arXiv subjects

Xiaopeng Peng

Publications and source records attributed to Xiaopeng Peng.

12 recordsLinked to original sources

arXiv2026

Improving Generalization Robustness of Multimodal RLVR

Reinforcement Learning with Verifiable Rewards (RLVR) makes Multimodal Large Language Models more accurate, but the gains are brittle: simply paraphrasing a question or changing the prompt template can degrade them, which challenges reliable deployment in high-stakes scenarios like medical VQA. We trace this to two issues of the standard RL objective. First, the binary verifier conflates format with content, so the reward signal cannot tell a wrong answer apart from a misformatted one. Second, the training distribution covers only a thin slice of the real-world prompts that the model might meet at deployment, so policies that perform well on the training distribution can behave differently under unseen prompts during test. Both failures call for a robust post-training method that helps the policy cover a broader distribution of semantically equivalent prompts, and we identify two measures that help achieve this objective: separating format from semantics in the reward, and applying policy invariance across perturbed prompts with equivalent semantics. We therefore propose Prompt-Invariant RLVR (PIRL), consisting of a dynamic trinary reward and a consistency regularizer based on an embedding-space adversary. Under stress testing, PIRL's average accuracy on benchmarks drops by only $\le 1\%$, where GRPO drops ~3%. On dynamic evaluation, PIRL also achieves the smallest performance drop.

cs.AI↗

arXiv2026

Unified Hallucination Fuzzing for Multimodal Large Language Models

Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications. Existing evaluations, predominantly based on static benchmarks, suffer from narrow taxonomical coverage and rapid performance saturation, failing to reflect model robustness in evolving real-world scenarios. To bridge this gap, we present a systematic evaluation framework integrating a comprehensive benchmark with self-evolving stress testing. First, we introduce UniHall, a fine-grained dataset grounded in a unified taxonomy spanning Object, Instruction, and Knowledge dimensions. Second, to address benchmark saturation, we propose Self-Adaptive Multimodal Fuzzing (SAMF), a self-adaptive framework that employs evolutionary mutation strategies to explore the boundaries of model hallucinations. Crucially, to ensure reliable assessment of dynamic inputs, SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles. Our extensive experiments reveal that state-of-the-art MLLMs exhibit significant performance degradation under fuzzing compared to conventional settings, exposing a dissociation between reasoning capabilities and factual grounding. Furthermore, we identify a helpfulness-hallucination trade-off, where reinforcement learning alignment inadvertently exacerbates sycophancy in instruction-following tasks. The framework, code and benchmark are available at https://github.com/LanceZPF/EvalHall.

cs.CL↗

arXiv2026

Neural-Driven Image Editing

Traditional image editing typically relies on manual prompting, making it labor-intensive and inaccessible to individuals with limited motor control or language abilities. Leveraging recent advances in brain-computer interfaces (BCIs) and generative models, we propose LoongX, a hands-free image editing approach driven by multimodal neurophysiological signals. LoongX utilizes state-of-the-art diffusion models trained on a comprehensive dataset of 23,928 image editing pairs, each paired with synchronized electroencephalography (EEG), functional near-infrared spectroscopy (fNIRS), photoplethysmography (PPG), and head motion signals that capture user intent. To effectively address the heterogeneity of these signals, LoongX integrates two key modules. The cross-scale state space (CS3) module encodes informative modality-specific features. The dynamic gated fusion (DGF) module further aggregates these features into a unified latent space, which is then aligned with edit semantics via fine-tuning on a diffusion transformer (DiT). Additionally, we pre-train the encoders using contrastive learning to align cognitive states with semantic intentions from embedded natural language. Extensive experiments demonstrate that LoongX achieves performance comparable to text-driven methods (CLIP-I: 0.6605 vs. 0.6558; DINO: 0.4812 vs. 0.4636) and outperforms them when neural signals are combined with speech (CLIP-T: 0.2588 vs. 0.2549). These results highlight the promise of neural-driven generative models in enabling accessible, intuitive image editing and open new directions for cognitive-driven creative technologies. The code and dataset are released on the project website: https://loongx1.github.io.

cs.CV↗

arXiv2025

Toward Human-Aligned Luminance Measurement for Large-Format LED Displays

Direct-view LED displays are widely adopted in large-format applications due to their high luminance and reliability. However, visual comfort and accurate performance evaluation remain challenging due to the complex interaction between pixel luminance, human visual perception, and measurement artifacts. In this work, we introduce a novel 2D imaging luminance meter that replicates key optical parameters of the human eye, including entrance pupil size and angular resolution, to assess perceived pixel luminance. We report comprehensive measurements across various visual field angles and distances and establish a refined luminance metric that aligns with foveal vision standards (1/120 degree). Furthermore, a new method to quantify and mitigate stray light effects significantly improves measurement precision by reducing luminance overestimation from 7\% to 2\%. Our findings provide a foundation for optimizing LED display design for perceptual comfort and advancing standardization in pixel luminance evaluation.

physics.optics↗

arXiv2025

Learning to See Through Flare

Machine vision systems are susceptible to laser flare, where unwanted intense laser illumination blinds and distorts its perception of the environment through oversaturation or permanent damage to sensor pixels. We introduce NeuSee, the first computational imaging framework for high-fidelity sensor protection across the full visible spectrum. It jointly learns a neural representation of a diffractive optical element (DOE) and a frequency-space Mamba-GAN network for image restoration. NeuSee system is adversarially trained end-to-end on 100K unique images to suppress the peak laser irradiance as high as $10^6$ times the sensor saturation threshold $I_{\textrm{sat}}$, the point at which camera sensors may experience damage without the DOE. Our system leverages heterogeneous data and model parallelism for distributed computing, integrating hyperspectral information and multiple neural networks for realistic simulation and image restoration. NeuSee takes into account open-world scenes with dynamically varying laser wavelengths, intensities, and positions, as well as lens flare effects, unknown ambient lighting conditions, and sensor noises. It outperforms other learned DOEs, achieving full-spectrum imaging and laser suppression for the first time, with a 10.1\% improvement in restored image quality.

eess.IV↗

arXiv2025

MDK12-Bench: A Comprehensive Evaluation of Multimodal Large Language Models on Multidisciplinary Exams

Multimodal large language models (MLLMs), which integrate language and visual cues for problem-solving, are crucial for advancing artificial general intelligence (AGI). However, current benchmarks for measuring the intelligence of MLLMs suffer from limited scale, narrow coverage, and unstructured knowledge, offering only static and undifferentiated evaluations. To bridge this gap, we introduce MDK12-Bench, a large-scale multidisciplinary benchmark built from real-world K-12 exams spanning six disciplines with 141K instances and 6,225 knowledge points organized in a six-layer taxonomy. Covering five question formats with difficulty and year annotations, it enables comprehensive evaluation to capture the extent to which MLLMs perform over four dimensions: 1) difficulty levels, 2) temporal (cross-year) shifts, 3) contextual shifts, and 4) knowledge-driven reasoning. We propose a novel dynamic evaluation framework that introduces unfamiliar visual, textual, and question form shifts to challenge model generalization while improving benchmark objectivity and longevity by mitigating data contamination. We further evaluate knowledge-point reference-augmented generation (KP-RAG) to examine the role of knowledge in problem-solving. Key findings reveal limitations in current MLLMs in multiple aspects and provide guidance for enhancing model robustness, interpretability, and AI-assisted education.

cs.AI↗

arXiv2025

MDK12-Bench: A Multi-Discipline Benchmark for Evaluating Reasoning in Multimodal Large Language Models

Multimodal reasoning, which integrates language and visual cues into problem solving and decision making, is a fundamental aspect of human intelligence and a crucial step toward artificial general intelligence. However, the evaluation of multimodal reasoning capabilities in Multimodal Large Language Models (MLLMs) remains inadequate. Most existing reasoning benchmarks are constrained by limited data size, narrow domain coverage, and unstructured knowledge distribution. To close these gaps, we introduce MDK12-Bench, a multi-disciplinary benchmark assessing the reasoning capabilities of MLLMs via real-world K-12 examinations. Spanning six disciplines (math, physics, chemistry, biology, geography, and information science), our benchmark comprises 140K reasoning instances across diverse difficulty levels from primary school to 12th grade. It features 6,827 instance-level knowledge point annotations based on a well-organized knowledge structure, detailed answer explanations, difficulty labels and cross-year partitions, providing a robust platform for comprehensive evaluation. Additionally, we present a novel dynamic evaluation framework to mitigate data contamination issues by bootstrapping question forms, question types, and image styles during evaluation. Extensive experiment on MDK12-Bench reveals the significant limitation of current MLLMs in multimodal reasoning. The findings on our benchmark provide insights into the development of the next-generation models. Our data and codes are available at https://github.com/LanceZPF/MDK12.

cs.CV↗

arXiv2025

OpenING: A Comprehensive Benchmark for Judging Open-ended Interleaved Image-Text Generation

Multimodal Large Language Models (MLLMs) have made significant strides in visual understanding and generation tasks. However, generating interleaved image-text content remains a challenge, which requires integrated multimodal understanding and generation abilities. While the progress in unified models offers new solutions, existing benchmarks are insufficient for evaluating these methods due to limitations in data size and diversity. To bridge this gap, we introduce OpenING, a comprehensive benchmark comprising 5,400 high-quality human-annotated instances across 56 real-world tasks. OpenING covers diverse daily scenarios such as travel guide, design, and brainstorming, offering a robust platform for challenging interleaved generation methods. In addition, we present IntJudge, a judge model for evaluating open-ended multimodal generation methods. Trained with a novel data pipeline, our IntJudge achieves an agreement rate of 82.42% with human judgments, outperforming GPT-based evaluators by 11.34%. Extensive experiments on OpenING reveal that current interleaved generation methods still have substantial room for improvement. Key findings on interleaved image-text generation are further presented to guide the development of next-generation models.

cs.CV↗

arXiv2024

Multi-Source EEG Emotion Recognition via Dynamic Contrastive Domain Adaptation

Electroencephalography (EEG) provides reliable indications of human cognition and mental states. Accurate emotion recognition from EEG remains challenging due to signal variations among individuals and across measurement sessions. We introduce a multi-source dynamic contrastive domain adaptation method (MS-DCDA) based on differential entropy (DE) features, in which coarse-grained inter-domain and fine-grained intra-class adaptations are modeled through a multi-branch contrastive neural network and contrastive sub-domain discrepancy learning. Leveraging domain knowledge from each individual source and a complementary source ensemble, our model uses dynamically weighted learning to achieve an optimal tradeoff between domain transferability and discriminability. The proposed MS-DCDA model was evaluated using the SEED and SEED-IV datasets, achieving respectively the highest mean accuracies of $90.84\%$ and $78.49\%$ in cross-subject experiments as well as $95.82\%$ and $82.25\%$ in cross-session experiments. Our model outperforms several alternative domain adaptation methods in recognition accuracy, inter-class margin, and intra-class compactness. Our study also suggests greater emotional sensitivity in the frontal and parietal brain lobes, providing insights for mental health interventions, personalized medicine, and preventive strategies.

eess.SP↗

arXiv2024

ContextDet: Temporal Action Detection with Adaptive Context Aggregation

Temporal action detection (TAD), which locates and recognizes action segments, remains a challenging task in video understanding due to variable segment lengths and ambiguous boundaries. Existing methods treat neighboring contexts of an action segment indiscriminately, leading to imprecise boundary predictions. We introduce a single-stage ContextDet framework, which makes use of large-kernel convolutions in TAD for the first time. Our model features a pyramid adaptive context aggragation (ACA) architecture, capturing long context and improving action discriminability. Each ACA level consists of two novel modules. The context attention module (CAM) identifies salient contextual information, encourages context diversity, and preserves context integrity through a context gating block (CGB). The long context module (LCM) makes use of a mixture of large- and small-kernel convolutions to adaptively gather long-range context and fine-grained local features. Additionally, by varying the length of these large kernels across the ACA pyramid, our model provides lightweight yet effective context aggregation and action discrimination. We conducted extensive experiments and compared our model with a number of advanced TAD methods on six challenging TAD benchmarks: MultiThumos, Charades, FineAction, EPIC-Kitchens 100, Thumos14, and HACS, demonstrating superior accuracy at reduced inference speed.

cs.CV↗

arXiv2024

Learning to See Through Dazzle

Machine vision is susceptible to laser dazzle, where intense laser light can blind and distort its perception of the environment through oversaturation or permanent damage to sensor pixels. Here we employ a wavefront-coded phase mask to diffuse the energy of laser light and introduce a sandwich generative adversarial network (SGAN) to restore images from complex image degradations, such as varying laser-induced image saturation, mask-induced image blurring, unknown lighting conditions, and various noise corruptions. The SGAN architecture combines discriminative and generative methods by wrapping two GANs around a learnable image deconvolution module. In addition, we make use of Fourier feature representations to reduce the spectral bias of neural networks and improve its learning of high-frequency image details. End-to-end training includes the realistic physics-based synthesis of a large set of training data from publicly available images. We trained the SGAN to suppress the peak laser irradiance as high as $10^6$ times the sensor saturation threshold - the point at which camera sensors may experience damage without the mask. The trained model was evaluated on both a synthetic data set and data collected from the laboratory. The proposed image restoration model quantitatively and qualitatively outperforms state-of-the-art methods for a wide range of scene contents, laser powers, incident laser angles, ambient illumination strengths, and noise characteristics.

cs.CV↗

arXiv2016

Randomized Aperture Imaging

Speckled images of a binary broad band light source (600-670 nm), generated by randomized reflections or transmissions, were used to reconstruct a binary image by use of multi-frame blind deconvolution algorithms. Craft store glitter was used as reflective elements. Another experiment used perforated foil. Also reported here are numerical models that afforded controlled tip-tilt and piston aberrations. These results suggest the potential importance of a poorly figured, randomly varying segmented imaging system.

astro-ph.IM↗