arXiv subjects

Radu Marculescu

Publications and source records attributed to Radu Marculescu.

At least 19 recordsLinked to original sources

arXiv2026

Pareto LoRA: Mitigating Modality Imbalance in Unified Multimodal Models via Pareto-Optimal Gradient Integration

Unified multimodal models (UMMs) have recently emerged as a promising paradigm for integrating multimodal understanding and generation within a single autoregressive transformer. However, during multimodal instruction tuning, these models often exhibit pronounced modality imbalance: language gradients dominate optimization, thus leading to lower image generation quality, especially under parameter-efficient fine-tuning such as LoRA. In this work, we systematically analyze modality imbalance in LoRA-based fine-tuning of UMMs for interleaved text-image generation. We show that vision modality performance degrades substantially more than text modality performance when compared to unimodal counterparts, and that modality-specific gradients can differ by orders of magnitude across various tasks and layers. Motivated by this observation, we reformulate the multimodal instruction tuning as a bi-objective optimization problem and propose Pareto LoRA, a Pareto-optimal gradient integration strategy that balances the text and image objectives by modulating the gradient direction and strength. Experiments on the CoMM benchmark with Emu2 demonstrate that Pareto LoRA consistently improves multimodal generation balance, achieving up to 44.9% gains in perceptual image quality over vanilla LoRA while maintaining comparable text performance.

cs.CV↗

arXiv2026

Fuel Gauge: Estimating Chain-of-Thought Length Ahead of Time in Large Multimodal Models

Reasoning Large Multi-modality Models (LMMs) have become the de facto choice for many applications. However, these models rely on a Chain-of-Thought (CoT) process that is lengthy and unpredictable at runtime, often resulting in inefficient use of computational resources (due to memory fragmentation) and sub-optimal accuracy (due to under- and over-thinking). We observe empirically that the CoT process follows a very simple form, whose behavior is independent of the specific generated samples. This suggests that the CoT length can be estimated ahead of time based on a hidden parameter representing the amount of "fuel" available to support the reasoning process. Based on this insight, we propose Fuel Gauge, the first method which extracts this hidden signal and predicts CoT length ahead of time. We demonstrate the utility on the Fuel Gauge on two downstream tasks: predictive KV cache allocation, which addresses memory fragmentation in LMM serving systems, and CoT length modulation, which mitigates under-thinking and over-thinking. Extensive experiments on LMMs across text-only, image-text, and video-text question answering benchmarks demonstrate the effectiveness, generalizability, and practical value of our Fuel Gauge. For example, on the GPQA-Diamond benchmark, our Fuel Gauge achieves less than half the CoT length prediction error compared to the baseline; this translates into a 13.37x reduction in the memory allocation frequency.

cs.CV↗

arXiv2026

Video Compression Meets Video Generation: Latent Inter-Frame Pruning with Attention Recovery

Current video generation models suffer from high computational latency, making real-time applications prohibitively costly. In this paper, we address this limitation by exploiting the temporal redundancy inherent in video latent patches. To this end, we propose the Latent Inter-frame Pruning with Attention Recovery (LIPAR) framework, which detects and skips recomputing duplicated latent patches. Additionally, we introduce a novel Attention Recovery mechanism that approximates the attention values of pruned tokens, thereby removing visual artifacts arising from naively applying the pruning method. Empirically, our method increases video editing throughput by $1.53\times$, achieving an average of 19.3 FPS on an NVIDIA RTX 4090 with the 1.3B Self-Forcing model (4-step denoising, FP16). The proposed method does not compromise generation quality and can be seamlessly integrated with the model without additional training. Our approach effectively bridges the gap between traditional compression algorithms and modern generative pipelines.

cs.CV↗

arXiv2025

PipeFlow: Pipelined Processing and Motion-Aware Frame Selection for Long-Form Video Editing

Long-form video editing poses unique challenges due to the exponential increase in the computational cost from joint editing and Denoising Diffusion Implicit Models (DDIM) inversion across extended sequences. To address these limitations, we propose PipeFlow, a scalable, pipelined video editing method that introduces three key innovations: First, based on a motion analysis using Structural Similarity Index Measure (SSIM) and Optical Flow, we identify and propose to skip editing of frames with low motion. Second, we propose a pipelined task scheduling algorithm that splits a video into multiple segments and performs DDIM inversion and joint editing in parallel based on available GPU memory. Lastly, we leverage a neural network-based interpolation technique to smooth out the border frames between segments and interpolate the previously skipped frames. Our method uniquely scales to longer videos by dividing them into smaller segments, allowing PipeFlow's editing time to increase linearly with video length. In principle, this enables editing of infinitely long videos without the growing per-frame computational overhead encountered by other methods. PipeFlow achieves up to a 9.6X speedup compared to TokenFlow and a 31.7X speedup over Diffusion Motion Transfer (DMT).

cs.CV↗

arXiv2025

Single-Round Scalable Analytic Federated Learning

Federated Learning (FL) is plagued by two key challenges: high communication overhead and performance collapse on heterogeneous (non-IID) data. Analytic FL (AFL) provides a single-round, data distribution invariant solution, but is limited to linear models. Subsequent non-linear approaches, like DeepAFL, regain accuracy but sacrifice the single-round benefit. In this work, we break this trade-off. We propose SAFLe, a framework that achieves scalable non-linear expressivity by introducing a structured head of bucketed features and sparse, grouped embeddings. We prove this non-linear architecture is mathematically equivalent to a high-dimensional linear regression. This key equivalence allows SAFLe to be solved with AFL's single-shot, invariant aggregation law. Empirically, SAFLe establishes a new state-of-the-art for analytic FL, significantly outperforming both linear AFL and multi-round DeepAFL in accuracy across all benchmarks, demonstrating a highly efficient and scalable solution for federated vision.

cs.LG↗

arXiv2025

Mitigating Intra- and Inter-modal Forgetting in Continual Learning of Unified Multimodal Models

Unified Multimodal Generative Models (UMGMs) unify visual understanding and image generation within a single autoregressive framework. However, their ability to continually learn new tasks is severely hindered by catastrophic forgetting, both within a modality (intra-modal) and across modalities (inter-modal). While intra-modal forgetting has been studied in prior continual learning (CL) work, inter-modal forgetting remains largely unexplored. In this paper, we identify and empirically validate this phenomenon in UMGMs and provide a theoretical explanation rooted in gradient conflict between modalities. To address both intra- and inter-modal forgetting, we propose Modality-Decoupled Experts (MoDE), a lightweight and scalable architecture that isolates modality-specific updates to mitigate the gradient conflict and leverages knowledge distillation to prevent catastrophic forgetting and preserve pre-trained capabilities. Unlike previous CL methods that remain modality-coupled and suffer from modality gradient conflict, MoDE explicitly decouples modalities to prevent interference. Experiments across diverse benchmarks demonstrate that MoDE significantly mitigates both inter- and intra-modal forgetting, outperforming prior CL baselines in unified multimodal generation settings. Codes will be publicly available: https://github.com/Christina200/MoDE-official.git

cs.LG↗

arXiv2025

ObjectAlign: Neuro-Symbolic Object Consistency Verification and Correction

Video editing and synthesis often introduce object inconsistencies, such as frame flicker and identity drift that degrade perceptual quality. To address these issues, we introduce ObjectAlign, a novel framework that seamlessly blends perceptual metrics with symbolic reasoning to detect, verify, and correct object-level and temporal inconsistencies in edited video sequences. The novel contributions of ObjectAlign are as follows: First, we propose learnable thresholds for metrics characterizing object consistency (i.e. CLIP-based semantic similarity, LPIPS perceptual distance, histogram correlation, and SAM-derived object-mask IoU). Second, we introduce a neuro-symbolic verifier that combines two components: (a) a formal, SMT-based check that operates on masked object embeddings to provably guarantee that object identity does not drift, and (b) a temporal fidelity check that uses a probabilistic model checker to verify the video's formal representation against a temporal logic specification. A frame transition is subsequently deemed "consistent" based on a single logical assertion that requires satisfying both the learned metric thresholds and this unified neuro-symbolic constraint, ensuring both low-level stability and high-level temporal correctness. Finally, for each contiguous block of flagged frames, we propose a neural network based interpolation for adaptive frame repair, dynamically choosing the interpolation depth based on the number of frames to be corrected. This enables reconstruction of the corrupted frames from the last valid and next valid keyframes. Our results show up to 1.4 point improvement in CLIP Score and up to 6.1 point improvement in warp error compared to SOTA baselines on the DAVIS and Pexels video datasets.

cs.CV↗

arXiv2025

AdaptViG: Adaptive Vision GNN with Exponential Decay Gating

Vision Graph Neural Networks (ViGs) offer a new direction for advancements in vision architectures. While powerful, ViGs often face substantial computational challenges stemming from their graph construction phase, which can hinder their efficiency. To address this issue we propose AdaptViG, an efficient and powerful hybrid Vision GNN that introduces a novel graph construction mechanism called Adaptive Graph Convolution. This mechanism builds upon a highly efficient static axial scaffold and a dynamic, content-aware gating strategy called Exponential Decay Gating. This gating mechanism selectively weighs long-range connections based on feature similarity. Furthermore, AdaptViG employs a hybrid strategy, utilizing our efficient gating mechanism in the early stages and a full Global Attention block in the final stage for maximum feature aggregation. Our method achieves a new state-of-the-art trade-off between accuracy and efficiency among Vision GNNs. For instance, our AdaptViG-M achieves 82.6% top-1 accuracy, outperforming ViG-B by 0.3% while using 80% fewer parameters and 84% fewer GMACs. On downstream tasks, AdaptViG-M obtains 45.8 mIoU, 44.8 APbox, and 41.1 APmask, surpassing the much larger EfficientFormer-L7 by 0.7 mIoU, 2.2 APbox, and 2.1 APmask, respectively, with 78% fewer parameters.

cs.CV↗

arXiv2025

LoMix: Learnable Weighted Multi-Scale Logits Mixing for Medical Image Segmentation

U-shaped networks output logits at multiple spatial scales, each capturing a different blend of coarse context and fine detail. Yet, training still treats these logits in isolation - either supervising only the final, highest-resolution logits or applying deep supervision with identical loss weights at every scale - without exploring mixed-scale combinations. Consequently, the decoder output misses the complementary cues that arise only when coarse and fine predictions are fused. To address this issue, we introduce LoMix (Logits Mixing), a NAS-inspired, differentiable plug-and-play module that generates new mixed-scale outputs and learns how exactly each of them should guide the training process. More precisely, LoMix mixes the multi-scale decoder logits with four lightweight fusion operators: addition, multiplication, concatenation, and attention-based weighted fusion, yielding a rich set of synthetic mutant maps. Every original or mutant map is given a softplus loss weight that is co-optimized with network parameters, mimicking a one-step architecture search that automatically discovers the most useful scales, mixtures, and operators. Plugging LoMix into recent U-shaped architectures (i.e., PVT-V2-B2 backbone with EMCAD decoder) on Synapse 8-organ dataset improves DICE by +4.2% over single-output supervision, +2.2% over deep supervision, and +1.5% over equally weighted additive fusion, all with zero inference overhead. When training data are scarce (e.g., one or two labeled scans), the advantage grows to +9.23%, underscoring LoMix's data efficiency. Across four benchmarks and diverse U-shaped networks, LoMiX improves DICE by up to +13.5% over single-output supervision, confirming that learnable weighted mixed-scale fusion generalizes broadly while remaining data efficient, fully interpretable, and overhead-free at inference. Our code is available at https://github.com/SLDGroup/LoMix.

cs.CV↗

arXiv2025

Multi-Scale High-Resolution Logarithmic Grapher Module for Efficient Vision GNNs

Vision graph neural networks (ViG) have demonstrated promise in vision tasks as a competitive alternative to conventional convolutional neural nets (CNN) and transformers (ViTs); however, common graph construction methods, such as k-nearest neighbor (KNN), can be expensive on larger images. While methods such as Sparse Vision Graph Attention (SVGA) have shown promise, SVGA's fixed step scale can lead to over-squashing and missing multiple connections to gain the same information that could be gained from a long-range link. Through this observation, we propose a new graph construction method, Logarithmic Scalable Graph Construction (LSGC) to enhance performance by limiting the number of long-range links. To this end, we propose LogViG, a novel hybrid CNN-GNN model that utilizes LSGC. Furthermore, inspired by the successes of multi-scale and high-resolution architectures, we introduce and apply a high-resolution branch and fuse features between our high-resolution and low-resolution branches for a multi-scale high-resolution Vision GNN network. Extensive experiments show that LogViG beats existing ViG, CNN, and ViT architectures in terms of accuracy, GMACs, and parameters on image classification and semantic segmentation tasks. Our smallest model, Ti-LogViG, achieves an average top-1 accuracy on ImageNet-1K of 79.9% with a standard deviation of 0.2%, 1.7% higher average accuracy than Vision GNN with a 24.3% reduction in parameters and 35.3% reduction in GMACs. Our work shows that leveraging long-range links in graph construction for ViGs through our proposed LSGC can exceed the performance of current state-of-the-art ViGs. Code is available at https://github.com/mmunir127/LogViG-Official.

cs.CV↗

arXiv2025

AttentionViG: Cross-Attention-Based Dynamic Neighbor Aggregation in Vision GNNs

Vision Graph Neural Networks (ViGs) have demonstrated promising performance in image recognition tasks against Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). An essential part of the ViG framework is the node-neighbor feature aggregation method. Although various graph convolution methods, such as Max-Relative, EdgeConv, GIN, and GraphSAGE, have been explored, a versatile aggregation method that effectively captures complex node-neighbor relationships without requiring architecture-specific refinements is needed. To address this gap, we propose a cross-attention-based aggregation method in which the query projections come from the node, while the key projections come from its neighbors. Additionally, we introduce a novel architecture called AttentionViG that uses the proposed cross-attention aggregation scheme to conduct non-local message passing. We evaluated the image recognition performance of AttentionViG on the ImageNet-1K benchmark, where it achieved SOTA performance. Additionally, we assessed its transferability to downstream tasks, including object detection and instance segmentation on MS COCO 2017, as well as semantic segmentation on ADE20K. Our results demonstrate that the proposed method not only achieves strong performance, but also maintains efficiency, delivering competitive accuracy with comparable FLOPs to prior vision GNN architectures.

cs.CV↗

arXiv2025

MK-UNet: Multi-kernel Lightweight CNN for Medical Image Segmentation

In this paper, we introduce MK-UNet, a paradigm shift towards ultra-lightweight, multi-kernel U-shaped CNNs tailored for medical image segmentation. Central to MK-UNet is the multi-kernel depth-wise convolution block (MKDC) we design to adeptly process images through multiple kernels, while capturing complex multi-resolution spatial relationships. MK-UNet also emphasizes the images salient features through sophisticated attention mechanisms, including channel, spatial, and grouped gated attention. Our MK-UNet network, with a modest computational footprint of only 0.316M parameters and 0.314G FLOPs, represents not only a remarkably lightweight, but also significantly improved segmentation solution that provides higher accuracy over state-of-the-art (SOTA) methods across six binary medical imaging benchmarks. Specifically, MK-UNet outperforms TransUNet in DICE score with nearly 333$\times$ and 123$\times$ fewer parameters and FLOPs, respectively. Similarly, when compared against UNeXt, MK-UNet exhibits superior segmentation performance, improving the DICE score up to 6.7% margins while operating with 4.7$\times$ fewer #Params. Our MK-UNet also outperforms other recent lightweight networks, such as MedT, CMUNeXt, EGE-UNet, and Rolling-UNet, with much lower computational resources. This leap in performance, coupled with drastic computational gains, positions MK-UNet as an unparalleled solution for real-time, high-fidelity medical diagnostics in resource-limited settings, such as point-of-care devices. Our implementation is available at https://github.com/SLDGroup/MK-UNet.

cs.CV↗

arXiv2025

NeuS-QA: Grounding Long-Form Video Understanding in Temporal Logic and Neuro-Symbolic Reasoning

While vision-language models (VLMs) excel at tasks involving single images or short videos, they still struggle with Long Video Question Answering (LVQA) due to its demand for complex multi-step temporal reasoning. Vanilla approaches, which simply sample frames uniformly and feed them to a VLM along with the question, incur significant token overhead. This forces aggressive downsampling of long videos, causing models to miss fine-grained visual structure, subtle event transitions, and key temporal cues. Recent works attempt to overcome these limitations through heuristic approaches; however, they lack explicit mechanisms for encoding temporal relationships and fail to provide any formal guarantees that the sampled context actually encodes the compositional or causal logic required by the question. To address these foundational gaps, we introduce NeuS-QA, a training-free, plug-and-play neuro-symbolic pipeline for LVQA. NeuS-QA first translates a natural language question into a logic specification that models the temporal relationship between frame-level events. Next, we construct a video automaton to model the video's frame-by-frame event progression, and finally employ model checking to compare the automaton against the specification to identify all video segments that satisfy the question's logical requirements. Only these logic-verified segments are submitted to the VLM, thus improving interpretability, reducing hallucinations, and enabling compositional reasoning without modifying or fine-tuning the model. Experiments on the LongVideoBench and CinePile LVQA benchmarks show that NeuS-QA significantly improves performance by over 10%, particularly on questions involving event ordering, causality, and multi-step reasoning. We open-source our code at https://utaustin-swarmlab.github.io/NeuS-QA/.

cs.CV↗

arXiv2025

VCMamba: Bridging Convolutions with Multi-Directional Mamba for Efficient Visual Representation

Recent advances in Vision Transformers (ViTs) and State Space Models (SSMs) have challenged the dominance of Convolutional Neural Networks (CNNs) in computer vision. ViTs excel at capturing global context, and SSMs like Mamba offer linear complexity for long sequences, yet they do not capture fine-grained local features as effectively as CNNs. Conversely, CNNs possess strong inductive biases for local features but lack the global reasoning capabilities of transformers and Mamba. To bridge this gap, we introduce \textit{VCMamba}, a novel vision backbone that integrates the strengths of CNNs and multi-directional Mamba SSMs. VCMamba employs a convolutional stem and a hierarchical structure with convolutional blocks in its early stages to extract rich local features. These convolutional blocks are then processed by later stages incorporating multi-directional Mamba blocks designed to efficiently model long-range dependencies and global context. This hybrid design allows for superior feature representation while maintaining linear complexity with respect to image resolution. We demonstrate VCMamba's effectiveness through extensive experiments on ImageNet-1K classification and ADE20K semantic segmentation. Our VCMamba-B achieves 82.6% top-1 accuracy on ImageNet-1K, surpassing PlainMamba-L3 by 0.3% with 37% fewer parameters, and outperforming Vision GNN-B by 0.3% with 64% fewer parameters. Furthermore, VCMamba-B obtains 47.1 mIoU on ADE20K, exceeding EfficientFormer-L7 by 2.0 mIoU while utilizing 62% fewer parameters. Code is available at https://github.com/Wertyuui345/VCMamba.

cs.CV↗

arXiv2025

The Effect of Network Topology on the Equilibria of Influence-Opinion Games

Online social networks exert a powerful influence on public opinion. Adversaries weaponize these networks to manipulate discourse, underscoring the need for more resilient social networks. To this end, we investigate the impact of network connectivity on Stackelberg equilibria in a two-player game to shape public opinion. We model opinion evolution as a repeated competitive influence-propagation process. Players iteratively inject \textit{messages} that diffuse until reaching a steady state, modeling the dispersion of two competing messages. Opinions then update according to the discounted sum of exposure to the messages. This bi-level model captures viral-media correlation effects omitted by standard opinion-dynamics models. To solve the resulting high-dimensional game, we propose a scalable, iterative algorithm based on linear-quadratic regulators that approximates local feedback Stackelberg strategies for players with limited cognition. We analyze how the network topology shapes equilibrium outcomes through experiments on synthetic networks and real Facebook data. Our results identify structural characteristics that improve a network's resilience to adversarial influence, guiding the design of more resilient social networks.

cs.SI↗

arXiv2025

Simulating Rumor Spreading in Social Networks using LLM Agents

With the rise of social media, misinformation has become increasingly prevalent, fueled largely by the spread of rumors. This study explores the use of Large Language Model (LLM) agents within a novel framework to simulate and analyze the dynamics of rumor propagation across social networks. To this end, we design a variety of LLM-based agent types and construct four distinct network structures to conduct these simulations. Our framework assesses the effectiveness of different network constructions and agent behaviors in influencing the spread of rumors. Our results demonstrate that the framework can simulate rumor spreading across more than one hundred agents in various networks with thousands of edges. The evaluations indicate that network structure, personas, and spreading schemes can significantly influence rumor dissemination, ranging from no spread to affecting 83\% of agents in iterations, thereby offering a realistic simulation of rumor spread in social networks.

cs.SI↗

arXiv2025

Graph Learning for Bidirectional Disease Contact Tracing on Real Human Mobility Data

For rapidly spreading diseases where many cases show no symptoms, swift and effective contact tracing is essential. While exposure notification applications provide alerts on potential exposures, a fully automated system is needed to track the infectious transmission routes. To this end, our research leverages large-scale contact networks from real human mobility data to identify the path of transmission. More precisely, we introduce a new Infectious Path Centrality network metric that informs a graph learning edge classifier to identify important transmission events, achieving an F1-score of 94%. Additionally, we explore bidirectional contact tracing, which quarantines individuals both retroactively and proactively, and compare its effectiveness against traditional forward tracing, which only isolates individuals after testing positive. Our results indicate that when only 30% of symptomatic individuals are tested, bidirectional tracing can reduce infectious effective reproduction rate by 71%, thus significantly controlling the outbreak.

cs.SI↗

arXiv2024

RapidNet: Multi-Level Dilated Convolution Based Mobile Backbone

Vision transformers (ViTs) have dominated computer vision in recent years. However, ViTs are computationally expensive and not well suited for mobile devices; this led to the prevalence of convolutional neural network (CNN) and ViT-based hybrid models for mobile vision applications. Recently, Vision GNN (ViG) and CNN hybrid models have also been proposed for mobile vision tasks. However, all of these methods remain slower compared to pure CNN-based models. In this work, we propose Multi-Level Dilated Convolutions to devise a purely CNN-based mobile backbone. Using Multi-Level Dilated Convolutions allows for a larger theoretical receptive field than standard convolutions. Different levels of dilation also allow for interactions between the short-range and long-range features in an image. Experiments show that our proposed model outperforms state-of-the-art (SOTA) mobile CNN, ViT, ViG, and hybrid architectures in terms of accuracy and/or speed on image classification, object detection, instance segmentation, and semantic segmentation. Our fastest model, RapidNet-Ti, achieves 76.3\% top-1 accuracy on ImageNet-1K with 0.9 ms inference latency on an iPhone 13 mini NPU, which is faster and more accurate than MobileNetV2x1.4 (74.7\% top-1 with 1.0 ms latency). Our work shows that pure CNN architectures can beat SOTA hybrid and ViT models in terms of accuracy and speed when designed properly.

cs.CV↗