arXiv subjects

Zheng Zhang

Publications and source records attributed to Zheng Zhang.

At least 199 records · Page 11Linked to original sources

arXiv2025

TreePO: Bridging the Gap of Policy Optimization and Efficacy and Inference Efficiency with Heuristic Tree-based Modeling

Recent advancements in aligning large language models via reinforcement learning have achieved remarkable gains in solving complex reasoning problems, but at the cost of expensive on-policy rollouts and limited exploration of diverse reasoning paths. In this work, we introduce TreePO, involving a self-guided rollout algorithm that views sequence generation as a tree-structured searching process. Composed of dynamic tree sampling policy and fixed-length segment decoding, TreePO leverages local uncertainty to warrant additional branches. By amortizing computation across common prefixes and pruning low-value paths early, TreePO essentially reduces the per-update compute burden while preserving or enhancing exploration diversity. Key contributions include: (1) a segment-wise sampling algorithm that alleviates the KV cache burden through contiguous segments and spawns new branches along with an early-stop mechanism; (2) a tree-based segment-level advantage estimation that considers both global and local proximal policy optimization. and (3) analysis on the effectiveness of probability and quality-driven dynamic divergence and fallback strategy. We empirically validate the performance gain of TreePO on a set reasoning benchmarks and the efficiency saving of GPU hours from 22\% up to 43\% of the sampling design for the trained models, meanwhile showing up to 40\% reduction at trajectory-level and 35\% at token-level sampling compute for the existing models. While offering a free lunch of inference efficiency, TreePO reveals a practical path toward scaling RL-based post-training with fewer samples and less compute. Home page locates at https://m-a-p.ai/TreePO.

cs.LG↗

arXiv2025

ContrastAlign: Toward Robust BEV Feature Alignment via Contrastive Learning for Multi-Modal 3D Object Detection

In the field of 3D object detection tasks, fusing heterogeneous features from LiDAR and camera sensors into a unified Bird's Eye View (BEV) representation is a widely adopted paradigm. However, existing methods often suffer from imprecise sensor calibration, leading to feature misalignment in LiDAR-camera BEV fusion. Moreover, such inaccuracies cause errors in depth estimation for the camera branch, aggravating misalignment between LiDAR and camera BEV features. In this work, we propose a novel ContrastAlign approach that utilizes contrastive learning to enhance the alignment of heterogeneous modalities, thereby improving the robustness of the fusion process. Specifically, our approach comprises three key components: (1) the L-Instance module, which extracts LiDAR instance features within the LiDAR BEV features; (2) the C-Instance module, which predicts camera instance features through Region of Interest (RoI) pooling on the camera BEV features; (3) the InstanceFusion module, which employs contrastive learning to generate consistent instance features across heFterogeneous modalities. Subsequently, we use graph matching to calculate the similarity between the neighboring camera instance features and the similarity instance features to complete the alignment of instance features. Our method achieves SOTA performance, with an mAP of 71.5%, surpassing GraphBEV by 1.4% on the nuScenes val set. Importantly, our method excels BEVFusion under conditions with spatial & temporal misalignment noise, improving mAP by 1.4% and 11.1% on nuScenes dataset. Notably, on the Argoverse2 dataset, ContrastAlign outperforms GraphBEV by 1.0% in mAP, indicating that the farther the distance, the more severe the feature misalignment and the more effective.

cs.CV↗

arXiv2025

Ultra Memory-Efficient On-FPGA Training of Transformers via Tensor-Compressed Optimization

Transformer models have achieved state-of-the-art performance across a wide range of machine learning tasks. There is growing interest in training transformers on resource-constrained edge devices due to considerations such as privacy, domain adaptation, and on-device scientific machine learning. However, the significant computational and memory demands required for transformer training often exceed the capabilities of an edge device. Leveraging low-rank tensor compression, this paper presents the first on-FPGA accelerator for end-to-end transformer training. On the algorithm side, we present a bi-directional contraction flow for tensorized transformer training, significantly reducing the computational FLOPS and intra-layer memory costs compared to existing tensor operations. On the hardware side, we store all highly compressed model parameters and gradient information on chip, creating an on-chip-memory-only framework for each stage in training. This reduces off-chip communication and minimizes latency and energy costs. Additionally, we implement custom computing kernels for each training stage and employ intra-layer parallelism and pipe-lining to further enhance run-time and memory efficiency. Through experiments on transformer models within $36.7$ to $93.5$ MB using FP-32 data formats on the ATIS dataset, our tensorized FPGA accelerator could conduct single-batch end-to-end training on the AMD Alevo U50 FPGA, with a memory budget of less than $6$-MB BRAM and $22.5$-MB URAM. Compared to uncompressed training on the NVIDIA RTX 3090 GPU, our on-FPGA training achieves a memory reduction of $30\times$ to $51\times$. Our FPGA accelerator also achieves up to $3.6\times$ less energy cost per epoch compared with tensor Transformer training on an NVIDIA RTX 3090 GPU.

cs.LG↗

arXiv2025

JanusNet: Hierarchical Slice-Block Shuffle and Displacement for Semi-Supervised 3D Multi-Organ Segmentation

Limited by the scarcity of training samples and annotations, weakly supervised medical image segmentation often employs data augmentation to increase data diversity, while randomly mixing volumetric blocks has demonstrated strong performance. However, this approach disrupts the inherent anatomical continuity of 3D medical images along orthogonal axes, leading to severe structural inconsistencies and insufficient training in challenging regions, such as small-sized organs, etc. To better comply with and utilize human anatomical information, we propose JanusNet}, a data augmentation framework for 3D medical data that globally models anatomical continuity while locally focusing on hard-to-segment regions. Specifically, our Slice-Block Shuffle step performs aligned shuffling of same-index slice blocks across volumes along a random axis, while preserving the anatomical context on planes perpendicular to the perturbation axis. Concurrently, the Confidence-Guided Displacement step uses prediction reliability to replace blocks within each slice, amplifying signals from difficult areas. This dual-stage, axis-aligned framework is plug-and-play, requiring minimal code changes for most teacher-student schemes. Extensive experiments on the Synapse and AMOS datasets demonstrate that JanusNet significantly surpasses state-of-the-art methods, achieving, for instance, a 4% DSC gain on the Synapse dataset with only 20% labeled data.

cs.CV↗

arXiv2025

Nexus-INR: Diverse Knowledge-guided Arbitrary-Scale Multimodal Medical Image Super-Resolution

Arbitrary-resolution super-resolution (ARSR) provides crucial flexibility for medical image analysis by adapting to diverse spatial resolutions. However, traditional CNN-based methods are inherently ill-suited for ARSR, as they are typically designed for fixed upsampling factors. While INR-based methods overcome this limitation, they still struggle to effectively process and leverage multi-modal images with varying resolutions and details. In this paper, we propose Nexus-INR, a Diverse Knowledge-guided ARSR framework, which employs varied information and downstream tasks to achieve high-quality, adaptive-resolution medical image super-resolution. Specifically, Nexus-INR contains three key components. A dual-branch encoder with an auxiliary classification task to effectively disentangle shared anatomical structures and modality-specific features; a knowledge distillation module using cross-modal attention that guides low-resolution modality reconstruction with high-resolution reference, enhanced by self-supervised consistency loss; an integrated segmentation module that embeds anatomical semantics to improve both reconstruction quality and downstream segmentation performance. Experiments on the BraTS2020 dataset for both super-resolution and downstream segmentation demonstrate that Nexus-INR outperforms state-of-the-art methods across various metrics.

eess.IV↗

arXiv2025

Contrastive Cross-Bag Augmentation for Multiple Instance Learning-based Whole Slide Image Classification

Recent pseudo-bag augmentation methods for Multiple Instance Learning (MIL)-based Whole Slide Image (WSI) classification sample instances from a limited number of bags, resulting in constrained diversity. To address this issue, we propose Contrastive Cross-Bag Augmentation ($C^2Aug$) to sample instances from all bags with the same class to increase the diversity of pseudo-bags. However, introducing new instances into the pseudo-bag increases the number of critical instances (e.g., tumor instances). This increase results in a reduced occurrence of pseudo-bags containing few critical instances, thereby limiting model performance, particularly on test slides with small tumor areas. To address this, we introduce a bag-level and group-level contrastive learning framework to enhance the discrimination of features with distinct semantic meanings, thereby improving model performance. Experimental results demonstrate that $C^2Aug$ consistently outperforms state-of-the-art approaches across multiple evaluation metrics.

cs.CV↗

arXiv2025

Seed Diffusion: A Large-Scale Diffusion Language Model with High-Speed Inference

We present Seed Diffusion Preview, a large-scale language model based on discrete-state diffusion, offering remarkably fast inference speed. Thanks to non-sequential, parallel generation, discrete diffusion models provide a notable speedup to mitigate the inherent latency of token-by-token decoding, as demonstrated recently (e.g., Mercury Coder, Gemini Diffusion). Seed Diffusion Preview achieves an inference speed of 2,146 token/s over H20 GPUs while maintaining competitive performance across a sweep of standard code evaluation benchmarks, significantly faster than contemporary Mercury and Gemini Diffusion, establishing new state of the art on the speed-quality Pareto frontier for code models.

cs.CL↗

arXiv2025

Uni-Layout: Integrating Human Feedback in Unified Layout Generation and Evaluation

Layout generation plays a crucial role in enhancing both user experience and design efficiency. However, current approaches suffer from task-specific generation capabilities and perceptually misaligned evaluation metrics, leading to limited applicability and ineffective measurement. In this paper, we propose \textit{Uni-Layout}, a novel framework that achieves unified generation, human-mimicking evaluation and alignment between the two. For universal generation, we incorporate various layout tasks into a single taxonomy and develop a unified generator that handles background or element contents constrained tasks via natural language prompts. To introduce human feedback for the effective evaluation of layouts, we build \textit{Layout-HF100k}, the first large-scale human feedback dataset with 100,000 expertly annotated layouts. Based on \textit{Layout-HF100k}, we introduce a human-mimicking evaluator that integrates visual and geometric information, employing a Chain-of-Thought mechanism to conduct qualitative assessments alongside a confidence estimation module to yield quantitative measurements. For better alignment between the generator and the evaluator, we integrate them into a cohesive system by adopting Dynamic-Margin Preference Optimization (DMPO), which dynamically adjusts margins based on preference strength to better align with human judgments. Extensive experiments show that \textit{Uni-Layout} significantly outperforms both task-specific and general-purpose methods. Our code is publicly available at https://github.com/JD-GenX/Uni-Layout.

cs.CV↗

arXiv2025

H$_2$O$_2$RAM: A High-Performance Hierarchical Doubly Oblivious RAM

The combination of Oblivious RAM (ORAM) with Trusted Execution Environments (TEE) has found numerous real-world applications due to their complementary nature. TEEs alleviate the performance bottlenecks of ORAM, such as network bandwidth and roundtrip latency, and ORAM provides general-purpose protection for TEE applications against attacks exploiting memory access patterns. The defining property of this combination, which sets it apart from traditional ORAM designs, is its ability to ensure that memory accesses, both inside and outside of TEEs, are made oblivious, thus termed doubly oblivious RAM (O$_2$RAM). Efforts to develop O$_2$RAM with enhanced performance are ongoing. In this work, we propose H$_2$O$_2$RAM, a high-performance doubly oblivious RAM construction. The distinguishing feature of our approach, compared to the existing tree-based doubly oblivious designs, is its first adoption of the hierarchical framework that enjoys inherently better data locality and parallelization. While the latest hierarchical solution, FutORAMa, achieves concrete efficiency in the classic client-server model by leveraging a relaxed assumption of sublinear-sized client-side private memory, adapting it to our scenario poses challenges due to the conflict between this relaxed assumption and our doubly oblivious requirement. To this end, we introduce several new efficient oblivious components to build a high-performance hierarchical O$_2$RAM (H$_2$O$_2$RAM). We implement our design and evaluate it on various scenarios. The results indicate that H$_2$O$_2$RAM reduces execution time by up to $\sim 10^3$ times and saves memory usage by $5\sim44$ times compared to state-of-the-art solutions.

cs.CR↗

arXiv2025

GRAG: Graph Retrieval-Augmented Generation

Naive Retrieval-Augmented Generation (RAG) focuses on individual documents during retrieval and, as a result, falls short in handling networked documents which are very popular in many applications such as citation graphs, social media, and knowledge graphs. To overcome this limitation, we introduce Graph Retrieval-Augmented Generation (GRAG), which tackles the fundamental challenges in retrieving textual subgraphs and integrating the joint textual and topological information into Large Language Models (LLMs) to enhance its generation. To enable efficient textual subgraph retrieval, we propose a novel divide-and-conquer strategy that retrieves the optimal subgraph structure in linear time. To achieve graph context-aware generation, incorporate textual graphs into LLMs through two complementary views-the text view and the graph view-enabling LLMs to more effectively comprehend and utilize the graph context. Extensive experiments on graph reasoning benchmarks demonstrate that in scenarios requiring multi-hop reasoning on textual graphs, our GRAG approach significantly outperforms current state-of-the-art RAG methods. Our datasets as well as codes of GRAG are available at https://github.com/HuieL/GRAG.

cs.LG↗

arXiv2025

Infinitesimal Torelli problems for special Gushel-Mukai and related Fano threefolds: Hodge theoretical and categorical perspectives

We investigate infinitesimal Torelli problems for some of the Fano threefolds of the following two types: (a) those which can be described as zero loci of sections of vector bundles on Grassmannians (for instance, ordinary Gushel-Mukai threefolds), and (b) double covers of rigid Fano threefolds branched along a $K3$ surface (such as, special Gushel-Mukai threefolds). The differential of the period map for ordinary Gushel-Mukai threefolds has been studied by Debarre, Iliev and Manivel; in particular, it has a $2$-dimensional kernel. The main result of this paper is that the invariant part of the infinitesimal period map for a special Gushel-Mukai threefold is injective. We prove this result using a Hodge theoretical argument as well as a categorical method. Through similar approaches, we also study infinitesimal Torelli problems for prime Fano threefolds with genus $7$, $8$, $9$, $10$, $12$ (type (a)) and for special Verra threefolds (type (b)). Furthermore, a geometric description of the kernel of the differential of the period maps for Gushel-Mukai threefolds (and for prime Fano threefolds of genus $8$) is given via a Bridgeland moduli space in the Kuznetsov components.

math.AG↗

arXiv2025

Perception, Reason, Think, and Plan: A Survey on Large Multimodal Reasoning Models

Reasoning lies at the heart of intelligence, shaping the ability to make decisions, draw conclusions, and generalize across domains. In artificial intelligence, as systems increasingly operate in open, uncertain, and multimodal environments, reasoning becomes essential for enabling robust and adaptive behavior. Large Multimodal Reasoning Models (LMRMs) have emerged as a promising paradigm, integrating modalities such as text, images, audio, and video to support complex reasoning capabilities and aiming to achieve comprehensive perception, precise understanding, and deep reasoning. As research advances, multimodal reasoning has rapidly evolved from modular, perception-driven pipelines to unified, language-centric frameworks that offer more coherent cross-modal understanding. While instruction tuning and reinforcement learning have improved model reasoning, significant challenges remain in omni-modal generalization, reasoning depth, and agentic behavior. To address these issues, we present a comprehensive and structured survey of multimodal reasoning research, organized around a four-stage developmental roadmap that reflects the field's shifting design philosophies and emerging capabilities. First, we review early efforts based on task-specific modules, where reasoning was implicitly embedded across stages of representation, alignment, and fusion. Next, we examine recent approaches that unify reasoning into multimodal LLMs, with advances such as Multimodal Chain-of-Thought (MCoT) and multimodal reinforcement learning enabling richer and more structured reasoning chains. Finally, drawing on empirical insights from challenging benchmarks and experimental cases of OpenAI O3 and O4-mini, we discuss the conceptual direction of native large multimodal reasoning models (N-LMRMs), which aim to support scalable, agentic, and adaptive reasoning and planning in complex, real-world environments.

cs.CV↗

arXiv2025

A New and Efficient Debiased Estimation of General Treatment Models by Balanced Neural Networks Weighting

Estimation and inference of treatment effects under unconfounded treatment assignments often suffer from bias and the `curse of dimensionality' due to the nonparametric estimation of nuisance parameters for high-dimensional confounders. Although debiased state-of-the-art methods have been proposed for binary treatments under particular treatment models, they can be unstable for small sample sizes. Moreover, directly extending them to general treatment models can lead to computational complexity. We propose a balanced neural networks weighting method for general treatment models, which leverages deep neural networks to alleviate the curse of dimensionality while retaining optimal covariate balance through calibration, thereby achieving debiased and robust estimation. Our method accommodates a wide range of treatment models, including average, quantile, distributional, and asymmetric least squares treatment effects, for discrete, continuous, and mixed treatments. Under regularity conditions, we show that our estimator achieves rate double robustness and $\sqrt{N}$-asymptotic normality, and its asymptotic variance achieves the semiparametric efficiency bound. We further develop a statistical inference procedure based on weighted bootstrap, which avoids estimating the efficient influence/score functions. Simulation results reveal that the proposed method consistently outperforms existing alternatives, especially when the sample size is small. Applications to the 401(k) dataset and the Mother's Significant Features dataset further illustrate the practical value of the method for estimating both average and quantile treatment effects under binary and continuous treatments, respectively.

stat.ME↗

arXiv2025

MPipeMoE: Memory Efficient MoE for Pre-trained Models with Adaptive Pipeline Parallelism

Recently, Mixture-of-Experts (MoE) has become one of the most popular techniques to scale pre-trained models to extraordinarily large sizes. Dynamic activation of experts allows for conditional computation, increasing the number of parameters of neural networks, which is critical for absorbing the vast amounts of knowledge available in many deep learning areas. However, despite the existing system and algorithm optimizations, there are significant challenges to be tackled when it comes to the inefficiencies of communication and memory consumption. In this paper, we present the design and implementation of MPipeMoE, a high-performance library that accelerates MoE training with adaptive and memory-efficient pipeline parallelism. Inspired by that the MoE training procedure can be divided into multiple independent sub-stages, we design adaptive pipeline parallelism with an online algorithm to configure the granularity of the pipelining. Further, we analyze the memory footprint breakdown of MoE training and identify that activations and temporary buffers are the primary contributors to the overall memory footprint. Toward memory efficiency, we propose memory reusing strategies to reduce memory requirements by eliminating memory redundancies, and develop an adaptive selection component to determine the optimal strategy that considers both hardware capacities and model characteristics at runtime. We implement MPipeMoE upon PyTorch and evaluate it with common MoE models in a physical cluster consisting of 8 NVIDIA DGX A100 servers. Compared with the state-of-art approach, MPipeMoE achieves up to 2.8x speedup and reduces memory footprint by up to 47% in training large models.

cs.DC↗

arXiv2025

MCFuser: High-Performance and Rapid Fusion of Memory-Bound Compute-Intensive Operators

Operator fusion, a key technique to improve data locality and alleviate GPU memory bandwidth pressure, often fails to extend to the fusion of multiple compute-intensive operators due to saturated computation throughput. However, the dynamicity of tensor dimension sizes could potentially lead to these operators becoming memory-bound, necessitating the generation of fused kernels, a task hindered by limited search spaces for fusion strategies, redundant memory access, and prolonged tuning time, leading to sub-optimal performance and inefficient deployment. We introduce MCFuser, a pioneering framework designed to overcome these obstacles by generating high-performance fused kernels for what we define as memory-bound compute-intensive (MBCI) operator chains. Leveraging high-level tiling expressions to delineate a comprehensive search space, coupled with Directed Acyclic Graph (DAG) analysis to eliminate redundant memory accesses, MCFuser streamlines kernel optimization. By implementing guidelines to prune the search space and incorporating an analytical performance model with a heuristic search, MCFuser not only significantly accelerates the tuning process but also demonstrates superior performance. Benchmarked against leading compilers like Ansor on NVIDIA A100 and RTX3080 GPUs, MCFuser achieves up to a 5.9x speedup in kernel performance and outpaces other baselines while reducing tuning time by over 70-fold, showcasing its agility.

cs.DC↗

arXiv2025

PsyLite Technical Report

With the rapid development of digital technology, AI-driven psychological counseling has gradually become an important research direction in the field of mental health. However, existing models still have deficiencies in dialogue safety, detailed scenario handling, and lightweight deployment. To address these issues, this study proposes PsyLite, a lightweight psychological counseling large language model agent developed based on the base model InternLM2.5-7B-chat. Through a two-stage training strategy (hybrid distillation data fine-tuning and ORPO preference optimization), PsyLite enhances the model's deep-reasoning ability, psychological counseling ability, and safe dialogue ability. After deployment using Ollama and Open WebUI, a custom workflow is created with Pipelines. An innovative conditional RAG is designed to introduce crosstalk humor elements at appropriate times during psychological counseling to enhance user experience and decline dangerous requests to strengthen dialogue safety. Evaluations show that PsyLite outperforms the baseline models in the Chinese general evaluation (CEval), psychological counseling professional evaluation (CPsyCounE), and dialogue safety evaluation (SafeDialBench), particularly in psychological counseling professionalism (CPsyCounE score improvement of 47.6\%) and dialogue safety (\safe{} score improvement of 2.4\%). Additionally, the model uses quantization technology (GGUF q4\_k\_m) to achieve low hardware deployment (5GB memory is sufficient for operation), providing a feasible solution for psychological counseling applications in resource-constrained environments.

cs.AI↗

arXiv2025

Kimi-VL Technical Report

We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities - all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, and multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking-2506. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), the latest model exhibits strong long-horizon reasoning capabilities (64.0 on MMMU, 46.3 on MMMU-Pro, 56.9 on MathVision, 80.1 on MathVista, 65.2 on VideoMMMU) while obtaining robust general abilities. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.

cs.CV↗

arXiv2025

Joint Optimization of Age of Information and Energy Consumption in NR-V2X System based on Deep Reinforcement Learning

Autonomous driving may be the most important application scenario of next generation, the development of wireless access technologies enabling reliable and low-latency vehicle communication becomes crucial. To address this, 3GPP has developed Vehicle-to-Everything (V2X) specifications based on 5G New Radio (NR) technology, where Mode 2 Side-Link (SL) communication resembles Mode 4 in LTE-V2X, allowing direct communication between vehicles. This supplements SL communication in LTE-V2X and represents the latest advancement in cellular V2X (C-V2X) with improved performance of NR-V2X. However, in NR-V2X Mode 2, resource collisions still occur, and thus degrade the age of information (AOI). Therefore, a interference cancellation method is employed to mitigate this impact by combining NR-V2X with Non-Orthogonal multiple access (NOMA) technology. In NR-V2X, when vehicles select smaller resource reservation interval (RRI), higher-frequency transmissions take ore energy to reduce AoI. Hence, it is important to jointly consider AoI and communication energy consumption based on NR-V2X communication. Then, we formulate such an optimization problem and employ the Deep Reinforcement Learning (DRL) algorithm to compute the optimal transmission RRI and transmission power for each transmitting vehicle to reduce the energy consumption of each transmitting vehicle and the AoI of each receiving vehicle. Extensive simulations have demonstrated the performance of our proposed algorithm.

cs.LG↗