covo-Audio-Chat, the dialogue-oriented variant, demonstrates strong spoken conversational abilities, including understanding, contextual reasoning, instruction following, and generating contextually appropriate and empathetic responses, validating its applicability to real-world conversational assistant scenarios. Covo-Audio-Chat-FD, the evolved full-duplex model, achieves substantially superior performance on both spoken dialogue capabilities and full-duplex interaction behaviors, demonstrating its competence in practical robustness. To mitigate the high cost of deploying end-to-end LALMs for natural conversational systems, we propose an intelligence-speaker decoupling strategy that separates dialogue intelligence from voice rendering, enabling flexible voice customization with minimal text-to-speech (TTS) data while preserving dialogue performance. Overall, our results highlight the strong potential of 7B-scale models to integrate sophisticated audio intelligence with high-level semantic reasoning, and suggest a scalable path toward more capable and versatile LALMs.

cs.SD↗

arXiv2026

Edit Content, Preserve Acoustics: Imperceptible Text-Based Speech Editing via Self-Consistency Rewards

Imperceptible text-based speech editing modifies spoken content through transcript manipulation while preserving acoustic continuity. Prior acoustic-space approaches suffer from content-style entanglement, causing unstable generation and boundary artifacts. We introduce a framework guided by the principle of "Edit Content, Preserve Acoustics". Editing is conducted in a stable semantic space, while acoustic realization is handled by a Flow Matching decoder. To ensure perceptual consistency, we propose Self-Consistency Rewards Group Relative Policy Optimization, which leverages a pre-trained Text-to-Speech model as an implicit critic, together with intelligibility and duration constraints. Experiments demonstrate consistent improvements over state-of-the-art autoregressive and non-autoregressive baselines in intelligibility, robustness, and perceptual quality.

cs.SD↗

arXiv2026

OV-InstructTTS: Towards Open-Vocabulary Instruct Text-to-Speech

Instruct Text-to-Speech (InstructTTS) leverages natural language descriptions as style prompts to guide speech synthesis. However, existing InstructTTS methods mainly rely on a direct combination of audio-related labels or their diverse rephrasings, making it difficult to handle flexible, high-level instructions. Such rigid control is insufficient for users such as content creators who wish to steer generation with descriptive instructions. To address these constraints, we introduce OV-InstructTTS, a new paradigm for open-vocabulary InstructTTS. We propose a comprehensive solution comprising a newly curated dataset, OV-Speech, and a novel reasoning-driven framework. The OV-Speech dataset pairs speech with open-vocabulary instructions, each augmented with a reasoning process that connects high-level instructions to acoustic features. The reasoning-driven framework infers emotional, acoustic, and paralinguistic information from open-vocabulary instructions before synthesizing speech. Evaluations show that this reasoning-driven approach significantly improves instruction-following fidelity and speech expressiveness. We believe this work can inspire the next user-friendly InstructTTS systems with stronger generalization and real-world applicability. The dataset and demos are publicly available on our project page.

cs.SD↗

arXiv2025

GaussianDWM: 3D Gaussian Driving World Model for Unified Scene Understanding and Multi-Modal Generation

Driving World Models (DWMs) have been developing rapidly with the advances of generative models. However, existing DWMs lack 3D scene understanding capabilities and can only generate content conditioned on input data, without the ability to interpret or reason about the driving environment. Moreover, current approaches represent 3D spatial information with point cloud or BEV features do not accurately align textual information with the underlying 3D scene. To address these limitations, we propose a novel unified DWM framework based on 3D Gaussian scene representation, which enables both 3D scene understanding and multi-modal scene generation, while also enabling contextual enrichment for understanding and generation tasks. Our approach directly aligns textual information with the 3D scene by embedding rich linguistic features into each Gaussian primitive, thereby achieving early modality alignment. In addition, we design a novel task-aware language-guided sampling strategy that removes redundant 3D Gaussians and injects accurate and compact 3D tokens into LLM. Furthermore, we design a dual-condition multi-modal generation model, where the information captured by our vision-language model is leveraged as a high-level language condition in combination with a low-level image condition, jointly guiding the multi-modal generation process. We conduct comprehensive studies on the nuScenes, and NuInteract datasets to validate the effectiveness of our framework. Our method achieves state-of-the-art performance. We will release the code publicly on GitHub https://github.com/dtc111111/GaussianDWM.

cs.CV↗

arXiv2025

OmniDrive-R1: Reinforcement-driven Interleaved Multi-modal Chain-of-Thought for Trustworthy Vision-Language Autonomous Driving

The deployment of Vision-Language Models (VLMs) in safety-critical domains like autonomous driving (AD) is critically hindered by reliability failures, most notably object hallucination. This failure stems from their reliance on ungrounded, text-based Chain-of-Thought (CoT) reasoning. While existing multi-modal CoT approaches attempt mitigation, they suffer from two fundamental flaws: (1) decoupled perception and reasoning stages that prevent end-to-end joint optimization, and (2) reliance on expensive, dense localization labels. Thus we introduce OmniDrive-R1, an end-to-end VLM framework designed for autonomous driving, which unifies perception and reasoning through an interleaved Multi-modal Chain-of-Thought (iMCoT) mechanism. Our core innovation is an Reinforcement-driven visual grounding capability, enabling the model to autonomously direct its attention and "zoom in" on critical regions for fine-grained analysis. This capability is enabled by our pure two-stage reinforcement learning training pipeline and Clip-GRPO algorithm. Crucially, Clip-GRPO introduces an annotation-free, process-based grounding reward. This reward not only eliminates the need for dense labels but also circumvents the instability of external tool calls by enforcing real-time cross-modal consistency between the visual focus and the textual reasoning. Extensive experiments on DriveLMM-o1 demonstrate our model's significant improvements. Compared to the baseline Qwen2.5VL-7B, OmniDrive-R1 improves the overall reasoning score from 51.77% to 80.35%, and the final answer accuracy from 37.81% to 73.62%.

cs.CV↗

arXiv2025

Enhancing Robustness of Offline Reinforcement Learning Under Data Corruption via Sharpness-Aware Minimization

Offline reinforcement learning (RL) is vulnerable to real-world data corruption, with even robust algorithms failing under challenging observation and mixture corruptions. We posit this failure stems from data corruption creating sharp minima in the loss landscape, leading to poor generalization. To address this, we are the first to apply Sharpness-Aware Minimization (SAM) as a general-purpose, plug-and-play optimizer for offline RL. SAM seeks flatter minima, guiding models to more robust parameter regions. We integrate SAM into strong baselines for data corruption: IQL, a top-performing offline RL algorithm in this setting, and RIQL, an algorithm designed specifically for data-corruption robustness. We evaluate them on D4RL benchmarks with both random and adversarial corruption. Our SAM-enhanced methods consistently and significantly outperform the original baselines. Visualizations of the reward surface confirm that SAM finds smoother solutions, providing strong evidence for its effectiveness in improving the robustness of offline RL agents.

cs.LG↗

arXiv2025

SAM2-3dMed: Empowering SAM2 for 3D Medical Image Segmentation

Accurate segmentation of 3D medical images is critical for clinical applications like disease assessment and treatment planning. While the Segment Anything Model 2 (SAM2) has shown remarkable success in video object segmentation by leveraging temporal cues, its direct application to 3D medical images faces two fundamental domain gaps: 1) the bidirectional anatomical continuity between slices contrasts sharply with the unidirectional temporal flow in videos, and 2) precise boundary delineation, crucial for morphological analysis, is often underexplored in video tasks. To bridge these gaps, we propose SAM2-3dMed, an adaptation of SAM2 for 3D medical imaging. Our framework introduces two key innovations: 1) a Slice Relative Position Prediction (SRPP) module explicitly models bidirectional inter-slice dependencies by guiding SAM2 to predict the relative positions of different slices in a self-supervised manner; 2) a Boundary Detection (BD) module enhances segmentation accuracy along critical organ and tissue boundaries. Extensive experiments on three diverse medical datasets (the Lung, Spleen, and Pancreas in the Medical Segmentation Decathlon (MSD) dataset) demonstrate that SAM2-3dMed significantly outperforms state-of-the-art methods, achieving superior performance in segmentation overlap and boundary precision. Our approach not only advances 3D medical image segmentation performance but also offers a general paradigm for adapting video-centric foundation models to spatial volumetric data.

eess.IV↗

arXiv2025

PolyServe: Efficient Multi-SLO Serving at Scale

Advances in Large Language Models (LLMs) have led to a surge of LLM-powered applications. These applications have diverse token-generation latency requirements. As a result, simply classifying workloads as latency-sensitive (LS) or best-effort (BE) overlooks the nuances within the latency-sensitive category and results in suboptimal user experiences and scheduling opportunities. However, efficiently serving requests with multiple SLO requirements poses significant challenges. First, all requests within a batch generate new tokens simultaneously, which can misalign them with their distinct SLO requirements. Moreover, while existing systems focus on auto-scaling for handling various overall request rates, the diversity of SLOs necessitates fine-grained auto-scaling among these SLO tiers. Finally, unlike LS/BE scenarios, where BE requests can be aborted at any time to ensure the SLO attainment of LS requests, those with different latency-sensitive SLOs cannot tolerate prolonged delays, and tail latency must be controlled. To tackle these challenges, we propose PolyServe, a novel multi-SLO scheduling policy at scale that maintains high SLO attainment while maximizing throughput. PolyServe first groups requests into multiple bins based on their per-token latency requirement, then schedules each bin to a subset of the server fleet. PolyServe routes requests to the highest-load but still SLO-attainable server to create a load gradient that facilitates auto-scaling. To increase utilization, PolyServe permits looser-SLO requests to share tighter-SLO instances when their own servers are saturated. PolyServe uses profiling data to guide scheduling decisions and manage tail latency through request-wait-time-aware scheduling, dynamic chunking, and continuous chunked prefill prediction. PolyServe achieves 1.23x goodput gain compared to existing policies, achieving up to 92.5% of optimal goodput.

cs.DC↗

arXiv2025

Perception Activator: An intuitive and portable framework for brain cognitive exploration

Recent advances in brain-vision decoding have driven significant progress, reconstructing with high fidelity perceived visual stimuli from neural activity, e.g., functional magnetic resonance imaging (fMRI), in the human visual cortex. Most existing methods decode the brain signal using a two-level strategy, i.e., pixel-level and semantic-level. However, these methods rely heavily on low-level pixel alignment yet lack sufficient and fine-grained semantic alignment, resulting in obvious reconstruction distortions of multiple semantic objects. To better understand the brain's visual perception patterns and how current decoding models process semantic objects, we have developed an experimental framework that uses fMRI representations as intervention conditions. By injecting these representations into multi-scale image features via cross-attention, we compare both downstream performance and intermediate feature changes on object detection and instance segmentation tasks with and without fMRI information. Our results demonstrate that incorporating fMRI signals enhances the accuracy of downstream detection and segmentation, confirming that fMRI contains rich multi-object semantic cues and coarse spatial localization information-elements that current models have yet to fully exploit or integrate.

cs.CV↗

arXiv2025

Mitigating Audiovisual Mismatch in Visual-Guide Audio Captioning

Current vision-guided audio captioning systems frequently fail to address audiovisual misalignment in real-world scenarios, such as dubbed content or off-screen sounds. To bridge this critical gap, we present an entropy-aware gated fusion framework that dynamically modulates visual information flow through cross-modal uncertainty quantification. Our novel approach employs attention entropy analysis in cross-attention layers to automatically identify and suppress misleading visual cues during modal fusion. Complementing this architecture, we develop a batch-wise audiovisual shuffling technique that generates synthetic mismatched training pairs, greatly enhancing model resilience against alignment noise. Evaluations on the AudioCaps benchmark demonstrate our system's superior performance over existing baselines, especially in mismatched modality scenarios. Furthermore, our solution demonstrates an approximately 6x improvement in inference speed compared to the baseline.

cs.MM↗