arXiv subjects

Cen Chen

Publications and source records attributed to Cen Chen.

At least 37 records · Page 2Linked to original sources

arXiv2026

Spectral Evolution Search: Efficient Inference-Time Scaling for Reward-Aligned Image Generation

Inference-time scaling offers a versatile paradigm for aligning visual generative models with downstream objectives without parameter updates. However, existing approaches that optimize the high-dimensional initial noise suffer from severe inefficiency, as many search directions exert negligible influence on the final generation. We show that this inefficiency is closely related to a spectral bias in generative dynamics: model sensitivity to initial perturbations diminishes rapidly as frequency increases. Building on this insight, we propose Spectral Evolution Search (SES), a plug-and-play framework for initial noise optimization that executes gradient-free evolutionary search within a low-frequency subspace. Theoretically, we derive the Spectral Scaling Prediction from perturbation propagation dynamics, which explains the systematic differences in the impact of perturbations across frequencies. Extensive experiments demonstrate that SES significantly advances the Pareto frontier of generation quality versus computational cost, consistently outperforming strong baselines under equivalent budgets.

cs.LG↗

arXiv2026

VIRAL: Visual In-Context Reasoning via Analogy in Diffusion Transformers

Replicating In-Context Learning (ICL) in computer vision remains challenging due to task heterogeneity. We propose \textbf{VIRAL}, a framework that elicits visual reasoning from a pre-trained image editing model by formulating ICL as conditional generation via visual analogy ($x_s : x_t :: x_q : y_q$). We adapt a frozen Diffusion Transformer (DiT) using role-aware multi-image conditioning and introduce a Mixture-of-Experts LoRA to mitigate gradient interference across diverse tasks. Additionally, to bridge the gaps in current visual context datasets, we curate a large-scale dataset spanning perception, restoration, and editing. Experiments demonstrate that VIRAL outperforms existing methods, validating that a unified V-ICL paradigm can handle the majority of visual tasks, including open-domain editing. Our code is available at https://anonymous.4open.science/r/VIRAL-744A

cs.CV↗

arXiv2026

When Sharpening Becomes Collapse: Sampling Bias and Semantic Coupling in RL with Verifiable Rewards

Reinforcement Learning with Verifiable Rewards (RLVR) is a central paradigm for turning large language models (LLMs) into reliable problem solvers, especially in logic-heavy domains. Despite its empirical success, it remains unclear whether RLVR elicits novel capabilities or merely sharpens the distribution over existing knowledge. We study this by formalizing over-sharpening, a phenomenon where the policy collapses onto limited modes, suppressing valid alternatives. At a high level, we discover finite-batch updates intrinsically bias learning toward sampled modes, triggering a collapse that propagates globally via semantic coupling. To mitigate this, we propose inverse-success advantage calibration to prioritize difficult queries and distribution-level calibration to diversify sampling via a memory network. Empirical evaluations validate that our strategies can effectively improve generalization.

cs.LG↗

arXiv2026

RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior

Large Language Models (LLMs) with long chain-of-thought (CoT) capability, termed Reasoning Models, demonstrate superior intricate problem-solving abilities through multi-step long CoT reasoning. To create a dual-capability model with long CoT capability and domain-specific knowledge without substantial computational and data costs, model merging emerges as a highly resource-efficient method. However, significant challenges lie in merging domain-specific LLMs with long CoT ones since nowadays merging methods suffer from reasoning capability degradation, even gibberish output and output collapse. To overcome this, we introduce RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior, a novel merging framework designed to integrate domain-specific LLMs with long CoT capability, meanwhile maintaining model performance in the original domain. Treating reasoning model weights as foundational prior, our method utilizes a reasoning capability indicator to preserve core long CoT capability model weights while selectively merging essential domain-specific weights. We conducted extensive experiments on Qwen2.5-7B, Llama3.1-8B, and Qwen2.5-1.5B models in BioMedicine and Finance domains. Our results show that RCP-Merging successfully merges a reasoning model with domain-specific ones, improving domain task performance by 9.5% and 9.2% over state-of-the-art methods, without significantly harming the original long CoT reasoning capability.

cs.CL↗

arXiv2026

LSSF: Safety Alignment for Large Language Models through Low-Rank Safety Subspace Fusion

The safety mechanisms of large language models (LLMs) exhibit notable fragility, as even fine-tuning on datasets without harmful content may still undermine their safety capabilities. Meanwhile, existing safety alignment methods predominantly rely on the fine-tuning process, which inadvertently leads to the increased complexity and computational resources required. To address these issues, we introduce LSSF, a novel safety re-alignment framework with \underline{L}ow-Rank \underline{S}afety \underline{S}ubspace \underline{F}usion. Our proposed method exploits the low-rank characteristics of safety information in LLMs by constructing a low-rank projection matrix to extract the principal components of safety vectors. Notably, this projection matrix represents the low-rank safety subspace of the LLMs, which we have observed to remain stable during fine-tuning process and is isolated from the model's general capabilities. These principal components are used to effectively restore safety alignment when combined with fine-tuned LLMs through linear arithmetic. Additionally, to account for the varying encoding densities of safety information across different layers of LLMs, we propose a novel metric called safety singular value entropy. This metric quantifies the encoding density and allows for the dynamic computation of the safety-critical rank for each safety vector. Extensive experiments demonstrate that our proposed post-hoc alignment method can effectively restore the safety alignment of fine-tuned models with minimal impact on their performance in downstream tasks.

cs.CY↗

arXiv2026

AttriCtrl: Fine-Grained Control of Aesthetic Attribute Intensity in Diffusion Models

Diffusion models have recently become the dominant paradigm for image generation, yet existing systems struggle to interpret and follow numeric instructions for adjusting semantic attributes. In real-world creative scenarios, especially when precise control over aesthetic attributes is required, current methods fail to provide such controllability. This limitation partly arises from the subjective and context-dependent nature of aesthetic judgments, but more fundamentally stems from the fact that current text encoders are designed for discrete tokens rather than continuous values. Meanwhile, efforts on aesthetic alignment, often leveraging reinforcement learning, direct preference optimization, or architectural modifications, primarily align models with a global notion of human preference. While these approaches improve user experience, they overlook the multifaceted and compositional nature of aesthetics, underscoring the need for explicit disentanglement and independent control of aesthetic attributes. To address this gap, we introduce AttriCtrl, a lightweight framework for continuous aesthetic intensity control in diffusion models. It first defines relevant aesthetic attributes, then quantifies them through a hybrid strategy that maps both concrete and abstract dimensions onto a unified $[0,1]$ scale. A plug-and-play value encoder is then used to transform user-specified values into model-interpretable embeddings for controllable generation. Experiments show that AttriCtrl achieves accurate and continuous control over both single and multiple aesthetic attributes, significantly enhancing personalization and diversity. Crucially, it is implemented as a lightweight adapter while keeping the diffusion model frozen, ensuring seamless integration with existing frameworks such as ControlNet at negligible computational cost.

cs.CV↗

arXiv2025

MixKVQ: Query-Aware Mixed-Precision KV Cache Quantization for Long-Context Reasoning

Long Chain-of-Thought (CoT) reasoning has significantly advanced the capabilities of Large Language Models (LLMs), but this progress is accompanied by substantial memory and latency overhead from the extensive Key-Value (KV) cache. Although KV cache quantization is a promising compression technique, existing low-bit quantization methods often exhibit severe performance degradation on complex reasoning tasks. Fixed-precision quantization struggles to handle outlier channels in the key cache, while current mixed-precision strategies fail to accurately identify components requiring high-precision representation. We find that an effective low-bit KV cache quantization strategy must consider two factors: a key channel's intrinsic quantization difficulty and its relevance to the query. Based on this insight, we propose MixKVQ, a novel plug-and-play method that introduces a lightweight, query-aware algorithm to identify and preserve critical key channels that need higher precision, while applying per-token quantization for value cache. Experiments on complex reasoning datasets demonstrate that our approach significantly outperforms existing low-bit methods, achieving performance comparable to a full-precision baseline at a substantially reduced memory footprint.

cs.LG↗

arXiv2025

REAL: Representation Enhanced Analytic Learning for Exemplar-free Class-incremental Learning

Exemplar-free class-incremental learning (EFCIL) aims to mitigate catastrophic forgetting in class-incremental learning (CIL) without available historical training samples as exemplars. Compared with its exemplar-based CIL counterpart that stores exemplars, EFCIL suffers more from forgetting issues. Recently, a new EFCIL branch named Analytic Continual Learning (ACL) introduces a gradient-free paradigm via Recursive Least-Square, achieving a forgetting-resistant classifier training with a frozen backbone during CIL. However, existing ACL suffers from ineffective representations and insufficient utilization of backbone knowledge. In this paper, we propose a representation-enhanced analytic learning (REAL) to address these problems. To enhance the representation, REAL constructs a dual-stream base pretraining followed by representation enhancing distillation process. The dual-stream base pretraining combines self-supervised contrastive learning for general features and supervised learning for class-specific knowledge, followed by the representation enhancing distillation to merge both streams, enhancing representations for subsequent CIL paradigm. To utilize more knowledge from the backbone, REAL presents a feature fusion buffer to multi-layer backbone features, providing informative features for the subsequent classifier training. Our method can be incorporated into existing ACL techniques and provides more competitive performance. Empirical results demonstrate that, REAL achieves state-of-the-art performance on CIFAR-100, ImageNet-100 and ImageNet-1k benchmarks, outperforming exemplar-free methods and rivaling exemplar-based approaches.

cs.LG↗

arXiv2025

ARGUS: Defending Against Multimodal Indirect Prompt Injection via Steering Instruction-Following Behavior

Multimodal Large Language Models (MLLMs) are increasingly vulnerable to multimodal Indirect Prompt Injection (IPI) attacks, which embed malicious instructions in images, videos, or audio to hijack model behavior. Existing defenses, designed primarily for text-only LLMs, are unsuitable for countering these multimodal threats, as they are easily bypassed, modality-dependent, or generalize poorly. Inspired by activation steering researches, we hypothesize that a robust, general defense independent of modality can be achieved by steering the model's behavior in the representation space. Through extensive experiments, we discover that the instruction-following behavior of MLLMs is encoded in a subspace. Steering along directions within this subspace can enforce adherence to user instructions, forming the basis of a defense. However, we also found that a naive defense direction could be coupled with a utility-degrading direction, and excessive intervention strength harms model performance. To address this, we propose ARGUS, which searches for an optimal defense direction within the safety subspace that decouples from the utility degradation direction, further combining adaptive strength steering to achieve a better safety-utility trade-off. ARGUS also introduces lightweight injection detection stage to activate the defense on-demand, and a post-filtering stage to verify defense success. Experimental results show that ARGUS can achieve robust defense against multimodal IPI while maximally preserving the MLLM's utility.

cs.CR↗

arXiv2025

CSMD: Curated Multimodal Dataset for Chinese Stock Analysis

The stock market is a complex and dynamic system, where it is non-trivial for researchers and practitioners to uncover underlying patterns and forecast stock movements. The existing studies for stock market analysis rely on leveraging various types of information to extract useful factors, which are highly conditional on the quality of the data used. However, the currently available resources are mainly based on the U.S. stock market in English, which is inapplicable to adapt to other countries. To address these issues, we propose CSMD, a multimodal dataset curated specifically for analyzing the Chinese stock market with meticulous processing for validated quality. In addition, we develop a lightweight and user-friendly framework LightQuant for researchers and practitioners with expertise in financial domains. Experimental results on top of our datasets and framework with various backbone models demonstrate their effectiveness compared with using existing datasets. The datasets and code are publicly available at the link: https://github.com/ECNU-CILAB/LightQuant.

cs.CE↗

arXiv2025

Kelp: A Streaming Safeguard for Large Models via Latent Dynamics-Guided Risk Detection

Large models (LMs) are powerful content generators, yet their open-ended nature can also introduce potential risks, such as generating harmful or biased content. Existing guardrails mostly perform post-hoc detection that may expose unsafe content before it is caught, and the latency constraints further push them toward lightweight models, limiting detection accuracy. In this work, we propose Kelp, a novel plug-in framework that enables streaming risk detection within the LM generation pipeline. Kelp leverages intermediate LM hidden states through a Streaming Latent Dynamics Head (SLD), which models the temporal evolution of risk across the generated sequence for more accurate real-time risk detection. To ensure reliable streaming moderation in real applications, we introduce an Anchored Temporal Consistency (ATC) loss to enforce monotonic harm predictions by embedding a benign-then-harmful temporal prior. Besides, for a rigorous evaluation of streaming guardrails, we also present StreamGuardBench-a model-grounded benchmark featuring on-the-fly responses from each protected model, reflecting real-world streaming scenarios in both text and vision-language tasks. Across diverse models and datasets, Kelp consistently outperforms state-of-the-art post-hoc guardrails and prior plug-in probes (15.61% higher average F1), while using only 20M parameters and adding less than 0.5 ms of per-token latency.

cs.LG↗

arXiv2025

ShieldMMU: Detecting and Defending against Controlled-Channel Attacks in Shielding Memory System

Intel SGX and hypervisors isolate non-privileged programs from other software, ensuring confidentiality and integrity. However, side-channel attacks continue to threaten Intel SGX's security, enabling malicious OS to manipulate PTE present bits, induce page faults, and steal memory access traces. Despite extensive research, existing defenses focus on detection or rely on impractical solutions. This paper presents ShieldMMU, a comprehensive solution for mitigating controlled channel attacks, balancing compatibility, performance, and usability. Leveraging a Merkle Tree-inspired Defense Tree (DD-Tree), ShieldMMU protects PTE integrity by detecting, locating, and restoring attacked PTEs. It identifies MMU page table lookup events and side-channel attacks, promptly restoring PTE parameters to prevent page fault traps and ensure secure non-privileged application operation within SGX. Our experiments confirm ShieldMMU's enhanced security and acceptable latency performance.

cs.CR↗

arXiv2025

RewardDS: Privacy-Preserving Fine-Tuning for Large Language Models via Reward Driven Data Synthesis

The success of large language models (LLMs) has attracted many individuals to fine-tune them for domain-specific tasks by uploading their data. However, in sensitive areas like healthcare and finance, privacy concerns often arise. One promising solution is to generate synthetic data with Differential Privacy (DP) guarantees to replace private data. However, these synthetic data contain significant flawed data, which are considered as noise. Existing solutions typically rely on naive filtering by comparing ROUGE-L scores or embedding similarities, which are ineffective in addressing the noise. To address this issue, we propose \textit{RewardDS}, a novel privacy-preserving framework that fine-tunes a reward proxy model and uses reward signals to guide the synthetic data generation. Our \textit{RewardDS} introduces two key modules, Reward Guided Filtering and Self-Optimizing Refinement, to both filter and refine the synthetic data, effectively mitigating the noise. Extensive experiments across medical, financial, and code generation domains demonstrate the effectiveness of our method.

cs.CR↗

arXiv2025

Towards Instance-wise Personalized Federated Learning via Semi-Implicit Bayesian Prompt Tuning

Federated learning (FL) is a privacy-preserving machine learning paradigm that enables collaborative model training across multiple distributed clients without disclosing their raw data. Personalized federated learning (pFL) has gained increasing attention for its ability to address data heterogeneity. However, most existing pFL methods assume that each client's data follows a single distribution and learn one client-level personalized model for each client. This assumption often fails in practice, where a single client may possess data from multiple sources or domains, resulting in significant intra-client heterogeneity and suboptimal performance. To tackle this challenge, we propose pFedBayesPT, a fine-grained instance-wise pFL framework based on visual prompt tuning. Specifically, we formulate instance-wise prompt generation from a Bayesian perspective and model the prompt posterior as an implicit distribution to capture diverse visual semantics. We derive a variational training objective under the semi-implicit variational inference framework. Extensive experiments on benchmark datasets demonstrate that pFedBayesPT consistently outperforms existing pFL methods under both feature and label heterogeneity settings.

cs.LG↗

arXiv2025

AutoLoRA: Automatic LoRA Retrieval and Fine-Grained Gated Fusion for Text-to-Image Generation

Despite recent advances in photorealistic image generation through large-scale models like FLUX and Stable Diffusion v3, the practical deployment of these architectures remains constrained by their inherent intractability to parameter fine-tuning. While low-rank adaptation (LoRA) have demonstrated efficacy in enabling model customization with minimal parameter overhead, the effective utilization of distributed open-source LoRA modules faces three critical challenges: sparse metadata annotation, the requirement for zero-shot adaptation capabilities, and suboptimal fusion strategies for multi-LoRA fusion strategies. To address these limitations, we introduce a novel framework that enables semantic-driven LoRA retrieval and dynamic aggregation through two key components: (1) weight encoding-base LoRA retriever that establishes a shared semantic space between LoRA parameter matrices and text prompts, eliminating dependence on original training data, and (2) fine-grained gated fusion mechanism that computes context-specific fusion weights across network layers and diffusion timesteps to optimally integrate multiple LoRA modules during generation. Our approach achieves significant improvement in image generation perfermance, thereby facilitating scalable and data-efficient enhancement of foundational models. This work establishes a critical bridge between the fragmented landscape of community-developed LoRAs and practical deployment requirements, enabling collaborative model evolution through standardized adapter integration.

cs.CV↗

arXiv2025

Analytic Continual Test-Time Adaptation for Multi-Modality Corruption

Test-Time Adaptation (TTA) enables pre-trained models to bridge the gap between source and target datasets using unlabeled test data, addressing domain shifts caused by corruptions like weather changes, noise, or sensor malfunctions in test time. Multi-Modal Continual Test-Time Adaptation (MM-CTTA), as an extension of standard TTA, further allows models to handle multi-modal inputs and adapt to continuously evolving target domains. However, MM-CTTA faces critical challenges such as catastrophic forgetting and reliability bias, which are rarely addressed effectively under multi-modal corruption scenarios. In this paper, we propose a novel approach, Multi-modality Dynamic Analytic Adapter (MDAA), to tackle MM-CTTA tasks. MDAA introduces analytic learning,a closed-form training technique,through Analytic Classifiers (ACs) to mitigate catastrophic forgetting. Furthermore, we design the Dynamic Late Fusion Mechanism (DLFM) to dynamically select and integrate reliable information from different modalities. Extensive experiments show that MDAA achieves state-of-the-art performance across the proposed tasks.

cs.LG↗

arXiv2025

GSDNet: Revisiting Incomplete Multimodal-Diffusion from Graph Spectrum Perspective for Conversation Emotion Recognition

Multimodal emotion recognition in conversations (MERC) aims to infer the speaker's emotional state by analyzing utterance information from multiple sources (i.e., video, audio, and text). Compared with unimodality, a more robust utterance representation can be obtained by fusing complementary semantic information from different modalities. However, the modality missing problem severely limits the performance of MERC in practical scenarios. Recent work has achieved impressive performance on modality completion using graph neural networks and diffusion models, respectively. This inspires us to combine these two dimensions through the graph diffusion model to obtain more powerful modal recovery capabilities. Unfortunately, existing graph diffusion models may destroy the connectivity and local structure of the graph by directly adding Gaussian noise to the adjacency matrix, resulting in the generated graph data being unable to retain the semantic and topological information of the original graph. To this end, we propose a novel Graph Spectral Diffusion Network (GSDNet), which maps Gaussian noise to the graph spectral space of missing modalities and recovers the missing data according to its original distribution. Compared with previous graph diffusion methods, GSDNet only affects the eigenvalues of the adjacency matrix instead of destroying the adjacency matrix directly, which can maintain the global topological information and important spectral features during the diffusion process. Extensive experiments have demonstrated that GSDNet achieves state-of-the-art emotion recognition performance in various modality loss scenarios.

cs.SD↗

arXiv2025

Refiner: Data Refining against Gradient Leakage Attacks in Federated Learning

Recent works have brought attention to the vulnerability of Federated Learning (FL) systems to gradient leakage attacks. Such attacks exploit clients' uploaded gradients to reconstruct their sensitive data, thereby compromising the privacy protection capability of FL. In response, various defense mechanisms have been proposed to mitigate this threat by manipulating the uploaded gradients. Unfortunately, empirical evaluations have demonstrated limited resilience of these defenses against sophisticated attacks, indicating an urgent need for more effective defenses. In this paper, we explore a novel defensive paradigm that departs from conventional gradient perturbation approaches and instead focuses on the construction of robust data. Intuitively, if robust data exhibits low semantic similarity with clients' raw data, the gradients associated with robust data can effectively obfuscate attackers. To this end, we design Refiner that jointly optimizes two metrics for privacy protection and performance maintenance. The utility metric is designed to promote consistency between the gradients of key parameters associated with robust data and those derived from clients' data, thus maintaining model performance. Furthermore, the privacy metric guides the generation of robust data towards enlarging the semantic gap with clients' data. Theoretical analysis supports the effectiveness of Refiner, and empirical evaluations on multiple benchmark datasets demonstrate the superior defense effectiveness of Refiner at defending against state-of-the-art attacks.

cs.LG↗