arXiv subjects

Zhibin Wang

Publications and source records attributed to Zhibin Wang.

At least 19 recordsLinked to original sources

arXiv2026

LOCAL: Enabling Learning On-device Contiguously for Agent LLMs

On-device LLM agents interact repeatedly with users on local hardware, producing private traces that are valuable for adaptation but should not be sent to a remote trainer. Ideally, such agents would learn contiguously---adapting from every interaction without pausing or suspending user-facing inference---yet existing inference runtimes assume stable weights and existing RL systems assume separated resources, so neither can support this continuity. We present LOCAL, the first single-GPU runtime that enables contiguous on-device learning for LLM agents. The key insight is that GPU scheduling, adapter version management, and KV-cache validity cannot be handled by independent subsystems: adapter updates invalidate cached KV tensors from older versions, and cache retention affects the memory available for training. LOCAL makes adapter version, task priority, and cache state visible to three cooperating components---a cooperative scheduler, a version-aware KV-cache manager, and a multi-agent model runtime---that share this state to keep scheduling, execution, and cache maintenance mutually consistent. On a single 24 GB GPU with 7B-class models, LOCAL lowers foreground queue-wait p95 by 3.1x over FIFO, lowers p95 time-to-first-token (TTFT) by 1.55x versus non-preemptible training, cuts post-publish first-hit prefill p99 by 25.6% and cross-agent TTFT p99 by 21.9%, and keeps background learning progressing under tight KV budgets.

cs.DC↗

arXiv2026

Model the Edit, Not the Image: Visual Autoregressive Editing from a Source-Centric Perspective

Next-scale visual autoregressive models (VARs) have emerged as a powerful generative paradigm, producing high-quality images through efficient coarse-to-fine prediction. However, their potential for text-guided image editing remains largely underexplored. Existing training-free VAR editing approaches often formulate editing as target-conditioned regeneration guided or constrained by the source image, and may rely on inversion, test-time optimization, attention control, or user-provided masks. This generation-centric formulation does not fully exploit the multiscale source representations provided by VARs and may introduce additional computation or intervention. We instead take a source-centric perspective on VAR editing, in which the encoded source image tokens serve as the primary visual state and the editing process focuses on condition-induced changes. Based on this perspective, we propose \textbf{EditMod}, which compares source- and target-conditioned predictions under a shared autoregressive context, treats their difference as a scale-wise editing direction, and applies it as a residual update to source tokens at selected scales. Experiments show that EditMod achieves leading source-image fidelity while maintaining strong text alignment, and completes end-to-end editing of a 1K image in only 1.57 seconds on a single A100 GPU without per-image preparation.

cs.CV↗

arXiv2026

ComBodied Agents: a New Paradigm of Human-Centric Agentic AI

After an older adult misses a medication dose, a software agent can send another reminder and an embodied agent can bring the medication. Yet neither explains whether the person forgot, is confused, has side effects, or deliberately refused, nor what support is appropriate. This reveals a structural gap in Agentic AI: Digital Agents primarily transform software states, while Embodied Agents transform physical states; neither makes a person's evolving state and agency the primary object of modeling, intervention, and evaluation. We introduce Combodied Agents, a human-centered paradigm that perceives, models, predicts, and supports individual human-state trajectories over time, using software tools, sensors, wearables, robots, and human services as action channels rather than end goals. We unify fragmented capabilities across personal assistants, health agents, AI companions, and adaptive human--AI systems into a closed loop: event-based multimodal perception reconstructs meaningful personal events; longitudinal, correctable memory provides temporal context; Personal World Models estimate future personal states and outcomes under alternative decisions and interventions; and an admissible intervention policy selects proportionate support under consent, uncertainty, safety, reversibility, and user control. Feedback from the person and environment updates the loop. Rather than requiring an exhaustive Human Digital Twin, the framework uses purpose-bounded, uncertainty-aware, user-correctable representations. We organize the design space by human-state targets, relational contexts, and agent roles, and propose scenario-centered evaluation, agency-preservation metrics, benchmark requirements, edge-native personal models, and governance directions. Combodied Agents shift Agentic AI from external task completion toward sustained human benefit.

cs.AI↗

arXiv2026

Adaptive Matrix Multiplication for Dynamic Shapes on Ascend NPUs

Matrix Multiplication (MatMul) faces a "generalization crisis" driven by highly dynamic tensor shapes. This crisis is particularly acute on Ascend NPUs, where explicitly controlled architectures and strict physical constraints render existing GPU-centric optimizations ineffective. To resolve this, we propose AdaptCore, an adaptive framework for universally high-performance MatMul on Ascend NPUs. AdaptCore systematically decouples operator optimization into spatial tiling and instruction orchestration. It first maps dynamic shapes into a hardware-aware 2D tiling taxonomy to balance on-chip capacity limits and multi-core parallelism. Furthermore, it integrates a composable optimization library with a deterministic analytical performance model. By mathematically evaluating hardware state mutations, AdaptCore proactively selects and caches optimal implementations, enabling O(1) overhead runtime dispatching. Evaluations demonstrate that AdaptCore delivers a remarkable 1.85x mean speedup across 80,000 input shapes, and achieves up to a 1.48x acceleration in representative end-to-end models over the highly-tuned native vendor library (ACLNN).

cs.AR↗

arXiv2026

MoE Proxy Models for Low-Cost Failure Reproduction and Diagnosis in LLM RL Post-Training

Reinforcement learning (RL) post-training of large language models (LLMs) is computationally intensive and involves complex system pipelines with substantial debugging overhead. In practice, factors such as framework adaptation, numerical precision, and operator implementation can cause failures, including gradient overflow and loss divergence. Reproducing such failures directly on large models requires considerable time and computational resources. This paper systematically analyzes failures encountered during large-scale RL training on the Huawei Ascend platform, summarizes representative failure types, and identifies three model-side factors relevant to fault reproduction. Based on these factors, we propose a proxy-model construction method for low-cost fault investigation and auxiliary diagnosis. It employs structure-preserving, clustering-based expert pruning to select representative experts while retaining the model's backbone architecture, routing mechanism, and basic task capabilities. Our experimental results show that the proxy models reduce accelerator requirements by 50%-87.5% and achieve up to a 33.3x reduction in per-step NPU-hour cost, while preserving major training dynamics and reproducing fault responses consistent with the original models. Overall, the proxy models can serve as low-cost surrogates for fault reproduction, targeted validation, and auxiliary diagnosis in RL post-training.

cs.LG↗

arXiv2026

Scheduling Mixed RL Rollouts Beyond Prefix Locality

Modern reinforcement learning (RL) post-training pipelines for large language models (LLMs) increasingly combine rollout workloads across multiple domains and feedback paradigms. Prefix-aware routing improves inference efficiency through cache reuse and load balancing, but it does not control how heterogeneous rollout sessions compete for KV-cache capacity. When reinforcement learning with verifiable rewards (RLVR), reinforcement learning from human feedback (RLHF), and agentic rollouts share an asynchronous inference service, their distinct sequence structures, interaction patterns, and KV-residency times create substantially different serving demands. Rollout scheduling must account for this heterogeneity without distorting the workload mixture specified by the trainer. We present MISA-T, a routing-layer admission policy for mixed rollout serving. MISA-T combines adaptive session admission, workload-aware KV-capacity allocation, and residency-time-aware KV accounting. In rollout-only ablations on Step3.7 and Qwen3.6-35B-A3B, MISA-T improves rollout throughput over a sweep-tuned cache-aware vLLM Router by 53.3% and 43.6%, respectively, while maintaining high prefix-cache hit rates. In a matched 50-iteration Step3.7 experiment, it increases rollout throughput by 35.6% and reduces mean iteration time by 22.8%, while keeping the consumed workload mixture close to the trainer target and achieving comparable task scores.

cs.DC↗

arXiv2026

TIDE-MC: Two-Sided Interpolative Decomposition for Billion-Scale GPU Matrix Completion

Matrix completion supports large-scale recommendation and scientific computing, yet existing GPU solvers commonly assume that the observed matrix or its dense factors fit in device memory. On real workloads, this assumption leads to out-of-memory failures or severe PCIe overhead under naive paging. We present TIDE-MC, a bounded-memory GPU framework built on Two-Sided Interpolative Decomposition (TSID). TSID uses a sampled template submatrix as an anchor for reconstructing the full low-rank matrix, allowing computation and storage to scale with the template and active data chunks rather than the complete matrix. TIDE-MC realizes this formulation through two execution stages. First, a conflict-free synchronization engine recovers the template using parallel factorization and hierarchical gradient aggregation. Second, a chunked reconstruction pipeline extends the recovered template to the remaining matrix while overlapping PCIe transfers with GPU computation. An asymmetric gradient-clipping scheme stabilizes mixed-precision Tensor Core execution. Across 15 benchmarks, TIDE-MC completes workloads that cause existing GPU solvers to run out of memory. Compared with the evaluated state-of-the-art baselines, it achieves up to 11,647x speedup, reduces peak memory usage by up to 8.5x, and lowers reconstruction error by up to 99.7%. These results show that template-anchored decomposition and stage-specific GPU execution can scale matrix completion beyond device-memory capacity.

cs.DC↗

arXiv2026

SPAE: Spectrally Guided Autoencoder for Pretrained Visual Latents

Latents from vision foundation models (VFMs) are semantically rich and well suited for visual understanding. Recent representation autoencoder methods such as RAE have shown that they can provide promising latent spaces for image generation. However, VFM latents remain difficult to model directly: DiT-generated latents exhibit spectral mismatch with encoder latents, especially in high-frequency components. Our channel-wise spectral analysis further reveals that these high-frequency components are diffusely distributed across latent channels and entangled with semantic information, making the latent space difficult for DiT to model. To address these challenges, we propose SPAE, latent adaptation framework for generation. Specifically, SPAE employs a compact bottleneck to distill stable semantic information while suppressing high-frequency components, thereby improving the alignment between DiT-generated latents and encoder latents. In addition, we apply a channel-wise masking strategy to promote the decoupling of semantic information and high-frequency details across bottleneck channels. Experiments show that SPAE achieves a favorable balance among visual understanding, generation quality, and reconstruction fidelity.

cs.CV↗

arXiv2026

SpecLA: Efficient Speculative Decoding for Linear-Attention Models

Linear-attention models replace the growing KV cache with recurrent states, but autoregressive decoding still reads, updates, and writes these states one token at a time. Speculative decoding can reduce this cost by verifying several draft tokens in one target pass, yet existing speculative systems are designed for Transformer KV caches. For stateful linear-attention targets, verification must follow recurrent dependencies across chains and branches, acceptance must update only the accepted state trajectory, and the drafter must avoid submitting candidates that waste stateful verification work. This paper presents SpecLA, a speculative decoding runtime for stateful linear-attention models. SpecLA verifies chains and trees with topology-aware kernels, stores compact factors produced during verification to recover accepted states, and uses confidence pruning plus a target-aligned EAGLE-style drafter to feed useful candidates to the verifier. On an NVIDIA H100 with a public GDN-1.3B target, SpecLA achieves up to 1.70x end-to-end speedup over autoregressive decoding.

cs.CL↗

arXiv2026

SmartSwap: Swap-Based Memory Optimization for LLM Training under Varying Operator Sequences

The increasing size of large language models (LLMs) has led to a surge in memory requirements during training, often exceeding the capacity of high-bandwidth memory (HBM). Swap-based memory optimization incurs neither accuracy loss nor additional end-to-end overhead when effectively overlapped, thus being an attractive solution. However, existing swap methods assume consistent operator sequences, which is impractical in Eager Mode, where operator sequences can vary during change. We propose Chameleon, which redesigns the end-to-end process of swap-based memory optimization and is the first work to consider varying operator sequences in Eager Mode. Chameleon (i) introduces a lightweight online profiler to enable continuous profiling for monitoring operator sequences, (ii) generates effective swap policies with limited operator information, and (iii) optimizes the policy execution module for accurate policy application and better performance. Experimental results demonstrate that Chameleon reduces profiling overhead by 84.25%, enables training models up to 4x larger than hardware memory while adapting to changes in operator sequences, improves performance by up to 38.94% compared to recomputation or high-degree parallelism.

cs.DC↗

arXiv2026

SSV: Sparse Speculative Verification for Efficient LLM Inference

Speculative decoding and dynamic sparse attention are two complementary approaches for accelerating long-context LLM inference: the former amortizes target-model execution across multiple verifier queries, while the latter reduces each query's KV-cache working set. Directly combining them, however, exposes a structural mismatch: speculative verification relies on cross-query commonality, whereas dynamic sparse attention assigns query-specific sparse layouts. This mismatch limits KV-block reuse, amplifies NSA's branch-wise overheads, and makes verification strategy selection input- and regime-dependent. We present SSV, a sparse speculative-verification framework that turns dynamic sparse attention into a verification-oriented workload. SSV combines overlap-aware grouped-query execution, refresh/reuse-based NSA kernel fusion, and profile-guided prompt-adaptive orchestration to improve cross-query reuse, reduce selected-index and branch-fusion overheads, and select effective draft-verification strategies under user-specified precision classes. Experiments on NVIDIA H100 GPUs show that SSV achieves up to 3.49x end-to-end throughput over autoregressive NSA decoding and up to 6.86x kernel speedups for sparse speculative verification.

cs.OS↗

arXiv2026

LEAPS: An LLM-Empowered Adaptive Plugin in Taobao AI Search

The rapid rise of large language models has shifted user search behavior from discrete keywords to natural-language, multi-constraint queries--a shift existing e-commerce search architectures struggle to accommodate. Users face a dilemma: precise natural-language queries often trigger zero-result scenarios, while forced simplification yields noisy, generic results that overwhelm decision-making. To address this, we propose LEAPS (LLM-Empowered Adaptive Plugin in Taobao AI Search), which upgrades traditional search pipelines via a "Broaden-and-Refine" paradigm by attaching plugins at both ends. (1) Upstream, a Query Expander generates adaptive, complementary query combinations to maximize the candidate set, trained via a three-stage strategy of inverse data augmentation, posterior-knowledge supervised fine-tuning, and diversity-aware reinforcement learning. (2) Downstream, a Relevance Verifier filters noise by synthesizing multi-source signals (e.g., OCR text, reviews) with chain-of-thought reasoning. Extensive offline experiments and online A/B testing show that LEAPS significantly enhances the conversational shopping experience, while its non-intrusive architecture preserves established short-text retrieval performance and enables low-cost integration with diverse back-ends. Fully deployed on Taobao AI Search since August 2025, LEAPS serves hundreds of millions of users monthly.

cs.IR↗

arXiv2026

STAR: Decode-Phase Rescheduling for LLM Inference

Large Language Model (LLM) inference has emerged as a fundamental paradigm, however, 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 STAR, a decode 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 75.1% and achieving 2.63 times higher goodput.

cs.DC↗

arXiv2026

DDA-Thinker: Decoupled Dual-Atomic Reinforcement Learning for Reasoning-Driven Image Editing

Recent image editing models have achieved strong visual fidelity but often struggle with tasks requiring complex reasoning. To investigate and enhance the reasoning-grounded planning for image editing, we propose DDA-Thinker, a Thinker-centric framework designed for the independent optimization of a planning module (Thinker) over a fixed generative model (Editor). This decoupled Thinker-centric paradigm facilitates a controlled analysis of the planning module and makes its contribution under a fixed Editor easier to assess. To effectively guide this Thinker, we introduce a dual-atomic reinforcement learning framework. This framework decomposes feedback into two distinct atomic rewards implemented through verifiable checklists: a cognitive-atomic reward to directly assess the quality of the Thinker's executable plan, which serves as the actionable outcome of the Thinker's reasoning, and a visual-atomic reward to assess the final image quality. To improve checklist quality, our checklist synthesis is grounded not only in the source image and user instruction but also in a rational reference description of the ideal post-edit scene. To support this training, we further develop a two-stage data curation pipeline that first synthesizes a diverse and reasoning-focused dataset, then applies difficulty-aware refinement to curate an effective training curriculum for reinforcement learning. Extensive experiments on reasoning-driven image editing benchmarks, including RISE-Bench and KRIS-Bench, demonstrate that our approach substantially improves overall performance. Our method enables a community model to achieve results competitive with strong proprietary models, highlighting the practical potential of Thinker-centric optimization under a fixed-editor setting.

cs.CV↗

arXiv2026

Stability and Decay for the 2D Anisotropic Navier-Stokes Equations with Fractional Horizontal Dissipation on $\mathbb{R}^2$

The stability problem for the 2D Navier-Stokes equations with dissipation in only one direction on $\mathbb R^2$ is not fully understood. This dissipation is in the intermediate regime between the fully dissipative Navier-Stokes and the inviscid Euler. Navier-Stokes solutions in $\mathbb R^2$ decay algebraically in time while Euler solutions can grow rather rapidly in time. This paper solves the fundamental stability and large-time behavior problem on the anisotropic Navier-Stokes with fractional dissipation $Λ_1^{2s}$ for all $0\leq s<1$. The case $s=1$ corresponds to the standard one directional dissipation $\partial_1^2$. Different techniques are developed to treat different ranges of fractional exponents: $0\leq s\leq \frac34$, $\frac34<s<\frac{11}{12}$, and $\frac{11}{12} \leq s <1$. The final range is the most difficult case, for which we introduce the spatial polynomial $A_2$ weights and exploit the boundedness of Riesz transforms on weighted $L^2$-spaces.

math.AP↗

arXiv2026

Chameleon: Adaptive Fault Tolerance for Distributed Training via Real-time Policy Selection

Training large language models faces frequent interruptions due to various faults, demanding robust fault-tolerance. Existing backup-free methods, such as redundant computation, dynamic parallelism, and data rerouting, each incur performance penalties, whether from ongoing overhead, lengthy reconfigurations, or post-recovery inefficiencies. We propose Chameleon, an adaptive fault-tolerant system that intelligently selects optimal recovery strategies when a failure occurs. Chameleon achieves this through a unified performance model, expedient execution plan search, accurate performance estimation, and efficient communication optimizations. Experiments on a 32-card cluster show that Chameleon maintains a performance gap of within 11.00% between post-recovery and failure-free training, while preserving model convergence and efficient memory usage. Compared to state-of-the-art methods, Chameleon achieves up to 1.229x and 1.355x higher average throughput than Oobleck and Recycle, respectively.

cs.DC↗

arXiv2026

AdaSpark: Adaptive Sparsity for Efficient Long-Video Understanding

Processing long-form videos with Video Large Language Models (Video-LLMs) is computationally prohibitive. Current efficiency methods often compromise fine-grained perception through irreversible information disposal or inhibit long-range temporal modeling via rigid, predefined sparse patterns. This paper introduces AdaSpark, an adaptive sparsity framework designed to address these limitations. AdaSpark first partitions video inputs into 3D spatio-temporal cubes. It then employs two co-designed, context-aware components: (1) Adaptive Cube-Selective Attention (AdaS-Attn), which adaptively selects a subset of relevant video cubes to attend for each query token, and (2) Adaptive Token-Selective FFN (AdaS-FFN), which selectively processes only the most salient tokens within each cube. An entropy-based (Top-p) selection mechanism adaptively allocates computational resources based on input complexity. Experiments demonstrate that AdaSpark significantly reduces computational load by up to 57% FLOPs while maintaining comparable performance to dense models and preserving fine-grained, long-range dependencies, as validated on challenging hour-scale video benchmarks.

cs.CV↗

arXiv2026

CoDec: Prefix-Shared Decoding Kernel for LLMs

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 CoDec. CoDec 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 CoDec achieves an average $1.9\times$ speedup and $120.9\times$ memory access reduction compared to the state-of-the-art FlashDecoding kernel regarding attention computation in the decode stage and $3.8\times$ end-to-end time per output token compared to the vLLM.

cs.LG↗