arXiv subjects

Menghao Li

Publications and source records attributed to Menghao Li.

13 recordsLinked to original sources

arXiv2026

CA-OPD: Confidence-Aware On-Policy Distillation for Structured Visual Prediction

Autoregressive vision language models unify heterogeneous perception tasks but are highly susceptible to compounding errors. On-policy distillation (OPD) bridges the training-inference mismatch by training students on their own rollouts. However, unreliable student predictions, especially early in training, can derail the trajectory and degrade the quality of teacher supervision. While recent interleaved distillation methods allow the teacher to verify and replace student tokens, they primarily rely on rigid ranking metrics rather than exact teacher confidence, and they overlook how intervention decisions can inform token-level supervision. To address this, we introduce Confidence-Aware On-Policy Distillation (CA-OPD), a framework that couples reliable rollout construction with adaptive supervision. CA-OPD utilizes teacher confidence to selectively correct unreliable student transitions, gradually transferring rollout control to the student via a strict-to-relaxed schedule. Crucially, CA-OPD aligns knowledge transfer with these intervention decisions: corrected positions receive direct cross-entropy supervision from the teacher's prediction, while retained positions benefit from the teacher's full predictive distribution. Evaluated in a multi-teacher setting for GUI grounding and optical character recognition, CA-OPD substantially improves the Qwen3.5-0.8B baseline across all six target benchmarks, including gains of $9.50$ points on ScreenSpot-Pro and $6.72$ points on OCRBench-v2 English. Controlled studies further show that the gains depend on intervention placement, progressive rollout control, and intervention-aligned supervision, rather than intervention frequency alone.

cs.CV↗

arXiv2026

MESA:Task-Adaptive Multi-Structure Evidence Selection for Long-Horizon Agent Memory

Long-horizon agents accumulate trajectories spanning hundreds of interleaved reasoning, action, and observation steps, where answering a query may depend on evidence buried far back in the history. External memory stores such trajectories as structured representations, yet each structure provides a distinct and incomplete view. Existing multi-memory systems either read a fixed set of structures for every query, inflating context and introducing noise, or route each query to a single structure, preventing the composition of complementary evidence. A controlled analysis on AMA-Bench shows that the optimal memory configuration is typically neither a single structure nor the full union, but a tailored composition of multiple structural memories that varies with query and task demands. Motivated by these findings, we formulate structure-level dynamic selection: selecting and fusing a query-adaptive subset from a library of specialized memory structures. We propose MESA (a Multi-structure Evidence Selection framework for long-horizon Agent), which builds five complementary structure views of each trajectory and learns from end-to-end answer-level feedback to select and fuse a query-specific subset for a frozen answer model. To learn under this weak supervision, MESA employs harness optimization with prior-guided search and UCB-guided scheduling to balance exploration and exploitation. On AMA-Bench, MESA outperforms the strongest baseline by 8.5% while using 41% fewer evidence tokens than the all-structure alternative.

cs.AI↗

arXiv2026

Native3D: End-to-End 3D Scene Generation via Unified Mesh-Texture Modeling and Semantic Alignment

This paper presents Native3D, the first end-to-end 3D scene generation framework that completely bypasses 2D intermediate representations. Traditional approaches typically require adapting 3D representations to the 2D domain to leverage pre-trained diffusion models, which inevitably introduces domain adaptation issues including geometric structural distortion and texture detail degradation. To address these limitations, we design a unified mesh-texture joint representation that simultaneously models both geometric structures and texture features through a Transformer-based scene encoder, effectively maintaining spatial relationships and visual consistency among objects within scenes. We further propose the 3D Representation Alignment Loss (3D REPA Loss), which employs an improved contrastive learning mechanism to align multi-level semantic representations in the latent space, significantly enhancing geometric and textural fidelity. Experimental results demonstrate that Native3D outperforms existing methods in both generation quality and editing flexibility, providing a novel solution for 3D scene editing.

cs.CV↗

arXiv2026

Beyond Semantic Organization: Memory as Execution State Management for Long-Horizon Agents

LLM-based agents increasingly tackle long-horizon tasks with interdependent decisions, where each action reshapes future constraints and intermediate errors can cascade. Existing RAG and agent memory systems organize histories by semantic similarity, retrieving content-relevant entries at decision time. We argue that this design mismatches execution-state dependencies: it fragments decision trajectories and mixes valid and erroneous traces, hindering coherent state reconstruction and error isolation. We propose MAGE (Memory as Agent-Guided Exploration), an active execution-state manager that stores interactions in a hierarchical state tree. The agent derives its state from the active root-to-current path, combining subgoal summaries, recent traces, and hints from prior branches. Four coupled operations maintain the tree: Grow records new traces, Compress summarizes completed subgoals, Maintain validates summaries, and Revise restores a target boundary and resumes on a new branch. This design bounds context growth while preserving state integrity and isolating flawed segments from the active path. Experiments on MemoryArena show that MAGE improves the average task success rate by 7.8--20.4 pp over baselines, while reducing token consumption by 55.1%.

cs.AI↗

arXiv2025

Scalable Distributed Vector Search via Accuracy Preserving Index Construction

Scaling Approximate Nearest Neighbor Search (ANNS) to billions of vectors requires distributed indexes that balance accuracy, latency, and throughput. Yet existing index designs struggle with this tradeoff. This paper presents SPIRE, a scalable vector index based on two design decisions. First, it identifies a balanced partition granularity that avoids read-cost explosion. Second, it introduces an accuracy-preserving recursive construction that builds a multi-level index with predictable search cost and stable accuracy. In experiments with up to 8 billion vectors across 46 nodes, SPIRE achieves high scalability and up to 9.64X higher throughput than state-of-the-art systems.

cs.DC↗

arXiv2025

DISTRIBUTEDANN: Efficient Scaling of a Single DISKANN Graph Across Thousands of Computers

We present DISTRIBUTEDANN, a distributed vector search service that makes it possible to search over a single 50 billion vector graph index spread across over a thousand machines that offers 26ms median query latency and processes over 100,000 queries per second. This is 6x more efficient than existing partitioning and routing strategies that route the vector query to a subset of partitions in a scale out vector search system. DISTRIBUTEDANN is built using two well-understood components: a distributed key-value store and an in-memory ANN index. DISTRIBUTEDANN has replaced conventional scale-out architectures for serving the Bing search engine, and we share our experience from making this transition.

cs.DC↗

arXiv2025

TransVDM: Motion-Constrained Video Diffusion Model for Transparent Video Synthesis

Recent developments in Video Diffusion Models (VDMs) have demonstrated remarkable capability to generate high-quality video content. Nonetheless, the potential of VDMs for creating transparent videos remains largely uncharted. In this paper, we introduce TransVDM, the first diffusion-based model specifically designed for transparent video generation. TransVDM integrates a Transparent Variational Autoencoder (TVAE) and a pretrained UNet-based VDM, along with a novel Alpha Motion Constraint Module (AMCM). The TVAE captures the alpha channel transparency of video frames and encodes it into the latent space of the VDMs, facilitating a seamless transition to transparent video diffusion models. To improve the detection of transparent areas, the AMCM integrates motion constraints from the foreground within the VDM, helping to reduce undesirable artifacts. Moreover, we curate a dataset containing 250K transparent frames for training. Experimental results demonstrate the effectiveness of our approach across various benchmarks.

cs.GR↗

arXiv2024

Tora: Trajectory-oriented Diffusion Transformer for Video Generation

Recent advancements in Diffusion Transformer (DiT) have demonstrated remarkable proficiency in producing high-quality video content. Nonetheless, the potential of transformer-based diffusion models for effectively generating videos with controllable motion remains an area of limited exploration. This paper introduces Tora, the first trajectory-oriented DiT framework that concurrently integrates textual, visual, and trajectory conditions, thereby enabling scalable video generation with effective motion guidance. Specifically, Tora consists of a Trajectory Extractor (TE), a Spatial-Temporal DiT, and a Motion-guidance Fuser (MGF). The TE encodes arbitrary trajectories into hierarchical spacetime motion patches with a 3D motion compression network. The MGF integrates the motion patches into the DiT blocks to generate consistent videos that accurately follow designated trajectories. Our design aligns seamlessly with DiT's scalability, allowing precise control of video content's dynamics with diverse durations, aspect ratios, and resolutions. Extensive experiments demonstrate that Tora excels in achieving high motion fidelity compared to the foundational DiT model, while also accurately simulating the complex movements of the physical world. Code is made available at https://github.com/alibaba/Tora .

cs.CV↗

arXiv2024

Deep Linear Array Pushbroom Image Restoration: A Degradation Pipeline and Jitter-Aware Restoration Network

Linear Array Pushbroom (LAP) imaging technology is widely used in the realm of remote sensing. However, images acquired through LAP always suffer from distortion and blur because of camera jitter. Traditional methods for restoring LAP images, such as algorithms estimating the point spread function (PSF), exhibit limited performance. To tackle this issue, we propose a Jitter-Aware Restoration Network (JARNet), to remove the distortion and blur in two stages. In the first stage, we formulate an Optical Flow Correction (OFC) block to refine the optical flow of the degraded LAP images, resulting in pre-corrected images where most of the distortions are alleviated. In the second stage, for further enhancement of the pre-corrected images, we integrate two jitter-aware techniques within the Spatial and Frequency Residual (SFRes) block: 1) introducing Coordinate Attention (CoA) to the SFRes block in order to capture the jitter state in orthogonal direction; 2) manipulating image features in both spatial and frequency domains to leverage local and global priors. Additionally, we develop a data synthesis pipeline, which applies Continue Dynamic Shooting Model (CDSM) to simulate realistic degradation in LAP images. Both the proposed JARNet and LAP image synthesis pipeline establish a foundation for addressing this intricate challenge. Extensive experiments demonstrate that the proposed two-stage method outperforms state-of-the-art image restoration models. Code is available at https://github.com/JHW2000/JARNet.

cs.CV↗

arXiv2023

Iterative Robust Visual Grounding with Masked Reference based Centerpoint Supervision

Visual Grounding (VG) aims at localizing target objects from an image based on given expressions and has made significant progress with the development of detection and vision transformer. However, existing VG methods tend to generate false-alarm objects when presented with inaccurate or irrelevant descriptions, which commonly occur in practical applications. Moreover, existing methods fail to capture fine-grained features, accurate localization, and sufficient context comprehension from the whole image and textual descriptions. To address both issues, we propose an Iterative Robust Visual Grounding (IR-VG) framework with Masked Reference based Centerpoint Supervision (MRCS). The framework introduces iterative multi-level vision-language fusion (IMVF) for better alignment. We use MRCS to ahieve more accurate localization with point-wised feature supervision. Then, to improve the robustness of VG, we also present a multi-stage false-alarm sensitive decoder (MFSD) to prevent the generation of false-alarm objects when presented with inaccurate expressions. The proposed framework is evaluated on five regular VG datasets and two newly constructed robust VG datasets. Extensive experiments demonstrate that IR-VG achieves new state-of-the-art (SOTA) results, with improvements of 25\% and 10\% compared to existing SOTA approaches on the two newly proposed robust VG datasets. Moreover, the proposed framework is also verified effective on five regular VG datasets. Codes and models will be publicly at https://github.com/cv516Buaa/IR-VG.

cs.CV↗

arXiv2023

Mobile Image Restoration via Prior Quantization

In digital images, the performance of optical aberration is a multivariate degradation, where the spectral of the scene, the lens imperfections, and the field of view together contribute to the results. Besides eliminating it at the hardware level, the post-processing system, which utilizes various prior information, is significant for correction. However, due to the content differences among priors, the pipeline that aligns these factors shows limited efficiency and unoptimized restoration. Here, we propose a prior quantization model to correct the optical aberrations in image processing systems. To integrate these messages, we encode various priors into a latent space and quantify them by the learnable codebooks. After quantization, the prior codes are fused with the image restoration branch to realize targeted optical aberration correction. Comprehensive experiments demonstrate the flexibility of the proposed method and validate its potential to accomplish targeted restoration for a specific camera. Furthermore, our model promises to analyze the correlation between the various priors and the optical aberration of devices, which is helpful for joint soft-hardware design.

cs.CV↗

arXiv2022

Look in Different Views: Multi-Scheme Regression Guided Cell Instance Segmentation

Cell instance segmentation is a new and challenging task aiming at joint detection and segmentation of every cell in an image. Recently, many instance segmentation methods have applied in this task. Despite their great success, there still exists two main weaknesses caused by uncertainty of localizing cell center points. First, densely packed cells can easily be recognized into one cell. Second, elongated cell can easily be recognized into two cells. To overcome these two weaknesses, we propose a novel cell instance segmentation network based on multi-scheme regression guidance. With multi-scheme regression guidance, the network has the ability to look each cell in different views. Specifically, we first propose a gaussian guidance attention mechanism to use gaussian labels for guiding the network's attention. We then propose a point-regression module for assisting the regression of cell center. Finally, we utilize the output of the above two modules to further guide the instance segmentation. With multi-scheme regression guidance, we can take full advantage of the characteristics of different regions, especially the central region of the cell. We conduct extensive experiments on benchmark datasets, DSB2018, CA2.5 and SCIS. The encouraging results show that our network achieves SOTA (state-of-the-art) performance. On the DSB2018 and CA2.5, our network surpasses previous methods by 1.2% (AP50). Particularly on SCIS dataset, our network performs stronger by large margin (3.0% higher AP50). Visualization and analysis further prove that our proposed method is interpretable.

cs.CV↗

arXiv2018

Understanding Android Obfuscation Techniques: A Large-Scale Investigation in the Wild

In this paper, we seek to better understand Android obfuscation and depict a holistic view of the usage of obfuscation through a large-scale investigation in the wild. In particular, we focus on four popular obfuscation approaches: identifier renaming, string encryption, Java reflection, and packing. To obtain the meaningful statistical results, we designed efficient and lightweight detection models for each obfuscation technique and applied them to our massive APK datasets (collected from Google Play, multiple third-party markets, and malware databases). We have learned several interesting facts from the result. For example, malware authors use string encryption more frequently, and more apps on third-party markets than Google Play are packed. We are also interested in the explanation of each finding. Therefore we carry out in-depth code analysis on some Android apps after sampling. We believe our study will help developers select the most suitable obfuscation approach, and in the meantime help researchers improve code analysis systems in the right direction.

cs.CR↗