arXiv subjects

Huimin Wu

Publications and source records attributed to Huimin Wu.

15 recordsLinked to original sources

arXiv2026

DEVIS-GRPO: Unleashing GRPO on Dynamic Extreme View Synthesis

Trajectory-controlled video generation has become essential for controllable video generation. While current methods perform well under small-view camera motions, they degrade significantly with large-view motions. Existing solutions for extreme-view synthesis typically require dedicated video pairs, demanding substantial annotation effort. To address these limitations, we propose Dynamic Extreme VIew Synthesis-GRPO (DEVIS-GRPO), a GRPO-based framework for trajectory-controlled video generation, the first online policy gradient method for extreme view video generation. Central to our approach is a novel sampling strategy: Accumulative Dynamic Extreme VIew Synthesis (ADEVIS), which achieves large-view camera motions by progressively accumulating small-view increments. This method delivers two key advantages: 1) enhanced training efficiency, as it eliminates the need to warm-start the policy model by collecting expensive paired large-view videos, and 2) increased sampling diversity, achieved by flexibly varying trajectory configurations. Finally, we designed a multi-level consistency-quality reward function to select high-quality samples for model optimization. Experiments on the Kubric-4D, iPhone, and DL3DV datasets demonstrate our method's superiority. On Kubric-4D, we achieve relative improvements of 21.57% in PSNR and 7.31% in SSIM over the second-best method in non-occlusion areas. On iPhone, LPIPS is reduced by 18.56%.

cs.CV↗

arXiv2026

Behavioral-Level Simulation of Digital Readout for COFFEE at LHCb Upstream Pixel Tracker

COFFEE series is a HVCMOS pixel sensor using the advanced 55 nm process, currently being developed for the Upstream Pixel (UP) tracker of the LHCb Upgrade II. To ensure that COFFEE will be able to handle the particle hit rates at UP tracker, which reach a maximum of 322.5 MHz/chip, detailed simulation of the digital readout circuitry was performed. Simulation results show that the column-drain readout mechanism achieves nearly 100\% efficiency when the single readout cycle does not exceed 100 ns. Meanwhile, the buffer depth and memory resources required for the peripheral readout adapted to the BXID-sharing data format are also evaluated. These provide guidance for the design of COFFEE. The column-drain readout mechanism was used in COFFEE3 (fabricated in 2025), while the peripheral readout architecture adapted to the BXID-sharing data format is implemented in CHiR (taped out in early 2026).

physics.ins-det↗

arXiv2026

Design and First Results of COFFEE3: A 55nm HVCMOS Pixel Sensor Prototype for High-Energy Physics Applications

Motivated by the stringent requirements of the Upstream Pixel (UP) tracker in the LHCb Upgrade II and the Inner Tracking detector (ITK) of the Circular Electron Positron Collider, the COFFEE series of pixel sensor chips have been developed using a 55nm High-Voltage CMOS (HVCMOS) process. The primary objective is to achieve a time resolution of a few nanoseconds under a hit density of up to 100 MHz/cm$^2$, while maintaining fine spatial resolution ($\sim$10 $μ$m) and reasonable power consumption ($<$200 mW/cm$^2$). Building on the process validation of the COFFEE2 prototype, this work presents the design and preliminary test results of COFFEE3-a prototype integrating two distinct readout architectures. Architecture 1, tailored for the current triple-well process, adopts NMOS-only in-pixel circuitry and innovative column-level readout to handle high hit densities. The time walk of pixel-level signal is controlled within 10 ns, and the Time of Arrival (TOA) and Time over Threshold (TOT) are measured with a system clock with the period of 25 ns in peripheral circuits. Architecture 2, developed for future possible processes with p-type buried layer isolation, features pixel-level time measurement and storage. A chip-level Time-to-Digital Converter (TDC) is used and the part of Voltage-Controlled Delay Line (VCDL) is copied in each pixel to get a high time resolution. The TOA resolution is estimated to be 4.2 ns and the TOT resolution 8.4 ns. COFFEE3, with a layout size of 3$\times$4 mm$^2$, was manufactured and has undergone preliminary tests. Charge injection tests for analog circuits, and laser tests for full readout chains, confirm that both architectures operate as expected. Next step work will focus on characterizing key performance such as the timing resolution, radiation hardness, and tracking performance of minimum ionising particles.

physics.ins-det↗

arXiv2026

Step 3.5 Flash: Open Frontier-Level Intelligence with 11B Active Parameters

We introduce Step 3.5 Flash, a sparse Mixture-of-Experts (MoE) model that bridges frontier-level agentic intelligence and computational efficiency. We focus on what matters most when building agents: sharp reasoning and fast, reliable execution. Step 3.5 Flash pairs a 196B-parameter foundation with 11B active parameters for efficient inference. It is optimized with interleaved 3:1 sliding-window/full attention and Multi-Token Prediction (MTP-3) to reduce the latency and cost of multi-round agentic interactions. To reach frontier-level intelligence, we design a scalable reinforcement learning framework that combines verifiable signals with preference feedback, while remaining stable under large-scale off-policy training, enabling consistent self-improvement across mathematics, code, and tool use. Step 3.5 Flash demonstrates strong performance across agent, coding, and math tasks, achieving 85.4% on IMO-AnswerBench, 86.4% on LiveCodeBench-v6 (2024.08-2025.05), 88.2% on tau2-Bench, 69.0% on BrowseComp (with context management), and 51.0% on Terminal-Bench 2.0, comparable to frontier models such as GPT-5.2 xHigh and Gemini 3.0 Pro. By redefining the efficiency frontier, Step 3.5 Flash provides a high-density foundation for deploying sophisticated agents in real-world industrial environments.

cs.CL↗

arXiv2026

MILR: Improving Multimodal Image Generation via Test-Time Latent Reasoning

Reasoning-augmented machine learning systems have shown improved performance in various domains, including image generation. However, existing reasoning-based methods for image generation either restrict reasoning to a single modality (image or text) or rely on high-quality reasoning data for fine-tuning. To tackle these limitations, we propose MILR, a test-time method that jointly reasons over image and text in a unified latent vector space. Reasoning in MILR is performed by searching through vector representations of discrete image and text tokens. Practically, this is implemented via the policy gradient method, guided by an image quality critic. We instantiate MILR within the unified multimodal understanding and generation (MUG) framework that natively supports language reasoning before image synthesis and thus facilitates cross-modal reasoning. The intermediate model outputs, which are to be optimized, serve as the unified latent space, enabling MILR to operate entirely at test time. We evaluate MILR on GenEval, T2I-CompBench, and WISE, achieving state-of-the-art results on all benchmarks. Notably, on knowledge-intensive WISE, MILR attains an overall score of 0.63, improving over the baseline by 80%. Our further analysis indicates that joint reasoning in the unified latent space is the key to its strong performance. Moreover, our qualitative studies reveal MILR's non-trivial ability in temporal and cultural reasoning, highlighting the efficacy of our reasoning method.

cs.CV↗

arXiv2025

A Study of Finetuning Video Transformers for Multi-view Geometry Tasks

This paper presents an investigation of vision transformer learning for multi-view geometry tasks, such as optical flow estimation, by fine-tuning video foundation models. Unlike previous methods that involve custom architectural designs and task-specific pretraining, our research finds that general-purpose models pretrained on videos can be readily transferred to multi-view problems with minimal adaptation. The core insight is that general-purpose attention between patches learns temporal and spatial information for geometric reasoning. We demonstrate that appending a linear decoder to the Transformer backbone produces satisfactory results, and iterative refinement can further elevate performance to stateof-the-art levels. This conceptually simple approach achieves top cross-dataset generalization results for optical flow estimation with end-point error (EPE) of 0.69, 1.78, and 3.15 on the Sintel clean, Sintel final, and KITTI datasets, respectively. Our method additionally establishes a new record on the online test benchmark with EPE values of 0.79, 1.88, and F1 value of 3.79. Applications to 3D depth estimation and stereo matching also show strong performance, illustrating the versatility of video-pretrained models in addressing geometric vision tasks.

cs.CV↗

arXiv2025

NEP: Autoregressive Image Editing via Next Editing Token Prediction

Text-guided image editing involves modifying a source image based on a language instruction and, typically, requires changes to only small local regions. However, existing approaches generate the entire target image rather than selectively regenerate only the intended editing areas. This results in (1) unnecessary computational costs and (2) a bias toward reconstructing non-editing regions, which compromises the quality of the intended edits. To resolve these limitations, we propose to formulate image editing as Next Editing-token Prediction (NEP) based on autoregressive image generation, where only regions that need to be edited are regenerated, thus avoiding unintended modification to the non-editing areas. To enable any-region editing, we propose to pre-train an any-order autoregressive text-to-image (T2I) model. Once trained, it is capable of zero-shot image editing and can be easily adapted to NEP for image editing, which achieves a new state-of-the-art on widely used image editing benchmarks. Moreover, our model naturally supports test-time scaling (TTS) through iteratively refining its generation in a zero-shot manner. The project page is: https://nep-bigai.github.io/

cs.CV↗

arXiv2025

Falcon: Fast Visuomotor Policies via Partial Denoising

Diffusion policies are widely adopted in complex visuomotor tasks for their ability to capture multimodal action distributions. However, the multiple sampling steps required for action generation significantly harm real-time inference efficiency, which limits their applicability in real-time decision-making scenarios. Existing acceleration techniques either require retraining or degrade performance under low sampling steps. Here we propose Falcon, which mitigates this speed-performance trade-off and achieves further acceleration. The core insight is that visuomotor tasks exhibit sequential dependencies between actions. Falcon leverages this by reusing partially denoised actions from historical information rather than sampling from Gaussian noise at each step. By integrating current observations, Falcon reduces sampling steps while preserving performance. Importantly, Falcon is a training-free algorithm that can be applied as a plug-in to further improve decision efficiency on top of existing acceleration techniques. We validated Falcon in 48 simulated environments and 2 real-world robot experiments. demonstrating a 2-7x speedup with negligible performance degradation, offering a promising direction for efficient visuomotor policy design.

cs.RO↗

arXiv2024

Zeroth-Order Hard-Thresholding: Gradient Error vs. Expansivity

$\ell_0$ constrained optimization is prevalent in machine learning, particularly for high-dimensional problems, because it is a fundamental approach to achieve sparse learning. Hard-thresholding gradient descent is a dominant technique to solve this problem. However, first-order gradients of the objective function may be either unavailable or expensive to calculate in a lot of real-world problems, where zeroth-order (ZO) gradients could be a good surrogate. Unfortunately, whether ZO gradients can work with the hard-thresholding operator is still an unsolved problem. To solve this puzzle, in this paper, we focus on the $\ell_0$ constrained black-box stochastic optimization problems, and propose a new stochastic zeroth-order gradient hard-thresholding (SZOHT) algorithm with a general ZO gradient estimator powered by a novel random support sampling. We provide the convergence analysis of SZOHT under standard assumptions. Importantly, we reveal a conflict between the deviation of ZO estimators and the expansivity of the hard-thresholding operator, and provide a theoretical minimal value of the number of random directions in ZO gradients. In addition, we find that the query complexity of SZOHT is independent or weakly dependent on the dimensionality under different settings. Finally, we illustrate the utility of our method on a portfolio optimization problem as well as black-box adversarial attacks.

cs.LG↗

arXiv2023

Randomized Quantization: A Generic Augmentation for Data Agnostic Self-supervised Learning

Self-supervised representation learning follows a paradigm of withholding some part of the data and tasking the network to predict it from the remaining part. Among many techniques, data augmentation lies at the core for creating the information gap. Towards this end, masking has emerged as a generic and powerful tool where content is withheld along the sequential dimension, e.g., spatial in images, temporal in audio, and syntactic in language. In this paper, we explore the orthogonal channel dimension for generic data augmentation by exploiting precision redundancy. The data for each channel is quantized through a non-uniform quantizer, with the quantized value sampled randomly within randomly sampled quantization bins. From another perspective, quantization is analogous to channel-wise masking, as it removes the information within each bin, but preserves the information across bins. Our approach significantly surpasses existing generic data augmentation methods, while showing on par performance against modality-specific augmentations. We comprehensively evaluate our approach on vision, audio, 3D point clouds, as well as the DABS benchmark which is comprised of various data modalities. The code is available at https: //github.com/microsoft/random_quantize.

cs.CV↗

arXiv2023

Compete to Win: Enhancing Pseudo Labels for Barely-supervised Medical Image Segmentation

This study investigates barely-supervised medical image segmentation where only few labeled data, i.e., single-digit cases are available. We observe the key limitation of the existing state-of-the-art semi-supervised solution cross pseudo supervision is the unsatisfactory precision of foreground classes, leading to a degenerated result under barely-supervised learning. In this paper, we propose a novel Compete-to-Win method (ComWin) to enhance the pseudo label quality. In contrast to directly using one model's predictions as pseudo labels, our key idea is that high-quality pseudo labels should be generated by comparing multiple confidence maps produced by different networks to select the most confident one (a compete-to-win strategy). To further refine pseudo labels at near-boundary areas, an enhanced version of ComWin, namely, ComWin+, is proposed by integrating a boundary-aware enhancement module. Experiments show that our method can achieve the best performance on three public medical image datasets for cardiac structure segmentation, pancreas segmentation and colon tumor segmentation, respectively. The source code is now available at https://github.com/Huiimin5/comwin.

cs.CV↗

arXiv2023

Exploring Feature Representation Learning for Semi-supervised Medical Image Segmentation

This paper presents a simple yet effective two-stage framework for semi-supervised medical image segmentation. Unlike prior state-of-the-art semi-supervised segmentation methods that predominantly rely on pseudo supervision directly on predictions, such as consistency regularization and pseudo labeling, our key insight is to explore the feature representation learning with labeled and unlabeled (i.e., pseudo labeled) images to regularize a more compact and better-separated feature space, which paves the way for low-density decision boundary learning and therefore enhances the segmentation performance. A stage-adaptive contrastive learning method is proposed, containing a boundary-aware contrastive loss that takes advantage of the labeled images in the first stage, as well as a prototype-aware contrastive loss to optimize both labeled and pseudo labeled images in the second stage. To obtain more accurate prototype estimation, which plays a critical role in prototype-aware contrastive learning, we present an aleatoric uncertainty-aware method, namely AUA, to generate higher-quality pseudo labels. AUA adaptively regularizes prediction consistency by taking advantage of image ambiguity, which, given its significance, is under-explored by existing works. Our method achieves the best results on three public medical image segmentation benchmarks.

cs.CV↗

arXiv2022

FedMix: Mixed Supervised Federated Learning for Medical Image Segmentation

The purpose of federated learning is to enable multiple clients to jointly train a machine learning model without sharing data. However, the existing methods for training an image segmentation model have been based on an unrealistic assumption that the training set for each local client is annotated in a similar fashion and thus follows the same image supervision level. To relax this assumption, in this work, we propose a label-agnostic unified federated learning framework, named FedMix, for medical image segmentation based on mixed image labels. In FedMix, each client updates the federated model by integrating and effectively making use of all available labeled data ranging from strong pixel-level labels, weak bounding box labels, to weakest image-level class labels. Based on these local models, we further propose an adaptive weight assignment procedure across local clients, where each client learns an aggregation weight during the global model update. Compared to the existing methods, FedMix not only breaks through the constraint of a single level of image supervision, but also can dynamically adjust the aggregation weight of each local client, achieving rich yet discriminative feature representations. To evaluate its effectiveness, experiments have been carried out on two challenging medical image segmentation tasks, i.e., breast tumor segmentation and skin lesion segmentation. The results validate that our proposed FedMix outperforms the state-of-the-art method by a large margin.

cs.CV↗

arXiv2022

Deep Understanding based Multi-Document Machine Reading Comprehension

Most existing multi-document machine reading comprehension models mainly focus on understanding the interactions between the input question and documents, but ignore following two kinds of understandings. First, to understand the semantic meaning of words in the input question and documents from the perspective of each other. Second, to understand the supporting cues for a correct answer from the perspective of intra-document and inter-documents. Ignoring these two kinds of important understandings would make the models oversee some important information that may be helpful for inding correct answers. To overcome this deiciency, we propose a deep understanding based model for multi-document machine reading comprehension. It has three cascaded deep understanding modules which are designed to understand the accurate semantic meaning of words, the interactions between the input question and documents, and the supporting cues for the correct answer. We evaluate our model on two large scale benchmark datasets, namely TriviaQA Web and DuReader. Extensive experiments show that our model achieves state-of-the-art results on both datasets.

cs.CL↗

arXiv2021

Fast and Scalable Adversarial Training of Kernel SVM via Doubly Stochastic Gradients

Adversarial attacks by generating examples which are almost indistinguishable from natural examples, pose a serious threat to learning models. Defending against adversarial attacks is a critical element for a reliable learning system. Support vector machine (SVM) is a classical yet still important learning algorithm even in the current deep learning era. Although a wide range of researches have been done in recent years to improve the adversarial robustness of learning models, but most of them are limited to deep neural networks (DNNs) and the work for kernel SVM is still vacant. In this paper, we aim at kernel SVM and propose adv-SVM to improve its adversarial robustness via adversarial training, which has been demonstrated to be the most promising defense techniques. To the best of our knowledge, this is the first work that devotes to the fast and scalable adversarial training of kernel SVM. Specifically, we first build connection of perturbations of samples between original and kernel spaces, and then give a reduced and equivalent formulation of adversarial training of kernel SVM based on the connection. Next, doubly stochastic gradients (DSG) based on two unbiased stochastic approximations (i.e., one is on training points and another is on random features) are applied to update the solution of our objective function. Finally, we prove that our algorithm optimized by DSG converges to the optimal solution at the rate of O(1/t) under the constant and diminishing stepsizes. Comprehensive experimental results show that our adversarial training algorithm enjoys robustness against various attacks and meanwhile has the similar efficiency and scalability with classical DSG algorithm.

cs.LG↗