Chengfei Lv

Chengfei Lv, Chaoyue Niu, Renjie Gu et al.

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.

Xiang Chen, Duanzheng Song, Honghao Gui et al.

Despite their impressive generative capabilities, LLMs are hindered by fact-conflicting hallucinations in real-world applications. The accurate identification of hallucinations in texts generated by LLMs, especially in complex inferential scenarios, is a relatively unexplored area. To address this gap, we present FactCHD, a dedicated benchmark designed for the detection of fact-conflicting hallucinations from LLMs. FactCHD features a diverse dataset that spans various factuality patterns, including vanilla, multi-hop, comparison, and set operation. A distinctive element of FactCHD is its integration of fact-based evidence chains, significantly enhancing the depth of evaluating the detectors' explanations. Experiments on different LLMs expose the shortcomings of current approaches in detecting factual errors accurately. Furthermore, we introduce Truth-Triangulator that synthesizes reflective considerations by tool-enhanced ChatGPT and LoRA-tuning based on Llama2, aiming to yield more credible detection through the amalgamation of predictive results and evidence. The benchmark dataset is available at https://github.com/zjunlp/FactCHD.

Bowen Cai, Jinchi Huang, Rongfei Jia et al.

This paper studies implicit surface reconstruction leveraging differentiable ray casting. Previous works such as IDR and NeuS overlook the spatial context in 3D space when predicting and rendering the surface, thereby may fail to capture sharp local topologies such as small holes and structures. To mitigate the limitation, we propose a flexible neural implicit representation leveraging hierarchical voxel grids, namely Neural Deformable Anchor (NeuDA), for high-fidelity surface reconstruction. NeuDA maintains the hierarchical anchor grids where each vertex stores a 3D position (or anchor) instead of the direct embedding (or feature). We optimize the anchor grids such that different local geometry structures can be adaptively encoded. Besides, we dig into the frequency encoding strategies and introduce a simple hierarchical positional encoding method for the hierarchical anchor structure to flexibly exploit the properties of high-frequency and low-frequency geometry and appearance. Experiments on both the DTU and BlendedMVS datasets demonstrate that NeuDA can produce promising mesh surfaces.

Shuofei Qiao, Ningyu Zhang, Runnan Fang et al.

Language agents have achieved considerable performance on various complex question-answering tasks by planning with external tools. Despite the incessant exploration in this field, existing language agent systems still struggle with costly, non-reproducible data reliance and face the challenge of compelling a single model for multiple functions. To this end, we introduce AutoAct, an automatic agent learning framework for QA that does not rely on large-scale annotated data and synthetic planning trajectories from closed-source models (e.g., GPT-4). Given limited data with a tool library, AutoAct first automatically synthesizes planning trajectories without any assistance from humans or strong closed-source models. Then, AutoAct leverages a division-of-labor strategy to automatically differentiate based on the target task information and synthesized trajectories, producing a sub-agent group to complete the task. We conduct comprehensive experiments with different LLMs, which demonstrates that AutoAct yields better or parallel performance compared to various strong baselines. Further analysis demonstrates the effectiveness of the division-of-labor strategy, with the trajectory quality generated by AutoAct generally outperforming that of others. Code will be available at https://github.com/zjunlp/AutoAct.

Yuxin Zhang, Mengxue Hu, Zheng Lin et al.

Large language model (LLM) agents excel at solving complex long-horizon tasks through autonomous interaction with environments. However, their real-world deployment faces a fundamental device--cloud dilemma: on-device models are efficient but often brittle, while cloud models are stronger but costly in computation. State-of-the-art LLM device--cloud routers usually make coarse task-level decisions, which cannot adapt to the changing difficulty of multi-step agent interactions. To address this issue, we present Hera, a step-level device--cloud LLM agent coordinator for long-horizon tasks achieving a strong performance--cost Pareto frontier. Hera adopts a novel two-stage training paradigm: (1) imitation learning for cold-start, followed by (2) reinforcement learning that jointly optimizes task success and cloud usage efficiency. The first stage casts step-level routing as a supervised classification problem: the device agent is replayed on cloud trajectories, with each state labeled by the agreement between device and cloud actions. In the second stage, we perform cost-aware reinforcement learning by grouping identical states across trajectories and updating Hera with labels favoring higher expected return and fewer future cloud calls. We evaluate Hera on ALFWorld, WebShop, and AppWorld, where it consistently outperforms prior methods, achieving 92.5% of the cloud-only success rate with cloud use in only 46.3% of steps.

Yujie Sun, Zhuoqiang Cai, Chaoyue Niu et al.

We present FHAvatar, a novel framework for reconstructing 3D Gaussian avatars with composable face and hair components from an arbitrary number of views. Unlike previous approaches that couple facial and hair representations within a unified modeling process, we explicitly decouple two components in texture space by representing the face with planar Gaussians and the hair with strand-based Gaussians. To overcome the limitations of existing methods that rely on dense multi-view captures or costly per-identity optimization, we propose an aggregated transformer backbone to learn geometry-aware cross-view priors and head-hair structural coherence from multi-view datasets, enabling effective and efficient feature extraction and fusion from few casual captures. Extensive quantitative and qualitative experiments demonstrate that FHAvatar achieves state-of-the-art reconstruction quality from only a few observations of new identities within minutes, while supporting real-time animation, convenient hairstyle transfer, and stylized editing, broadening the accessibility and applicability of digital avatar creation.

Tianyu Zhan, Kairui Fu, Chengfei Lv et al.

Generative Recommendation (GR) has recently transitioned from atomic item-indexing to Semantic ID (SID)-based frameworks to capture intrinsic item relationships and enhance generalization. However, the adoption of high-granularity SIDs leads to two critical challenges: prohibitive training overhead due to sequence expansion and unstable performance reliability characterized by non-monotonic accuracy fluctuations. We identify that these disparate issues are fundamentally rooted in the Semantic Dilution Effect, where redundant tokens waste massive computation and dilute the already sparse learning signals in recommendation. To counteract this, we propose STAMP (Semantic Trimming and Auxiliary Multi-step Prediction), a framework utilizing a dual-end optimization strategy. We argue that effective SID learning requires simultaneously addressing low input information density and sparse output supervision. On the input side, Semantic Adaptive Pruning (SAP) dynamically filters redundancy during the forward pass, converting noise-laden sequences into compact, information-rich representations. On the output side, Multi-step Auxiliary Prediction (MAP) employs a multi-token objective to densify feedback, strengthening long-range dependency capture and ensuring robust learning signals despite compressed inputs. Unifying input purification and signal amplification, STAMP enhances both training efficiency and representation capability. Experiments on public Amazon and large-scale industrial datasets show STAMP achieves 1.23--1.38$\times$ speedup and 17.2\%--54.7\% VRAM reduction while maintaining or improving performance across multiple architectures.

Zhonghua Jiang, Kui Chen, Kunxi Li et al.

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.

Shuofei Qiao, Honghao Gui, Chengfei Lv et al.

Tools serve as pivotal interfaces that enable humans to understand and reshape the environment. With the advent of foundation models, AI systems can utilize tools to expand their capabilities and interact with the real world. Existing tool learning methodologies, encompassing supervised fine-tuning and prompt engineering approaches, often induce large language models to utilize tools indiscriminately, as complex tasks often exceed their own competencies. However, introducing tools for simple tasks, which the models themselves can readily resolve, can inadvertently propagate errors rather than enhance performance. This leads to the research question: can we teach language models when and how to use tools? To meet this need, we propose Tool leaRning wIth exeCution fEedback (TRICE), a two-stage end-to-end framework that enables the model to continually learn through feedback derived from tool execution, thereby learning when and how to use tools effectively. Experimental results, backed by further analysis, show that TRICE can make the large language model selectively use tools by improving the accuracy of tool usage while enhancing insufficient tool learning and mitigating excessive reliance on tools. Code is available at https://github.com/zjunlp/TRICE.

Xiaotang Jiang, Huan Wang, Yiliu Chen et al.

Deploying deep learning models on mobile devices draws more and more attention recently. However, designing an efficient inference engine on devices is under the great challenges of model compatibility, device diversity, and resource limitation. To deal with these challenges, we propose Mobile Neural Network (MNN), a universal and efficient inference engine tailored to mobile applications. In this paper, the contributions of MNN include: (1) presenting a mechanism called pre-inference that manages to conduct runtime optimization; (2)deliveringthorough kernel optimization on operators to achieve optimal computation performance; (3) introducing backend abstraction module which enables hybrid scheduling and keeps the engine lightweight. Extensive benchmark experiments demonstrate that MNN performs favorably against other popular lightweight deep learning frameworks. MNN is available to public at: https://github.com/alibaba/MNN.

Xi Chen, Sida Peng, Dongchen Yang et al.

This paper aims to recover object materials from posed images captured under an unknown static lighting condition. Recent methods solve this task by optimizing material parameters through differentiable physically based rendering. However, due to the coupling between object geometry, materials, and environment lighting, there is inherent ambiguity during the inverse rendering process, preventing previous methods from obtaining accurate results. To overcome this ill-posed problem, our key idea is to learn the material prior with a generative model for regularizing the optimization process. We observe that the general rendering equation can be split into diffuse and specular shading terms, and thus formulate the material prior as diffusion models of albedo and specular. Thanks to this design, our model can be trained using the existing abundant 3D object data, and naturally acts as a versatile tool to resolve the ambiguity when recovering material representations from RGB images. In addition, we develop a coarse-to-fine training strategy that leverages estimated materials to guide diffusion models to satisfy multi-view consistent constraints, leading to more stable and accurate results. Extensive experiments on real-world and synthetic datasets demonstrate that our approach achieves state-of-the-art performance on material recovery. The code will be available at https://zju3dv.github.io/IntrinsicAnything.

Bin Zhang, Weipeng Huang, Dimin Wang et al.

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.

Hongyun Yu, Zhan Qu, Qihang Yu et al.

Recent works on audio-driven talking head synthesis using Neural Radiance Fields (NeRF) have achieved impressive results. However, due to inadequate pose and expression control caused by NeRF implicit representation, these methods still have some limitations, such as unsynchronized or unnatural lip movements, and visual jitter and artifacts. In this paper, we propose GaussianTalker, a novel method for audio-driven talking head synthesis based on 3D Gaussian Splatting. With the explicit representation property of 3D Gaussians, intuitive control of the facial motion is achieved by binding Gaussians to 3D facial models. GaussianTalker consists of two modules, Speaker-specific Motion Translator and Dynamic Gaussian Renderer. Speaker-specific Motion Translator achieves accurate lip movements specific to the target speaker through universalized audio feature extraction and customized lip motion generation. Dynamic Gaussian Renderer introduces Speaker-specific BlendShapes to enhance facial detail representation via a latent pose, delivering stable and realistic rendered videos. Extensive experimental results suggest that GaussianTalker outperforms existing state-of-the-art methods in talking head synthesis, delivering precise lip synchronization and exceptional visual quality. Our method achieves rendering speeds of 130 FPS on NVIDIA RTX4090 GPU, significantly exceeding the threshold for real-time rendering performance, and can potentially be deployed on other hardware platforms.

Yufei Yin, Jie Zheng, Qianke Meng et al.

Zero-shot 3D Visual Grounding (3DVG) is a critical capability for open-world embodied AI. However, existing methods are fundamentally bottlenecked by the poor quality of open-vocabulary 3D proposals, suffering from inaccurate categories and imprecise geometries, as well as the spatial redundancy of exhaustive multi-view reasoning. To address these challenges, we propose MCM-VG, a novel framework that achieves robust zero-shot 3DVG by explicitly establishing Multiple Consistent 2D-3D Mappings. Instead of passively relying on noisy 3D segments, MCM-VG enforces 2D-3D consistency across three fundamental dimensions to achieve precise target localization and reliable reasoning. First, a Semantic Alignment module corrects category mismatches via LLM-driven query parsing and coarse-to-fine 2D-3D matching. Second, an Instance Rectification module leverages VLM-guided 2D segmentations to reconstruct missing targets, back-projecting these reliable visual priors to establish accurate 3D geometries. Finally, to eliminate spatial redundancy, a Viewpoint Distillation module clusters 3D camera directions to extract optimal frames. By pairing these optimal RGB frames with Bird's Eye View maps into concise visual prompt sets, we formulate the final target disambiguation as a multiple-choice reasoning task for Vision-Language Models. Extensive evaluations on ScanRefer and Nr3D benchmarks demonstrate that MCM-VG sets a new state-of-the-art for zero-shot 3D visual grounding. Remarkably, it achieves 62.0\% and 53.6\% in Acc@0.25 and Acc@0.5 on ScanRefer, outperforming previous baselines by substantial margins of 6.4\% and 4.0\%.

Xingze Zou, Jing Wang, Yuhua Zheng et al.

Large language models (LLMs) have demonstrated remarkable capabilities in code generation, yet their potential for generating kernels specifically for mobile de- vices 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 inher-ent 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 K ernel A gent (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 kernelsto deliver measurable speedups over native libraries.

Zhaode Wang, Jingbang Yang, Xinyu Qian et al.

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.

Jianchuan Chen, Jingchuan Hu, Gaige Wang et al.

Realistic 3D full-body talking avatars hold great potential in AR, with applications ranging from e-commerce live streaming to holographic communication. Despite advances in 3D Gaussian Splatting (3DGS) for lifelike avatar creation, existing methods struggle with fine-grained control of facial expressions and body movements in full-body talking tasks. Additionally, they often lack sufficient details and cannot run in real-time on mobile devices. We present TaoAvatar, a high-fidelity, lightweight, 3DGS-based full-body talking avatar driven by various signals. Our approach starts by creating a personalized clothed human parametric template that binds Gaussians to represent appearances. We then pre-train a StyleUnet-based network to handle complex pose-dependent non-rigid deformation, which can capture high-frequency appearance details but is too resource-intensive for mobile devices. To overcome this, we "bake" the non-rigid deformations into a lightweight MLP-based network using a distillation technique and develop blend shapes to compensate for details. Extensive experiments show that TaoAvatar achieves state-of-the-art rendering quality while running in real-time across various devices, maintaining 90 FPS on high-definition stereo devices such as the Apple Vision Pro.

Zhen Lv, Yangqi Long, Congzhentao Huang et al.

Stereo video synthesis from a monocular input is a demanding task in the fields of spatial computing and virtual reality. The main challenges of this task lie on the insufficiency of high-quality paired stereo videos for training and the difficulty of maintaining the spatio-temporal consistency between frames. Existing methods primarily address these issues by directly applying novel view synthesis (NVS) techniques to video, while facing limitations such as the inability to effectively represent dynamic scenes and the requirement for large amounts of training data. In this paper, we introduce a novel self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which meets the challenges head-on. Firstly, to address the stereo video data insufficiency, we propose a Depth based Video Generation module DVG, which employs a forward-backward rendering mechanism to generate paired videos with geometric and temporal priors. Leveraging data generated by DVG, we propose RefinerNet along with a self-supervised synthetic framework designed to facilitate efficient and dedicated training. More importantly, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL for geometric and temporal consistency ensurance respectively. We evaluated the proposed method against various benchmark methods, with the results showcasing its superior performance.

Dian Zheng, Cheng Zhang, Xiao-Ming Wu et al.

Generating 360-degree panoramas from narrow field of view (NFoV) image is a promising computer vision task for Virtual Reality (VR) applications. Existing methods mostly assess the generated panoramas with InceptionNet or CLIP based metrics, which tend to perceive the image quality and is \textbf{not suitable for evaluating the distortion}. In this work, we first propose a distortion-specific CLIP, named Distort-CLIP to accurately evaluate the panorama distortion and discover the \textbf{``visual cheating''} phenomenon in previous works (\ie, tending to improve the visual results by sacrificing distortion accuracy). This phenomenon arises because prior methods employ a single network to learn the distinct panorama distortion and content completion at once, which leads the model to prioritize optimizing the latter. To address the phenomenon, we propose \textbf{PanoDecouple}, a decoupled diffusion model framework, which decouples the panorama generation into distortion guidance and content completion, aiming to generate panoramas with both accurate distortion and visual appeal. Specifically, we design a DistortNet for distortion guidance by imposing panorama-specific distortion prior and a modified condition registration mechanism; and a ContentNet for content completion by imposing perspective image information. Additionally, a distortion correction loss function with Distort-CLIP is introduced to constrain the distortion explicitly. The extensive experiments validate that PanoDecouple surpasses existing methods both in distortion and visual metrics.

Shuichang Lai, Letian Huang, Jie Guo et al.

Reconstructing objects from posed images is a crucial and complex task in computer graphics and computer vision. While NeRF-based neural reconstruction methods have exhibited impressive reconstruction ability, they tend to be time-comsuming. Recent strategies have adopted 3D Gaussian Splatting (3D-GS) for inverse rendering, which have led to quick and effective outcomes. However, these techniques generally have difficulty in producing believable geometries and materials for glossy objects, a challenge that stems from the inherent ambiguities of inverse rendering. To address this, we introduce GlossyGS, an innovative 3D-GS-based inverse rendering framework that aims to precisely reconstruct the geometry and materials of glossy objects by integrating material priors. The key idea is the use of micro-facet geometry segmentation prior, which helps to reduce the intrinsic ambiguities and improve the decomposition of geometries and materials. Additionally, we introduce a normal map prefiltering strategy to more accurately simulate the normal distribution of reflective surfaces. These strategies are integrated into a hybrid geometry and material representation that employs both explicit and implicit methods to depict glossy objects. We demonstrate through quantitative analysis and qualitative visualization that the proposed method is effective to reconstruct high-fidelity geometries and materials of glossy objects, and performs favorably against state-of-the-arts.

Chaoyue Niu, Yucheng Ding, Junhui Lu et al.

The conventional cloud-based large model learning framework is increasingly constrained by latency, cost, personalization, and privacy concerns. In this survey, we explore an emerging paradigm: collaborative learning between on-device small model and cloud-based large model, which promises low-latency, cost-efficient, and personalized intelligent services while preserving user privacy. We provide a comprehensive review across hardware, system, algorithm, and application layers. At each layer, we summarize key problems and recent advances from both academia and industry. In particular, we categorize collaboration algorithms into data-based, feature-based, and parameter-based frameworks. We also review publicly available datasets and evaluation metrics with user-level or device-level consideration tailored to collaborative learning settings. We further highlight real-world deployments, ranging from recommender systems and mobile livestreaming to personal intelligent assistants. We finally point out open research directions to guide future development in this rapidly evolving field.

Hang Zeng, Chaoyue Niu, Fan Wu et al.

Existing work on large language model (LLM) personalization assigned different responding roles to LLMs, but overlooked the diversity of queriers. In this work, we propose a new form of querier-aware LLM personalization, generating different responses even for the same query from different queriers. We design a dual-tower model architecture with a cross-querier general encoder and a querier-specific encoder. We further apply contrastive learning with multi-view augmentation, pulling close the dialogue representations of the same querier, while pulling apart those of different queriers. To mitigate the impact of query diversity on querier-contrastive learning, we cluster the dialogues based on query similarity and restrict the scope of contrastive learning within each cluster. To address the lack of datasets designed for querier-aware personalization, we also build a multi-querier dataset from English and Chinese scripts, as well as WeChat records, called MQDialog, containing 173 queriers and 12 responders. Extensive evaluations demonstrate that our design significantly improves the quality of personalized response generation, achieving relative improvement of 8.4% to 48.7% in ROUGE-L scores and winning rates ranging from 54% to 82% compared with various baseline methods.

Kunxi Li, Yufan Xiong, Zhonghua Jiang et al.

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.

Yuqi Tang, Kehua Feng, Yunfeng Wang et al.

Evaluating the conversational abilities of large language models (LLMs) remains a challenging task. Current mainstream approaches primarily rely on the "LLM-as-a-judge" paradigm, where an LLM is prompted to serve as an evaluator to assess dialogue quality. However, such methods often suffer from various biases, which undermine the reliability and consistency of the evaluation results. To mitigate these biases, recent methods employ multiple LLMs as judges and aggregate their judgments to select the optimal assessment. Although effective, this multi-judge approach incurs significant computational overhead during inference. In this paper, we propose an efficient multi-turn dialogue evaluator that captures the collective wisdom of multiple LLM judges by aggregating their preference knowledge into a single model. Our approach preserves the advantages of diverse multi-judge feedback while drastically reducing the evaluation cost, enabling fast and flexible dialogue quality assessment. Extensive experiments on seven single rating and pairwise comparison dialogue evaluation benchmarks demonstrate that our method outperforms existing baselines across diverse scenarios, showcasing its efficiency and robustness.

Zhengxiang Huang, Chaoyue Niu, Zhaode Wang et al.

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.

Yucheng Ding, Chaoyue Niu, Fan Wu et al.

We study practical data characteristics underlying federated learning, where non-i.i.d. data from clients have sparse features, and a certain client's local data normally involves only a small part of the full model, called a submodel. Due to data sparsity, the classical federated averaging (FedAvg) algorithm or its variants will be severely slowed down, because when updating the global model, each client's zero update of the full model excluding its submodel is inaccurately aggregated. Therefore, we propose federated submodel averaging (FedSubAvg), ensuring that the expectation of the global update of each model parameter is equal to the average of the local updates of the clients who involve it. We theoretically proved the convergence rate of FedSubAvg by deriving an upper bound under a new metric called the element-wise gradient norm. In particular, this new metric can characterize the convergence of federated optimization over sparse data, while the conventional metric of squared gradient norm used in FedAvg and its variants cannot. We extensively evaluated FedSubAvg over both public and industrial datasets. The evaluation results demonstrate that FedSubAvg significantly outperforms FedAvg and its variants.

Hongtao Lv, Zhenzhe Zheng, Tie Luo et al.

Federated learning (FL) trains a machine learning model on mobile devices in a distributed manner using each device's private data and computing resources. A critical issues is to evaluate individual users' contributions so that (1) users' effort in model training can be compensated with proper incentives and (2) malicious and low-quality users can be detected and removed. The state-of-the-art solutions require a representative test dataset for the evaluation purpose, but such a dataset is often unavailable and hard to synthesize. In this paper, we propose a method called Pairwise Correlated Agreement (PCA) based on the idea of peer prediction to evaluate user contribution in FL without a test dataset. PCA achieves this using the statistical correlation of the model parameters uploaded by users. We then apply PCA to designing (1) a new federated learning algorithm called Fed-PCA, and (2) a new incentive mechanism that guarantees truthfulness. We evaluate the performance of PCA and Fed-PCA using the MNIST dataset and a large industrial product recommendation dataset. The results demonstrate that our Fed-PCA outperforms the canonical FedAvg algorithm and other baseline methods in accuracy, and at the same time, PCA effectively incentivizes users to behave truthfully.

Yihao Xue, Chaoyue Niu, Zhenzhe Zheng et al.

Federated learning allows mobile clients to jointly train a global model without sending their private data to a central server. Extensive works have studied the performance guarantee of the global model, however, it is still unclear how each individual client influences the collaborative training process. In this work, we defined a new notion, called {\em Fed-Influence}, to quantify this influence over the model parameters, and proposed an effective and efficient algorithm to estimate this metric. In particular, our design satisfies several desirable properties: (1) it requires neither retraining nor retracing, adding only linear computational overhead to clients and the server; (2) it strictly maintains the tenets of federated learning, without revealing any client's local private data; and (3) it works well on both convex and non-convex loss functions, and does not require the final model to be optimal. Empirical results on a synthetic dataset and the FEMNIST dataset demonstrate that our estimation method can approximate Fed-Influence with small bias. Further, we show an application of Fed-Influence in model debugging.

Yiwu Yao, Yuchao Li, Chengyu Wang et al.

The intensive computation of Automatic Speech Recognition (ASR) models obstructs them from being deployed on mobile devices. In this paper, we present a novel quantized Winograd optimization pipeline, which combines the quantization and fast convolution to achieve efficient inference acceleration on mobile devices for ASR models. To avoid the information loss due to the combination of quantization and Winograd convolution, a Range-Scaled Quantization (RSQ) training method is proposed to expand the quantized numerical range and to distill knowledge from high-precision values. Moreover, an improved Conv1D equipped DFSMN (ConvDFSMN) model is designed for mobile deployment. We conduct extensive experiments on both ConvDFSMN and Wav2letter models. Results demonstrate the models can be effectively optimized with the proposed pipeline. Especially, Wav2letter achieves 1.48* speedup with an approximate 0.07% WER decrease on ARMv7-based mobile devices.

Chaoyue Niu, Fan Wu, Shaojie Tang et al.

Federated learning was proposed with an intriguing vision of achieving collaborative machine learning among numerous clients without uploading their private data to a cloud server. However, the conventional framework requires each client to leverage the full model for learning, which can be prohibitively inefficient for resource-constrained clients and large-scale deep learning tasks. We thus propose a new framework, called federated submodel learning, where clients download only the needed parts of the full model, namely submodels, and then upload the submodel updates. Nevertheless, the "position" of a client's truly required submodel corresponds to her private data, and its disclosure to the cloud server during interactions inevitably breaks the tenet of federated learning. To integrate efficiency and privacy, we have designed a secure federated submodel learning scheme coupled with a private set union protocol as a cornerstone. Our secure scheme features the properties of randomized response, secure aggregation, and Bloom filter, and endows each client with a customized plausible deniability, in terms of local differential privacy, against the position of her desired submodel, thus protecting her private data. We further instantiated our scheme with the e-commerce recommendation scenario in Alibaba, implemented a prototype system, and extensively evaluated its performance over 30-day Taobao user data. The analysis and evaluation results demonstrate the feasibility and scalability of our scheme from model accuracy and convergency, practical communication, computation, and storage overheads, as well as manifest its remarkable advantages over the conventional federated learning framework.