arXiv subjects

Bo Zhao

Publications and source records attributed to Bo Zhao.

At least 145 records · Page 8Linked to original sources

arXiv2025

Megrez2 Technical Report

We present Megrez2, a novel lightweight and high-performance language model architecture optimized for device native deployment. Megrez2 introduces a novel cross-layer expert sharing mechanism, which significantly reduces total parameter count by reusing expert modules across adjacent transformer layers while maintaining most of the model's capacity. It also incorporates pre-gated routing, enabling memory-efficient expert loading and faster inference. As the first instantiation of the Megrez2 architecture, we introduce the Megrez2-Preview model, which is pre-trained on a 5-trillion-token corpus and further enhanced through supervised fine-tuning and reinforcement learning with verifiable rewards. With only 3B activated and 7.5B stored parameters, Megrez2-Preview demonstrates competitive or superior performance compared to larger models on a wide range of tasks, including language understanding, instruction following, mathematical reasoning, and code generation. These results highlight the effectiveness of the Megrez2 architecture to achieve a balance between accuracy, efficiency, and deployability, making it a strong candidate for real-world, resource-constrained applications.

cs.CL↗

arXiv2025

STI-Bench: Are MLLMs Ready for Precise Spatial-Temporal World Understanding?

The use of Multimodal Large Language Models (MLLMs) as an end-to-end solution for Embodied AI and Autonomous Driving has become a prevailing trend. While MLLMs have been extensively studied for visual semantic understanding tasks, their ability to perform precise and quantitative spatial-temporal understanding in real-world applications remains largely unexamined, leading to uncertain prospects. To evaluate models' Spatial-Temporal Intelligence, we introduce STI-Bench, a benchmark designed to evaluate MLLMs' spatial-temporal understanding through challenging tasks such as estimating and predicting the appearance, pose, displacement, and motion of objects. Our benchmark encompasses a wide range of robot and vehicle operations across desktop, indoor, and outdoor scenarios. The extensive experiments reveals that the state-of-the-art MLLMs still struggle in real-world spatial-temporal understanding, especially in tasks requiring precise distance estimation and motion analysis.

cs.CV↗

arXiv2025

MMCR: Benchmarking Cross-Source Reasoning in Scientific Papers

Fully comprehending scientific papers by machines reflects a high level of Artificial General Intelligence, requiring the ability to reason across fragmented and heterogeneous sources of information, presenting a complex and practically significant challenge. While Vision-Language Models (VLMs) have made remarkable strides in various tasks, particularly those involving reasoning with evidence source from single image or text page, their ability to use cross-source information for reasoning remains an open problem. This work presents MMCR, a high-difficulty benchmark designed to evaluate VLMs' capacity for reasoning with cross-source information from scientific papers. The benchmark comprises 276 high-quality questions, meticulously annotated by humans across 7 subjects and 10 task types. Experiments with 18 VLMs demonstrate that cross-source reasoning presents a substantial challenge for existing models. Notably, even the top-performing model, GPT-4o, achieved only 48.55% overall accuracy, with only 20% accuracy in multi-table comprehension tasks, while the second-best model, Qwen2.5-VL-72B, reached 39.86% overall accuracy. Furthermore, we investigated the impact of the Chain-of-Thought (CoT) technique on cross-source reasoning and observed a detrimental effect on small models, whereas larger models demonstrated substantially enhanced performance. These results highlight the pressing need to develop VLMs capable of effectively utilizing cross-source information for reasoning.

cs.CL↗

arXiv2025

Task-Aware KV Compression For Cost-Effective Long Video Understanding

Long-video understanding (LVU) remains a severe challenge for existing multimodal large language models (MLLMs), primarily due to the prohibitive computational cost. Recent approaches have explored KV compression to mitigate this issue, but they often suffer from significant information loss at high compression ratios. In this paper, we introduce Video-X^2L, which flexibly preserves critical video information for each LVU task. Video-X^2L involves two key operations. The first one is called bi-level KV compression. During the MLLM's pre-filling stage, Video-X^2L generates two types of compressed KVs: low-compression KVs (L-KVs) to capture fine-grained video details and high-compression KVs (H-KVs) to offer compact video representations. The second one is called selective KV re-loading. During the MLLM's decoding stage, Video-X^2L selectively re-loads L-KVs for the most critical video chunks while using H-KVs for other less important ones. This allows the MLLM to fully utilize task-specific information while maintaining the overall compactness. Video-X^2L is simple yet effective: it is free from additional training and directly compatible with existing KV-compressible MLLMs. We evaluate Video-X^2L with a variety of popular LVU benchmarks, including VideoMME, MLVU, LongVideoBench, and VNBench. Our experiment result shows that Video-X^2L outperforms existing KV-compression methods by a huge advantage while substantially saving the computation cost.

cs.CV↗

arXiv2025

Video-XL-2: Towards Very Long-Video Understanding Through Task-Aware KV Sparsification

Multi-modal large language models (MLLMs) models have made significant progress in video understanding over the past few years. However, processing long video inputs remains a major challenge due to high memory and computational costs. This makes it difficult for current models to achieve both strong performance and high efficiency in long video understanding. To address this challenge, we propose Video-XL-2, a novel MLLM that delivers superior cost-effectiveness for long-video understanding based on task-aware KV sparsification. The proposed framework operates with two key steps: chunk-based pre-filling and bi-level key-value decoding. Chunk-based pre-filling divides the visual token sequence into chunks, applying full attention within each chunk and sparse attention across chunks. This significantly reduces computational and memory overhead. During decoding, bi-level key-value decoding selectively reloads either dense or sparse key-values for each chunk based on its relevance to the task. This approach further improves memory efficiency and enhances the model's ability to capture fine-grained information. Video-XL-2 achieves state-of-the-art performance on various long video understanding benchmarks, outperforming existing open-source lightweight models. It also demonstrates exceptional efficiency, capable of processing over 10,000 frames on a single NVIDIA A100 (80GB) GPU and thousands of frames in just a few seconds.

cs.CV↗

arXiv2025

Image Captions are Natural Prompts for Text-to-Image Models

With the rapid development of Artificial Intelligence Generated Content (AIGC), it has become a common practice to train models on synthetic data due to data-scarcity and privacy leakage problems. Owing to massive and diverse information conveyed in real images, it is challenging for text-to-image generative models to synthesize informative training data with hand-crafted prompts. Considering the impressive ability of large generative models, could such models directly synthesize good training images for prediction tasks with proper prompts? We offer an affirmative response to this question by proposing a simple yet effective method, validated through ImageNet classification. Specifically, we caption each real image with the advanced captioning model to obtain informative and faithful prompts that extract class-relevant information and clarify the polysemy of class names. The image captions and class names are concatenated to prompt generative models for training image synthesis. We show that this simple caption incorporation significantly boosts the informativeness of synthetic data therefore enhancing downstream model generalization. More importantly, besides improvements in data augmentation and privacy preservation, our experiments demonstrate that synthesized images can exceed real data in terms of out-of-distribution robustness.

cs.CV↗

arXiv2025

Additive Manufacturing of Lunar Regolith for Reconfigurable Building Blocks toward Lunar Habitation

Utilizing locally available materials is a crucial step towards sustainable planetary habitation. Lunar regolith has gained tremendous interest in additive manufacturing in the past decades. However, due to the constrained manufacturing facilities and materials on the moon, many existing additive manufacturing methods are not suitable for practical on-site manufacturing. Here, we envision that light-based direct sintering of lunar regolith can be a feasible approach. Instead of directly manufacturing large structures, we hypothesize that small-scale, reconfigurable building blocks can be an alternative to form large and complex structures. To verify the feasibility, we conducted laser sintering of lunar regolith simulants as a proof of concept, following a simple theoretical calculation for direct sintering using the light available in space. Different laser processing parameters are investigated to obtain controllable lunar regolith sintering. We further designed Lego-like interlocking bricks that are reconfigurable for different structure assemblies without additional material. Mechanical performance (compressive strength) of sintered cubic blocks is evaluated, showing a peak stress of ~1.5 MPa. We hope this work will inspire other in-space manufacturing techniques and enable low-cost space habitation.

astro-ph.IM↗

arXiv2025

Understanding Mode Connectivity via Parameter Space Symmetry

Neural network minima are often connected by curves along which train and test loss remain nearly constant, a phenomenon known as mode connectivity. While this property has enabled applications such as model merging and fine-tuning, its theoretical explanation remains unclear. We propose a new approach to exploring the connectedness of minima using parameter space symmetry. By linking the topology of symmetry groups to that of the minima, we derive the number of connected components of the minima of linear networks and show that skip connections reduce this number. We then examine when mode connectivity and linear mode connectivity hold or fail, using parameter symmetries which account for a significant part of the minimum. Finally, we provide explicit expressions for connecting curves in the minima induced by symmetry. Using the curvature of these curves, we derive conditions under which linear mode connectivity approximately holds. Our findings highlight the role of continuous symmetries in understanding the neural network loss landscape.

cs.LG↗

arXiv2025

HiDream-I1: A High-Efficient Image Generative Foundation Model with Sparse Diffusion Transformer

Recent advancements in image generative foundation models have prioritized quality improvements but often at the cost of increased computational complexity and inference latency. To address this critical trade-off, we introduce HiDream-I1, a new open-source image generative foundation model with 17B parameters that achieves state-of-the-art image generation quality within seconds. HiDream-I1 is constructed with a new sparse Diffusion Transformer (DiT) structure. Specifically, it starts with a dual-stream decoupled design of sparse DiT with dynamic Mixture-of-Experts (MoE) architecture, in which two separate encoders are first involved to independently process image and text tokens. Then, a single-stream sparse DiT structure with dynamic MoE architecture is adopted to trigger multi-model interaction for image generation in a cost-efficient manner. To support flexiable accessibility with varied model capabilities, we provide HiDream-I1 in three variants: HiDream-I1-Full, HiDream-I1-Dev, and HiDream-I1-Fast. Furthermore, we go beyond the typical text-to-image generation and remould HiDream-I1 with additional image conditions to perform precise, instruction-based editing on given images, yielding a new instruction-based image editing model namely HiDream-E1. Ultimately, by integrating text-to-image generation and instruction-based image editing, HiDream-I1 evolves to form a comprehensive image agent (HiDream-A1) capable of fully interactive image creation and refinement. To accelerate multi-modal AIGC research, we have open-sourced all the codes and model weights of HiDream-I1-Full, HiDream-I1-Dev, HiDream-I1-Fast, HiDream-E1 through our project websites: https://github.com/HiDream-ai/HiDream-I1 and https://github.com/HiDream-ai/HiDream-E1. All features can be directly experienced via https://vivago.ai/studio.

cs.CV↗

arXiv2025

Video-XL-Pro: Reconstructive Token Compression for Extremely Long Video Understanding

Despite advanced token compression techniques, existing multimodal large language models (MLLMs) still struggle with hour-long video understanding. In this work, we propose Video-XL-Pro, an efficient method for extremely long video understanding, built upon Reconstructive Compression of Tokens (ReCoT), a learnable module that leverages self-supervised learning to generate comprehensive and compact video tokens. ReCoT introduces two key components: (i) Dynamic Token Synthesizer (DTS): DTS generates pseudo-video tokens from static image tokens by learning intra-token relationships, which are then used in masked video modeling. (ii) Semantic-Guided Masking (SGM): SGM adaptively masks redundant visual tokens to facilitate more effective reconstructive learning. To improve training efficiency in MLLMs fine-tuning, we introduce a video-specific dataset pruning strategy and design a simple yet Query-aware Selector that enables the model to precisely locate query-relevant video tokens. With only 3B parameters, Video-XL-Pro outperforms most 7B models trained on larger datasets across multiple long video understanding benchmarks. Moreover, it can process over 8K frames on a single A100 GPU while maintaining high-quality performance.

cs.CV↗

arXiv2025

CR-LSO: Convex Neural Architecture Optimization in the Latent Space of Graph Variational Autoencoder with Input Convex Neural Networks

In neural architecture search (NAS) methods based on latent space optimization (LSO), a deep generative model is trained to embed discrete neural architectures into a continuous latent space. In this case, different optimization algorithms that operate in the continuous space can be implemented to search neural architectures. However, the optimization of latent variables is challenging for gradient-based LSO since the mapping from the latent space to the architecture performance is generally non-convex. To tackle this problem, this paper develops a convexity regularized latent space optimization (CR-LSO) method, which aims to regularize the learning process of latent space in order to obtain a convex architecture performance mapping. Specifically, CR-LSO trains a graph variational autoencoder (G-VAE) to learn the continuous representations of discrete architectures. Simultaneously, the learning process of latent space is regularized by the guaranteed convexity of input convex neural networks (ICNNs). In this way, the G-VAE is forced to learn a convex mapping from the architecture representation to the architecture performance. Hereafter, the CR-LSO approximates the performance mapping using the ICNN and leverages the estimated gradient to optimize neural architecture representations. Experimental results on three popular NAS benchmarks show that CR-LSO achieves competitive evaluation results in terms of both computational complexity and architecture performance.

cs.LG↗

arXiv2025

FX-DARTS: Designing Topology-unconstrained Architectures with Differentiable Architecture Search and Entropy-based Super-network Shrinking

Strong priors are imposed on the search space of Differentiable Architecture Search (DARTS), such that cells of the same type share the same topological structure and each intermediate node retains two operators from distinct nodes. While these priors reduce optimization difficulties and improve the applicability of searched architectures, they hinder the subsequent development of automated machine learning (Auto-ML) and prevent the optimization algorithm from exploring more powerful neural networks through improved architectural flexibility. This paper aims to reduce these prior constraints by eliminating restrictions on cell topology and modifying the discretization mechanism for super-networks. Specifically, the Flexible DARTS (FX-DARTS) method, which leverages an Entropy-based Super-Network Shrinking (ESS) framework, is presented to address the challenges arising from the elimination of prior constraints. Notably, FX-DARTS enables the derivation of neural architectures without strict prior rules while maintaining the stability in the enlarged search space. Experimental results on image classification benchmarks demonstrate that FX-DARTS is capable of exploring a set of neural architectures with competitive trade-offs between performance and computational complexity within a single search procedure.

cs.LG↗

arXiv2025

DNAD: Differentiable Neural Architecture Distillation

To meet the demand for designing efficient neural networks with appropriate trade-offs between model performance (e.g., classification accuracy) and computational complexity, the differentiable neural architecture distillation (DNAD) algorithm is developed based on two cores, namely search by deleting and search by imitating. Primarily, to derive neural architectures in a space where cells of the same type no longer share the same topology, the super-network progressive shrinking (SNPS) algorithm is developed based on the framework of differentiable architecture search (DARTS), i.e., search by deleting. Unlike conventional DARTS-based approaches which yield neural architectures with simple structures and derive only one architecture during the search procedure, SNPS is able to derive a Pareto-optimal set of architectures with flexible structures by forcing the dynamic super-network shrink from a dense structure to a sparse one progressively. Furthermore, since knowledge distillation (KD) has shown great effectiveness to train a compact network with the assistance of an over-parameterized model, we integrate SNPS with KD to formulate the DNAD algorithm, i.e., search by imitating. By minimizing behavioral differences between the super-network and teacher network, the over-fitting of one-level DARTS is avoided and well-performed neural architectures are derived. Experiments on CIFAR-10 and ImageNet classification tasks demonstrate that both SNPS and DNAD are able to derive a set of architectures which achieve similar or lower error rates with fewer parameters and FLOPs. Particularly, DNAD achieves the top-1 error rate of 23.7% on ImageNet classification with a model of 6.0M parameters and 598M FLOPs, which outperforms most DARTS-based methods.

cs.LG↗

arXiv2025

Improving Learning to Optimize Using Parameter Symmetries

We analyze a learning-to-optimize (L2O) algorithm that exploits parameter space symmetry to enhance optimization efficiency. Prior work has shown that jointly learning symmetry transformations and local updates improves meta-optimizer performance. Supporting this, our theoretical analysis demonstrates that even without identifying the optimal group element, the method locally resembles Newton's method. We further provide an example where the algorithm provably learns the correct symmetry transformation during training. To empirically evaluate L2O with teleportation, we introduce a benchmark, analyze its success and failure cases, and show that enhancements like momentum further improve performance. Our results highlight the potential of leveraging neural network parameter space symmetry to advance meta-optimization.

cs.LG↗

arXiv2025

EDA-Q: Electronic Design Automation for Superconducting Quantum Chip

Electronic Design Automation (EDA) plays a crucial role in classical chip design and significantly influences the development of quantum chip design. However, traditional EDA tools cannot be directly applied to quantum chip design due to vast differences compared to the classical realm. Several EDA products tailored for quantum chip design currently exist, yet they only cover partial stages of the quantum chip design process instead of offering a fully comprehensive solution. Additionally, they often encounter issues such as limited automation, steep learning curves, challenges in integrating with actual fabrication processes, and difficulties in expanding functionality. To address these issues, we developed a full-stack EDA tool specifically for quantum chip design, called EDA-Q. The design workflow incorporates functionalities present in existing quantum EDA tools while supplementing critical design stages such as device mapping and fabrication process mapping, which users expect. EDA-Q utilizes a unique architecture to achieve exceptional scalability and flexibility. The integrated design mode guarantees algorithm compatibility with different chip components, while employing a specialized interactive processing mode to offer users a straightforward and adaptable command interface. Application examples demonstrate that EDA-Q significantly reduces chip design cycles, enhances automation levels, and decreases the time required for manual intervention. Multiple rounds of testing on the designed chip have validated the effectiveness of EDA-Q in practical applications.

cs.ET↗

arXiv2025

Fractionalized Kohn-Sham Scheme for Strongly Correlated Electrons

We propose to expand the territory of density functional theory to strongly correlated electrons by reformulating the Kohn-Sham scheme in the representation of fractionalized particles. We call it the ``KS* scheme.'' Using inhomogeneous $t$-$J$ chains as a test bed, we show that the KS* scheme with simple local density approximation is able to achieve accurate ground-state energy and density distribution comparable to the density matrix renormalization group method, while the computational complexity is much lower.

physics.comp-ph↗

arXiv2025

Domain Generalization for Face Anti-spoofing via Content-aware Composite Prompt Engineering

The challenge of Domain Generalization (DG) in Face Anti-Spoofing (FAS) is the significant interference of domain-specific signals on subtle spoofing clues. Recently, some CLIP-based algorithms have been developed to alleviate this interference by adjusting the weights of visual classifiers. However, our analysis of this class-wise prompt engineering suffers from two shortcomings for DG FAS: (1) The categories of facial categories, such as real or spoof, have no semantics for the CLIP model, making it difficult to learn accurate category descriptions. (2) A single form of prompt cannot portray the various types of spoofing. In this work, instead of class-wise prompts, we propose a novel Content-aware Composite Prompt Engineering (CCPE) that generates instance-wise composite prompts, including both fixed template and learnable prompts. Specifically, our CCPE constructs content-aware prompts from two branches: (1) Inherent content prompt explicitly benefits from abundant transferred knowledge from the instruction-based Large Language Model (LLM). (2) Learnable content prompts implicitly extract the most informative visual content via Q-Former. Moreover, we design a Cross-Modal Guidance Module (CGM) that dynamically adjusts unimodal features for fusion to achieve better generalized FAS. Finally, our CCPE has been validated for its effectiveness in multiple cross-domain experiments and achieves state-of-the-art (SOTA) results.

cs.CV↗

arXiv2025

High-resolution and ultra-low power nonlinear image processing with passive high-quality factor metasurfaces

Image processing is both one of the most exciting domains for applying artificial intelligence and the most computationally expensive. Nanostructured metasurfaces have opened the door to the ultimate energy saving by directly processing ambient image data via ultra-thin layers before detection. However, a key ingredient of universal computation - nonlinear thresholding functions - have yet to be demonstrated for low intensities without an external power source. Here, we present a passive, all-optical method for nonlinear image processing using Silicon nanoantenna arrays. We experimentally demonstrate an intensity thresholding filter capable of processing one-dimensional images with only Watt-level power. By leveraging the opto-thermal Kerr nonlinearity through high-Q guided mode resonance, we achieve an experimental threshold as low as 0.1 mW/μm^2 with a spatial resolution of 1.85 μm. Additional simulations indicate that the threshold can be further reduced while maintaining high spatial selectivity. Analog, pixel-wise, smoothed leaky ReLU activation filters promise to revolutionize image sensing.

physics.optics↗