arXiv subjects

Pengfei Sun

Publications and source records attributed to Pengfei Sun.

At least 19 recordsLinked to original sources

arXiv2026

MetaVideoAgent: Automated Video-Agent Evolution for Long-Form Video Understanding

Long-form video understanding requires locating sparse, question-relevant evidence in long, multimodal videos. Real-world video distributions differ in modality-specific information density, content structure, and evidence patterns, causing fixed video-agent designs to incur redundant processing or fail when mismatched. Extending automated agent evolution from text to video is challenging because full long-video execution makes candidate validation expensive, failures propagate across coupled evidence-processing stages, and complex preprocessing, perception tools, and localization strategies make code-level updates difficult to implement reliably. We introduce MetaVideoAgent, a framework that automatically evolves a video agent for a target distribution. It profiles information density and evidence requirements from sparsely sampled frames and associated queries to guide initial design, then compresses localized failures into independently executable minimal validation tasks. It constructs evidence-grounded Gold Paths, audits Student trajectories, aggregates recurring failures across samples, and attributes them to responsible modules. A modular agent representation constrains each update to the primary responsible module and its necessary dependencies. We further introduce VA-EvoBench, covering eight video distributions with separate evolution and held-out splits. With four evolution iterations per distribution, MetaVideoAgent improves every initial agent and raises macro-average accuracy from 38.44% to 51.47%, at an average evolution cost of 3.54M tokens per distribution. The evolved agents outperform the strongest prior fixed-design video agent by 6.39 percentage points while using the fewest tokens and video frames per question among the compared video agents. We will release all code and data to support reproducible research.

cs.CV↗

arXiv2026

Yuvion LLM: An Adversarially-Aware Large Language Model for Content And AI Safety

As large language models are increasingly deployed in real-world systems, safety failures can still lead to harmful outputs and dangerous misuse. We argue that the essence of safety is adversarial: many failures arise not from natural inputs alone, but from strategic attempts to evade model policies and safeguards. However, existing general-purpose model development largely overlook this adversarial nature, and often remain insufficient for realistic safety scenarios involving planning, tool use, and multi-step reasoning, causing measured safety performance to overestimate real deployment robustness. To address this gap, we present Yuvion LLM, a large language model built for adversarially robust content safety and broader AI safety. Yuvion LLM treats adversarial robustness and agentic capability as first-class objectives. Its pipeline combines adversarially aware data construction, knowledge-enhanced continued pretraining, and policy-grounded multi-task safety post-training, including risk-aware supervised fine-tuning and reinforcement learning-based policy optimization, together with safety-aware agentic reinforcement learning for tool use and multi-step reasoning in complex safety scenarios. We further introduce the Yuvion LLM RiskEval (YLRE), a collection of 93 benchmarks across four evaluation categories, covering diverse open and internal evaluations with a focus on safety, adversarial robustness, and real-world capability requirements. Across these evaluations, Yuvion LLM demonstrates clear advantages on safety-focused benchmarks and particularly strong robustness under adversarial conditions, while maintaining solid overall capability. Notably, Yuvion-8B outperforms most state-of-the-art baselines, including substantially larger models such as GPT-5.4 and Qwen3-MAX, on several safety tasks.

cs.CL↗

arXiv2026

Yuvion VL: A Multimodal Foundation Model for Adversarial Content and AI Safety

General-purpose models often struggle to reliably identify and understand real-world multimodal risks, largely due to the inherent multimodal adversarial nature of content and AI safety. We present Yuvion VL, a family of multimodal large language models purpose-built for content and AI safety, with both instruction-tuned and reasoning-oriented variants. Yuvion VL addresses this gap by treating safety as an inherently adversarial and multimodal problem and designing the entire pipeline around adversarial robustness. For data construction, we develop an automated pipeline integrating adversarial-aware data synthesis with multi-stage quality control, producing large-scale, high-quality multimodal samples augmented with domain knowledge and reasoning annotations. For training, we adopt a three-stage pipeline that includes continued pretraining for risk-concept cross-modal alignment, instruct post-training for production-grade safety tasks, and reasoning post-training for enhanced interpretability and performance in complex tasks. We further introduce Confuse-then-Contrast Fine-Tuning, a contrastive framework that mines model-specific confusions and constructs multi-image contrastive groups to enforce explicit discrimination of fine-grained visual-semantic elements, enabling the model to distinguish between visually similar cases with different safety implications in adversarial safety tasks. To support rigorous evaluation, we further introduce Yuvion VL RiskEval (YVRE), a collection of benchmarks covering diverse open and internal evaluations, with a focus on content and AI safety, adversarial robustness, and real-world capability requirements. Experiments show that Yuvion VL-32B achieves industry-leading safety performance, surpassing comparably sized open-source models and best closed-source commercial models, while maintaining comparable general capabilities.

cs.CV↗

arXiv2026

CogGen: A Cognitively Inspired Recursive Framework for Deep Research Report Generation

The autonomous synthesis of deep research reports represents a critical frontier for Large Language Models (LLMs), demanding sophisticated information orchestration and non-linear narrative logic. Current approaches rely on rigid predefined linear workflows, which cause error accumulation, preclude global restructuring from subsequent insights, and ultimately limit in-depth multimodal fusion and report quality. We propose CogGen, a Cognitively inspired recursive framework for deep research report Generation. Leveraging a Hierarchical Recursive Architecture to simulate cognitive writing, CogGen enables flexible planning and global restructuring. To extend this recursivity to multimodal content, we introduce Abstract Visual Representation (AVR): a concise intent-driven language that iteratively refines visual-text layouts without pixel-level regeneration overhead. We further present CLEF, a Cognitive Load Evaluation Framework, and curate a new benchmark from Our World in Data (OWID). Extensive experiments show CogGen achieves state-of-the-art results among open-source systems, generating reports comparable to professional analysts' outputs and surpassing Gemini Deep Research. Our code and dataset are available at https://github.com/NJUNLP/CogGen.

cs.MA↗

arXiv2025

Self-motion as a structural prior for coherent and robust formation of cognitive maps

Most computational accounts of cognitive maps assume that stability is achieved primarily through sensory anchoring, with self-motion contributing to incremental positional updates only. However, biological spatial representations often remain coherent even when sensory cues degrade or conflict, suggesting that self-motion may play a deeper organizational role. Here, we show that self-motion can act as a structural prior that actively organizes the geometry of learned cognitive maps. We embed a path-integration-based motion prior in a predictive-coding framework, implemented using a capacity-efficient, brain-inspired recurrent mechanism combining spiking dynamics, analog modulation and adaptive thresholds. Across highly aliased, dynamically changing and naturalistic environments, this structural prior consistently stabilizes map formation, improving local topological fidelity, global positional accuracy and next-step prediction under sensory ambiguity. Mechanistic analyses reveal that the motion prior itself encodes geometrically precise trajectories under tight constraints of internal states and generalizes zero-shot to unseen environments, outperforming simpler motion-based constraints. Finally, deployment on a quadrupedal robot demonstrates that motion-derived structural priors enhance online landmark-based navigation under real-world sensory variability. Together, these results reframe self-motion as an organizing scaffold for coherent spatial representations, showing how brain-inspired principles can systematically strengthen spatial intelligence in embodied artificial agents.

q-bio.NC↗

arXiv2025

Batch Normalization-Free Fully Integer Quantized Neural Networks via Progressive Tandem Learning

Quantised neural networks (QNNs) shrink models and reduce inference energy through low-bit arithmetic, yet most still depend on a running statistics batch normalisation (BN) layer, preventing true integer-only deployment. Prior attempts remove BN by parameter folding or tailored initialisation; while helpful, they rarely recover BN's stability and accuracy and often impose bespoke constraints. We present a BN-free, fully integer QNN trained via a progressive, layer-wise distillation scheme that slots into existing low-bit pipelines. Starting from a pretrained BN-enabled teacher, we use layer-wise targets and progressive compensation to train a student that performs inference exclusively with integer arithmetic and contains no BN operations. On ImageNet with AlexNet, the BN-free model attains competitive Top-1 accuracy under aggressive quantisation. The procedure integrates directly with standard quantisation workflows, enabling end-to-end integer-only inference for resource-constrained settings such as edge and embedded devices.

cs.LG↗

arXiv2025

Algorithm-hardware co-design of neuromorphic networks with dual memory pathways

Spiking neural networks excel at event-driven sensing. Yet, maintaining task-relevant context over long timescales both algorithmically and in hardware, while respecting both tight energy and memory budgets, remains a core challenge in the field. We address this challenge through an algorithm-hardware co-design effort. At the algorithm level, inspired by the cortical fast-slow organization in the brain, we introduce a neural network with an explicit slow memory pathway that, combined with fast spiking activity, enables a dual memory pathway (DMP) architecture in which each layer maintains a compact low-dimensional state that summarizes recent activity and modulates spiking dynamics. This explicit memory stabilizes learning while preserving event-driven sparsity, achieving competitive accuracy on long-sequence benchmarks with 40-60% fewer parameters than equivalent state-of-the-art spiking neural networks. At the hardware level, we introduce a near-memory-compute architecture that fully leverages the advantages of the DMP architecture by retaining its compact shared state while optimizing dataflow, across heterogeneous sparse-spike and dense-memory pathways. We show experimental results that demonstrate more than a 4X increase in throughput and over a 5X improvement in energy efficiency compared with state-of-the-art implementations. Together, these contributions demonstrate that biological principles can guide functional abstractions that are both algorithmically effective and hardware-efficient, establishing a scalable co-design framework for real-time neuromorphic computation and learning.

cs.NE↗

arXiv2025

Parallel Delayed Memory Units for Enhanced Temporal Modeling in Biomedical and Bioacoustic Signal Analysis

Advanced deep learning architectures, particularly recurrent neural networks (RNNs), have been widely applied in audio, bioacoustic, and biomedical signal analysis, especially in data-scarce environments. While gated RNNs remain effective, they can be relatively over-parameterised and less training-efficient in some regimes, while linear RNNs tend to fall short in capturing the complexity inherent in bio-signals. To address these challenges, we propose the Parallel Delayed Memory Unit (PDMU), a {delay-gated state-space module for short-term temporal credit assignment} targeting audio and bioacoustic signals, which enhances short-term temporal state interactions and memory efficiency via a gated delay-line mechanism. Unlike previous Delayed Memory Units (DMU) that embed temporal dynamics into the delay-line architecture, the PDMU further compresses temporal information into vector representations using Legendre Memory Units (LMU). This design serves as a form of causal attention, allowing the model to dynamically adjust its reliance on past states and improve real-time learning performance. Notably, in low-information scenarios, the gating mechanism behaves similarly to skip connections by bypassing state decay and preserving early representations, thereby facilitating long-term memory retention. The PDMU is modular, supporting parallel training and sequential inference, and can be easily integrated into existing linear RNN frameworks. Furthermore, we introduce bidirectional, efficient, and spiking variants of the architecture, each offering additional gains in performance or energy efficiency. Experimental results on diverse audio and biomedical benchmarks demonstrate that the PDMU significantly enhances both memory capacity and overall model performance.

cs.SD↗

arXiv2025

Exploiting heterogeneous delays for efficient computation in low-bit neural networks

Neural networks rely on learning synaptic weights. However, this overlooks other neural parameters that can also be learned and may be utilized by the brain. One such parameter is the delay: the brain exhibits complex temporal dynamics with heterogeneous delays, where signals are transmitted asynchronously between neurons. It has been theorized that this delay heterogeneity, rather than a cost to be minimized, can be exploited in embodied contexts where task-relevant information naturally sits contextually in the time domain. We test this hypothesis by training spiking neural networks to modify not only their weights but also their delays at different levels of precision. We find that delay heterogeneity enables state-of-the-art performance on temporally complex neuromorphic problems and can be achieved even when weights are extremely imprecise (1.58-bit ternary precision: just positive, negative, or absent). By enabling high performance with extremely low-precision weights, delay heterogeneity allows memory-efficient solutions that maintain state-of-the-art accuracy even when weights are compressed over an order of magnitude more aggressively than typically studied weight-only networks. We show how delays and time-constants adaptively trade-off, and reveal through ablation that task performance depends on task-appropriate delay distributions, with temporally-complex tasks requiring longer delays. Our results suggest temporal heterogeneity is an important principle for efficient computation, particularly when task-relevant information is temporal - as in the physical world - with implications for embodied intelligent systems and neuromorphic hardware.

cs.NE↗

arXiv2025

Beyond Rate Coding: Surrogate Gradients Enable Spike Timing Learning in Spiking Neural Networks

The surrogate gradient descent algorithm enabled spiking neural networks to be trained to carry out challenging sensory processing tasks, an important step in understanding how spikes contribute to neural computations. However, it is unclear the extent to which these algorithms fully explore the space of possible spiking solutions to problems. We investigated whether spiking networks trained with surrogate gradient descent can learn to make use of information that is only encoded in the timing and not the rate of spikes. We constructed synthetic datasets with a range of types of spike timing information (interspike intervals, spatio-temporal spike patterns or polychrony, and coincidence codes). We find that surrogate gradient descent training can extract all of these types of information. In more realistic speech-based datasets, both timing and rate information is present. We therefore constructed variants of these datasets in which all rate information is removed, and find that surrogate gradient descent can still perform well. We tested all networks both with and without trainable axonal delays. We find that delays can give a significant increase in performance, particularly for more challenging tasks. To determine what types of spike timing information are being used by the networks trained on the speech-based tasks, we test these networks on time-reversed spikes which perturb spatio-temporal spike patterns but leave interspike intervals and coincidence information unchanged. We find that when axonal delays are not used, networks perform well under time reversal, whereas networks trained with delays perform poorly. This suggests that spiking neural networks with delays are better able to exploit temporal structure. To facilitate further studies of temporal coding, we have released our modified speech-based datasets.

cs.NE↗

arXiv2024

EEG decoding with conditional identification information

Decoding EEG signals is crucial for unraveling human brain and advancing brain-computer interfaces. Traditional machine learning algorithms have been hindered by the high noise levels and inherent inter-person variations in EEG signals. Recent advances in deep neural networks (DNNs) have shown promise, owing to their advanced nonlinear modeling capabilities. However, DNN still faces challenge in decoding EEG samples of unseen individuals. To address this, this paper introduces a novel approach by incorporating the conditional identification information of each individual into the neural network, thereby enhancing model representation through the synergistic interaction of EEG and personal traits. We test our model on the WithMe dataset and demonstrated that the inclusion of these identifiers substantially boosts accuracy for both subjects in the training set and unseen subjects. This enhancement suggests promising potential for improving for EEG interpretability and understanding of relevant identification features.

eess.SP↗

arXiv2023

Delayed Memory Unit: Modelling Temporal Dependency Through Delay Gate

Recurrent Neural Networks (RNNs) are widely recognized for their proficiency in modeling temporal dependencies, making them highly prevalent in sequential data processing applications. Nevertheless, vanilla RNNs are confronted with the well-known issue of gradient vanishing and exploding, posing a significant challenge for learning and establishing long-range dependencies. Additionally, gated RNNs tend to be over-parameterized, resulting in poor computational efficiency and network generalization. To address these challenges, this paper proposes a novel Delayed Memory Unit (DMU). The DMU incorporates a delay line structure along with delay gates into vanilla RNN, thereby enhancing temporal interaction and facilitating temporal credit assignment. Specifically, the DMU is designed to directly distribute the input information to the optimal time instant in the future, rather than aggregating and redistributing it over time through intricate network dynamics. Our proposed DMU demonstrates superior temporal modeling capabilities across a broad range of sequential modeling tasks, utilizing considerably fewer parameters than other state-of-the-art gated RNN models in applications such as speech recognition, radar gesture recognition, ECG waveform segmentation, and permuted sequential image classification.

cs.NE↗

arXiv2023

Adaptive Axonal Delays in feedforward spiking neural networks for accurate spoken word recognition

Spiking neural networks (SNN) are a promising research avenue for building accurate and efficient automatic speech recognition systems. Recent advances in audio-to-spike encoding and training algorithms enable SNN to be applied in practical tasks. Biologically-inspired SNN communicates using sparse asynchronous events. Therefore, spike-timing is critical to SNN performance. In this aspect, most works focus on training synaptic weights and few have considered delays in event transmission, namely axonal delay. In this work, we consider a learnable axonal delay capped at a maximum value, which can be adapted according to the axonal delay distribution in each network layer. We show that our proposed method achieves the best classification results reported on the SHD dataset (92.45%) and NTIDIGITS dataset (95.09%). Our work illustrates the potential of training axonal delays for tasks with complex temporal structures.

cs.NE↗

arXiv2022

Spikemax: Spike-based Loss Methods for Classification

Spiking Neural Networks~(SNNs) are a promising research paradigm for low power edge-based computing. Recent works in SNN backpropagation has enabled training of SNNs for practical tasks. However, since spikes are binary events in time, standard loss formulations are not directly compatible with spike output. As a result, current works are limited to using mean-squared loss of spike count. In this paper, we formulate the output probability interpretation from the spike count measure and introduce spike-based negative log-likelihood measure which are more suited for classification tasks especially in terms of the energy efficiency and inference latency. We compare our loss measures with other existing alternatives and evaluate using classification performances on three neuromorphic benchmark datasets: NMNIST, DVS Gesture and N-TIDIGITS18. In addition, we demonstrate state of the art performances on these datasets, achieving faster inference speed and less energy consumption.

cs.NE↗

arXiv2022

Axonal Delay As a Short-Term Memory for Feed Forward Deep Spiking Neural Networks

The information of spiking neural networks (SNNs) are propagated between the adjacent biological neuron by spikes, which provides a computing paradigm with the promise of simulating the human brain. Recent studies have found that the time delay of neurons plays an important role in the learning process. Therefore, configuring the precise timing of the spike is a promising direction for understanding and improving the transmission process of temporal information in SNNs. However, most of the existing learning methods for spiking neurons are focusing on the adjustment of synaptic weight, while very few research has been working on axonal delay. In this paper, we verify the effectiveness of integrating time delay into supervised learning and propose a module that modulates the axonal delay through short-term memory. To this end, a rectified axonal delay (RAD) module is integrated with the spiking model to align the spike timing and thus improve the characterization learning ability of temporal features. Experiments on three neuromorphic benchmark datasets : NMNIST, DVS Gesture and N-TIDIGITS18 show that the proposed method achieves the state-of-the-art performance while using the fewest parameters.

cs.NE↗

arXiv2021

Let Your Camera See for You: A Novel Two-Factor Authentication Method against Real-Time Phishing Attacks

Today, two-factor authentication (2FA) is a widely implemented mechanism to counter phishing attacks. Although much effort has been investigated in 2FA, most 2FA systems are still vulnerable to carefully designed phishing attacks, and some even request special hardware, which limits their wide deployment. Recently, real-time phishing (RTP) has made the situation even worse because an adversary can effortlessly establish a phishing website replicating a target website without any background of the web page design technique. Traditional 2FA can be easily bypassed by such RTP attacks. In this work, we propose a novel 2FA system to counter RTP attacks. The main idea is to request a user to take a photo of the web browser with the domain name in the address bar as the 2nd authentication factor. The web server side extracts the domain name information based on Optical Character Recognition (OCR), and then determines if the user is visiting this website or a fake one, thus defeating the RTP attacks where an adversary must set up a fake website with a different domain. We prototyped our system and evaluated its performance in various environments. The results showed that PhotoAuth is an effective technique with good scalability. We also showed that compared to other 2FA systems, PhotoAuth has several advantages, especially no special hardware or software support is needed on the client side except a phone, making it readily deployable.

cs.CR↗

arXiv2021

NTIRE 2021 Challenge on Perceptual Image Quality Assessment

This paper reports on the NTIRE 2021 challenge on perceptual image quality assessment (IQA), held in conjunction with the New Trends in Image Restoration and Enhancement workshop (NTIRE) workshop at CVPR 2021. As a new type of image processing technology, perceptual image processing algorithms based on Generative Adversarial Networks (GAN) have produced images with more realistic textures. These output images have completely different characteristics from traditional distortions, thus pose a new challenge for IQA methods to evaluate their visual quality. In comparison with previous IQA challenges, the training and testing datasets in this challenge include the outputs of perceptual image processing algorithms and the corresponding subjective scores. Thus they can be used to develop and evaluate IQA methods on GAN-based distortions. The challenge has 270 registered participants in total. In the final testing stage, 13 participating teams submitted their models and fact sheets. Almost all of them have achieved much better results than existing IQA methods, while the winning method can demonstrate state-of-the-art performance.

eess.IV↗

arXiv2019

GLMNet: Graph Learning-Matching Networks for Feature Matching

Recently, graph convolutional networks (GCNs) have shown great potential for the task of graph matching. It can integrate graph node feature embedding, node-wise affinity learning and matching optimization together in a unified end-to-end model. One important aspect of graph matching is the construction of two matching graphs. However, the matching graphs we feed to existing graph convolutional matching networks are generally fixed and independent of graph matching, which thus are not guaranteed to be optimal for the graph matching task. Also, existing GCN matching method employs several general smoothing-based graph convolutional layers to generate graph node embeddings, in which extensive smoothing convolution operation may dilute the desired discriminatory information of graph nodes. To overcome these issues, we propose a novel Graph Learning-Matching Network (GLMNet) for graph matching problem. GLMNet has three main aspects. (1) It integrates graph learning into graph matching which thus adaptively learn a pair of optimal graphs that best serve graph matching task. (2) It further employs a Laplacian sharpening convolutional module to generate more discriminative node embeddings for graph matching. (3) A new constraint regularized loss is designed for GLMNet training which can encode the desired one-to-one matching constraints in matching optimization. Experiments on two benchmarks demonstrate the effectiveness of GLMNet and advantages of its main modules.

cs.CV↗