Searcharxiv⌕ Search

arXiv subjects

Zhaode Wang

Publications and source records attributed to Zhaode Wang.

14 recordsLinked to original sources

EngineAgent: Agentizing On-Device LLM Inference Engines with Inference Stack Hierarchy Awareness and Hardware Abilities

Rapidly evolving LLM architectures and heterogeneous mobile hardware make supporting new models in on-device inference engines a recurring and costly system challenge. Each model may require coordinated changes across conversion, graph construction, operators, and hardware backends. Existing coding agents often miss these cross-layer dependencies and struggle to diagnose failures across host--device execution paths. We present EngineAgent, an agentic system for automating on-device LLM model enablement. Its key insight is to separate reusable engine and hardware knowledge from model-specific implementation. Before enablement, EngineAgent constructs a hierarchy-aware engine index that captures engine abstraction layers and cross-layer dependencies, and pre-validates backend-specific build, execution, and debugging workflows using already supported models. For each new model, a builder uses this knowledge to coordinate cross-layer changes, while an auditor checks implementation completeness and drives iterative debugging through device-side validation. Across 12 tasks spanning three mobile LLMs and four engine--backend pairs on CPUs, GPUs, and NPUs, EngineAgent achieves 75% end-to-end success with either of two underlying LLMs, versus 33.3% for the strongest coding-agent baseline. Success requires compilation, on-device execution, and output-logit agreement with reference implementations. EngineAgent also reduces average per-task time and model-API cost relative to baselines using the same underlying LLMs.

cs.OS↗

Sibyl: An Efficient Small-large Model Collaboration Framework for Long-horizon Tasks

Small language models (SLMs) offer a promising foundation for on-device agents through low-latency, resource-efficient inference, yet limited reasoning and planning capabilities constrain their performance on long-horizon tasks requiring multi-step interaction with the environment. Step-level collaboration between SLMs and larger cloud-hosted models can bridge this gap, but identifying states that warrant cloud assistance remains challenging: the contribution of each cloud call is entangled with subsequent actions and can be assessed only from the final task outcome. Compounding this challenge, the SLM must balance two competing objectives: maximizing task success and minimizing cloud calls. To address this, we propose Sibyl, an algorithm that trains SLM agents to selectively consult cloud models at the step level and internalize their guidance for subsequent decisions, achieving strong task performance with minimal cloud reliance. Sibyl follows a three-stage training pipeline that (1) builds a robust base policy through consultation-free self-evolving reinforcement learning (RL); (2) cold-starts consultation behavior via decisive-disagreement state mining; and (3) jointly optimizes consultation decisions and guidance internalization through consultation-aware RL. Experiments on ALFWorld and WebShop demonstrate that Sibyl, using only a 0.6B-parameter model, outperforms state-of-the-art baselines, including agent training and routing methods, by 95.2% and 80.4% in success rate while averaging only 0.8 and 3.9 cloud calls per trajectory, respectively.

cs.AI↗

Dynamic Flow, Static Graph: KV Cache Reuse for Efficient LLM Serving on Mobile NPUs

On-device large language model (LLM) serving is a cornerstone of local-first personal intelligence, offering users data sovereignty, strong privacy guarantees, and freedom from cloud API latency and cost. Although KV caching is widely used to reduce latency in long-context inference, existing designs were primarily optimized for cloud GPUs with dynamic execution environments and abundant memory bandwidth. These architectural assumptions do not hold on mobile NPUs, where computation graphs must be statically compiled and both memory capacity and I/O bandwidth are severely constrained. In this work, we present a compute-storage co-design for mobile-centric prefix and non-prefix KV reuse. We first propose an intra-graph mechanism that maps selective KV recomputation onto static NPU graphs, reconciling algorithmic dynamicity with NPU staticity. We further develop an inter-graph scheduler to optimize chunk merging and minimize padding with dynamic programming. To address mobile bandwidth limitations, we introduce a hierarchical KV manager featuring a tree-hash-semantic hybrid structure, along with cost-aware prefetching and eviction policies. We also build a two-dimensional pipeline that overlaps KV loading, rerotation, and storage with NPU execution, hiding data-movement latency. Experiments across representative on-device workloads and LLMs show that our design reduces time-to-first-token (TTFT) by $40-60\%$ compared with no reuse and prefix-only caching.

cs.OS↗

RecGPT-Mobile-V2 Technical Report

Personalized Query prediction maps implicit behavioral signals---clicks, favorites, purchases, and post-purchase exploration---to explicit retrieval intent. On-device deployment makes this task particularly challenging: behavioral trajectories are noisy and multi-scale, multiple Queries may be valid for a single trajectory, and a uniform reasoning policy either expends unnecessary computation on simple instances or allocates insufficient capacity to complex ones. We introduce RecGPT-Mobile-V2, an end-to-end framework that treats intent quality and execution efficiency as coupled objectives within a staged design. The framework transforms heterogeneous interactions into an evidence-preserving trajectory, establishes a recommendation-native foundation through domain adaptation and supervised alignment, and applies reasoning-cost optimization only after grouped rollouts meet grounding and utility criteria. The resulting teacher is distilled into a compact student deployed with low-bit execution, structured compression, and budget-aware device--cloud routing. In an aligned CoT ablation, an evidence-focused short rationale increases ROUGE-L from 0.228 to 0.315 and Jaccard from 0.174 to 0.248, while slightly outperforming the full five-stage rationale. In the controlled RL comparison, the complete reward formulation improves Query quality from 73.2% under quality-only RL to 78.6%, lowers the hard-failure rate from 3.6% to 1.6%, and reduces the median CoT length from 62 to 14 tokens. Online retrieval analysis further indicates that the Query recall channel retrieves inventory complementary to that surfaced by established recall channels. Collectively, these findings support sufficiency-oriented rather than uniformly short reasoning: retain decision-relevant evidence and allocate additional computation only when it is likely to improve the predicted Query.

cs.IR↗

RecGPT-Mobile: On-Device Large Language Models for User Intent Understanding in Taobao Feed Recommendation

Predicting a user's next search query from recent interaction behaviors is a critical problem in modern e-commerce systems, particularly in scenarios where user intent evolves rapidly. Large Language Models (LLMs) offer strong semantic reasoning capabilities and have recently been adopted to enhance training data construction for next-query prediction. However, due to resource constraints on mobile devices, existing applications are deployed on cloud servers, resulting in high inference costs. In this paper, we propose RecGPT-Mobile, a framework that designs a lightweight LLM-based intent understanding agent to improve recommendation quality in mobile e-commerce scenarios. By deploying LLMs directly on mobile devices, our approach can capture evolving interests of users more quickly and adjust the recommendation results in real time. Extensive offline analyses and online experiments demonstrate that our method significantly improves the accuracy of recommendation results, laying a practical path for LLM deployment in production-scale recommendation systems on mobile devices, as well as a scalable solution for integrating LLMs into real-world next-query prediction systems.

cs.IR↗

RetentiveKV: State-Space Memory for Uncertainty-Aware Multimodal KV Cache Eviction

Multimodal Large Language Models face severe challenges in computational efficiency and memory consumption due to the substantial expansion of the visual KV cache when processing long visual contexts. Existing KV cache compression methods typically rely on the "persistence of importance" hypothesis to prune tokens. However, this approach proves fragile in multimodal settings due to two key issues: 1) Visual tokens display "deferred importance," initially exhibiting low salience but becoming pivotal during later decoding, which can lead to premature eviction. 2) Discrete pruning disrupts the inherent spatial continuity of visual cues. To address these challenges, we propose RetentiveKV, an entropy-driven KV cache optimization method that reformulates KV eviction from "discrete context truncation" to "continuous memory evolution" based on State Space Models. Our method leverages information entropy to quantify the information potential of low-attention tokens and integrates tokens scheduled for eviction into a continuous state space through entropy-guided state transitions, enabling their dynamic reactivation when semantic relevance arises during subsequent decoding. Extensive experiments on multimodal benchmarks demonstrate that RetentiveKV achieves 5.0 times KV cache compression and 1.5 times decoding acceleration.

cs.LG↗

MobileKernelBench: Can LLMs Write Efficient Kernels for Mobile Devices?

Large language models (LLMs) have demonstrated remarkable capabilities in code generation, yet their potential for generating kernels specifically for mobile devices remains largely unexplored. In this work, we extend the scope of automated kernel generation to the mobile domain to investigate the central question: Can LLMs write efficient kernels for mobile devices? To enable systematic investigation, we introduce MobileKernelBench, a comprehensive evaluation framework comprising a benchmark prioritizing operator diversity and cross-framework interoperability, coupled with an automated pipeline that bridges the host-device gap for on-device verification. Leveraging this framework, we conduct extensive evaluation on the CPU backend of Mobile Neural Network (MNN), revealing that current LLMs struggle with the engineering complexity and data scarcity inherent to mobile frameworks; standard models and even fine-tuned variants exhibit high compilation failure rates (over 54%) and negligible performance gains due to hallucinations and a lack of domain-specific grounding. To overcome these limitations, we propose the Mobile Kernel Agent (MoKA), a multi-agent system equipped with repository-aware reasoning and a plan-and-execute paradigm. Validated on MobileKernelBench, MoKA achieves state-of-the-art performance, boosting compilation success to 93.7% and enabling 27.4% of generated kernels to deliver measurable speedups over native libraries.

cs.LG↗

AccKV: Towards Efficient Audio-Video LLMs Inference via Adaptive-Focusing and Cross-Calibration KV Cache Optimization

Recent advancements in Audio-Video Large Language Models (AV-LLMs) have enhanced their capabilities in tasks like audio-visual question answering and multimodal dialog systems. Video and audio introduce an extended temporal dimension, resulting in a larger key-value (KV) cache compared to static image embedding. A naive optimization strategy is to selectively focus on and retain KV caches of audio or video based on task. However, in the experiment, we observed that the attention of AV-LLMs to various modalities in the high layers is not strictly dependent on the task. In higher layers, the attention of AV-LLMs shifts more towards the video modality. In addition, we also found that directly integrating temporal KV of audio and spatial-temporal KV of video may lead to information confusion and significant performance degradation of AV-LLMs. If audio and video are processed indiscriminately, it may also lead to excessive compression or reservation of a certain modality, thereby disrupting the alignment between modalities. To address these challenges, we propose AccKV, an Adaptive-Focusing and Cross-Calibration KV cache optimization framework designed specifically for efficient AV-LLMs inference. Our method is based on layer adaptive focusing technology, selectively focusing on key modalities according to the characteristics of different layers, and enhances the recognition of heavy hitter tokens through attention redistribution. In addition, we propose a Cross-Calibration technique that first integrates inefficient KV caches within the audio and video modalities, and then aligns low-priority modalities with high-priority modalities to selectively evict KV cache of low-priority modalities. The experimental results show that AccKV can significantly improve the computational efficiency of AV-LLMs while maintaining accuracy.

cs.MM↗

FlowMM: Cross-Modal Information Flow Guided KV Cache Merging for Efficient Multimodal Context Inference

Traditional KV cache eviction strategies, which discard less critical KV-pairs based on attention scores, often degrade generation quality, causing context loss or hallucinations. Recent efforts shift toward KV merging, merging eviction tokens with retention tokens based on similarity. However, in multimodal scenarios, distributional biases across modality tokens and attentional biases in cross-modal interactions limit its effectiveness. This work introduces FlowMM, an adaptive framework for cross-modal information flow-guided multimodal KV cache merging. FlowMM leverages cross-modal information flow to dynamically apply layer-specific merging strategies, capturing modality-specific patterns while preserving contextual integrity. Furthermore, we introduce a sensitivity-adaptive token matching mechanism that jointly evaluates token similarity and task-critical sensitivity, merging low-risk tokens while safeguarding high-sensitivity ones. Extensive experiments across diverse leading MLLMs show that FlowMM reduces KV cache memory by 80% to 95% and decoding latency by 1.3-1.8x, while maintaining competitive task performance.

cs.CL↗

PureKV: Plug-and-Play KV Cache Optimization with Spatial-Temporal Sparse Attention for Vision-Language Large Models

Vision-Language Large Models (VLLMs) face significant efficiency challenges when processing high-resolution inputs. The quadratic complexity in attention and autoregressive generation, as well as the constantly growing key value (KV) cache size, severely hinder the prefilling and decoding stages. Recent efforts have attempted to compress KV cache by identifying and pruning KV cache of less important tokens, but these methods typically rely on attention scores to estimate token importance, making them incompatible with efficient attention mechanisms such as FlashAttention and Sparse Attention, which do not explicitly compute attention matrices. Moreover, existing methods overlook how sparse attention, while accelerating the prefilling stage, alters the information structure of the KV cache, thereby compromising the effectiveness of downstream KV cache compression strategies. To address this issue, we propose PureKV, a plug-and-play framework for joint optimization of sparse attention and KV cache compression. We first introduce a KV cache compression strategy that is fully compatible with efficient attention accelerators. Our method utilizes lower layer attention scores to estimate the importance of high layers' KV cache, enabling active pruning without compromising accuracy. In addition, we have designed a Spatial-Temporal Sparse Attention (ST-SpAttn) module specifically tailored for video KV cache compression algorithms. This module combines spatial and temporal attention sparsity to improve the compression efficiency of KV cache optimization algorithms by purifying spatial noise and temporal redundancy in KV cache. At the same time, ST-SpAttn also accelerated the prefilling stage of VLLMs. Extensive experiments on VLLMs (VideoLLaMA2, Qwen2.5-VL) have shown that PureKV achieves 5.0 times KV cache compression and 3.16 times prefill acceleration, with negligible quality degradation.

cs.MM↗

MNN-AECS: Energy Optimization for LLM Decoding on Mobile Devices via Adaptive Core Selection

As the demand for on-device Large Language Model (LLM) inference grows, energy efficiency has become a major concern, especially for battery-limited mobile devices. Our analysis shows that the memory-bound LLM decode phase dominates energy use, and yet most existing works focus on accelerating the prefill phase, neglecting energy concerns. We introduce Adaptive Energy-Centric Core Selection (AECS) and integrate it into MNN to create the energy-efficient version, MNN-AECS, the first engine-level system solution without requiring root access or OS modifications for energy-efficient LLM decoding. MNN-AECS is designed to reduce LLM decoding energy while keeping decode speed within an acceptable slowdown threshold by dynamically selecting low-power CPU cores. MNN-AECS is evaluated across 5 Android and 2 iOS devices on 5 popular LLMs of various sizes. Compared to original MNN, MNN-AECS cuts down energy use by 23% without slowdown averaged over all 7 devices and 4 datasets. Against other engines, including llama.cpp, executorch, mllm, and MediaPipe, MNN-AECS delivers 39% to 78% energy saving and 12% to 363% speedup on average.

cs.OS↗

MNN-LLM: A Generic Inference Engine for Fast Large Language Model Deployment on Mobile Devices

Large language models (LLMs) have demonstrated exceptional performance across a variety of tasks. However, their substantial scale leads to significant computational resource consumption during inference, resulting in high costs. Consequently, edge device inference presents a promising solution. The primary challenges of edge inference include memory usage and inference speed. This paper introduces MNN-LLM, a framework specifically designed to accelerate the deployment of large language models on mobile devices. MNN-LLM addresses the runtime characteristics of LLMs through model quantization and DRAM-Flash hybrid storage, effectively reducing memory usage. It rearranges weights and inputs based on mobile CPU instruction sets and GPU characteristics while employing strategies such as multicore load balancing, mixed-precision floating-point operations, and geometric computations to enhance performance. Notably, MNN-LLM achieves up to a 8.6x speed increase compared to current mainstream LLM-specific frameworks.

cs.LG↗

MadaKV: Adaptive Modality-Perception KV Cache Eviction for Efficient Multimodal Long-Context Inference

This paper introduces MadaKV, a modality-adaptive key-value (KV) cache eviction strategy designed to enhance the efficiency of multimodal large language models (MLLMs) in long-context inference. In multimodal scenarios, attention heads exhibit varying preferences for different modalities, resulting in significant disparities in modality importance across attention heads. Traditional KV cache eviction methods, which are tailored for unimodal settings, fail to capture modality-specific information, thereby yielding suboptimal performance. MadaKV addresses these challenges through two key components: modality preference adaptation and hierarchical compression compensation. By dynamically sensing modality information within attention heads and adaptively retaining critical tokens, MadaKV achieves substantial reductions in KV cache memory footprint and model inference decoding latency (1.3 to 1.5 times improvement) while maintaining high accuracy across various multimodal long-context tasks. Extensive experiments on representative MLLMs and the MileBench benchmark demonstrate the effectiveness of MadaKV compared to existing KV cache eviction methods.

cs.LG↗

Walle: An End-to-End, General-Purpose, and Large-Scale Production System for Device-Cloud Collaborative Machine Learning

To break the bottlenecks of mainstream cloud-based machine learning (ML) paradigm, we adopt device-cloud collaborative ML and build the first end-to-end and general-purpose system, called Walle, as the foundation. Walle consists of a deployment platform, distributing ML tasks to billion-scale devices in time; a data pipeline, efficiently preparing task input; and a compute container, providing a cross-platform and high-performance execution environment, while facilitating daily task iteration. Specifically, the compute container is based on Mobile Neural Network (MNN), a tensor compute engine along with the data processing and model execution libraries, which are exposed through a refined Python thread-level virtual machine (VM) to support diverse ML tasks and concurrent task execution. The core of MNN is the novel mechanisms of operator decomposition and semi-auto search, sharply reducing the workload in manually optimizing hundreds of operators for tens of hardware backends and further quickly identifying the best backend with runtime optimization for a computation graph. The data pipeline introduces an on-device stream processing framework to enable processing user behavior data at source. The deployment platform releases ML tasks with an efficient push-then-pull method and supports multi-granularity deployment policies. We evaluate Walle in practical e-commerce application scenarios to demonstrate its effectiveness, efficiency, and scalability. Extensive micro-benchmarks also highlight the superior performance of MNN and the Python thread-level VM. Walle has been in large-scale production use in Alibaba, while MNN has been open source with a broad impact in the community.

cs.LG↗