Selected work

Representative Papers

MetaGDPO: Alleviating Catastrophic Forgetting with Metacognitive Knowledge through Group Direct Preference Optimization

Nov 15, 2025

Small language models (<8B) suffer from catastrophic forgetting during knowledge distillation due to (i) misalignment between training data and the model’s intrinsic capabilities, and (ii) the absence of explicit constraints preserving previously acquired knowledge in conventional objectives. Method: We propose a dual-path solution: (1) a novel 5K-sample multitask reasoning dataset—first to incorporate metacognitive knowledge annotations—paired with a task-capability matching data selection mechanism; and (2) GDPO (Group Direct Preference Optimization), a preference-based optimization framework that leverages a reference model to implicitly guide parameter updates, jointly optimizing knowledge retention and transfer. Results: Experiments demonstrate significant mitigation of catastrophic forgetting; under resource-constrained settings, our approach approaches the performance of GRPO while substantially enhancing both reasoning capability and knowledge retention in small models.

0 citationsRead paper

Value-Aligned Prompt Moderation via Zero-Shot Agentic Rewriting for Safe Image Generation

Nov 12, 2025

Generative vision-language models (e.g., Stable Diffusion) excel in creative image synthesis but remain vulnerable to adversarial prompts that elicit unsafe, offensive, or culturally inappropriate outputs; existing defenses often compromise image quality or incur substantial computational overhead. This paper introduces VALOR—a modular, zero-shot proxy framework that enhances safety and utility in text-to-image generation via hierarchical prompt analysis and value-aligned reasoning. Its core contributions include: (i) integrated multi-level NSFW detection, cultural-value alignment, and intent disambiguation; (ii) LLM-driven selective prompt rewriting and optional stylistic regeneration; and (iii) semantics-preserving safe regeneration with dynamic role-instruction adaptation. Experiments demonstrate that VALOR achieves up to 100% suppression of unsafe outputs across diverse adversarial and culturally sensitive prompts, while preserving prompt fidelity, creativity, and functional utility.

0 citationsRead paper

Neutral Agent-based Adversarial Policy Learning against Deep Reinforcement Learning in Multi-party Open Systems

Oct 12, 2025

In multi-agent open systems, deep reinforcement learning (DRL) agents are vulnerable to adversarial attacks; however, existing methods require either full environmental control or direct interaction with the victim agent, severely limiting practical applicability. Method: This paper proposes an indirect adversarial strategy learning framework leveraging a *neutral agent*—an autonomous entity deployed in the shared environment that neither interacts with the victim nor controls the environment, yet exerts implicit interference to mislead the victim’s decision-making. The approach integrates DRL with adversarial policy optimization to generate robust, generalizable deceptive behaviors. Results: Experiments on SMAC and Highway-env demonstrate cross-scenario effectiveness: the method consistently misleads high-performance victim agents while significantly improving attack stealthiness, transferability, and real-world feasibility compared to conventional approaches reliant on strong assumptions.

0 citationsRead paper

LPCVAE: A Conditional VAE with Long-Term Dependency and Probabilistic Time-Frequency Fusion for Time Series Anomaly Detection

Oct 12, 2025

Existing VAE-based time-series anomaly detection methods suffer from three key limitations: reliance on single-window modeling, insufficient capture of long-term temporal dependencies, and inadequate fusion of time-frequency information. To address these issues, this paper proposes a conditional variational autoencoder (CVAE) framework. The framework incorporates LSTM networks to model cross-window long-term dependencies, employs short-time Fourier transform (STFT) to extract time-frequency features, and introduces a distribution-level adaptive fusion mechanism based on Product of Experts (PoE) to enable probabilistic collaborative modeling of time-domain and frequency-domain latent representations. Extensive experiments on four public benchmark datasets demonstrate that the proposed method consistently outperforms state-of-the-art baselines. Ablation studies further confirm that both long-term dependency modeling and adaptive time-frequency fusion are critical for enhancing detection accuracy and robustness.

0 citationsRead paper

AILoRA: Function-Aware Asymmetric Initialization for Low-Rank Adaptation of Large Language Models

Oct 09, 2025

To address suboptimal performance and slow convergence of Low-Rank Adaptation (LoRA) in fine-tuning large language models, this paper proposes a function-aware asymmetric LoRA initialization. We first identify the functional asymmetry between query (W^Q) and value (W^V) projection matrices in self-attention: W^Q primarily governs task-specific modeling, whereas W^V emphasizes general feature extraction. Leveraging this insight, we introduce a principal–auxiliary component co-initialization strategy—assigning high signal-to-noise-ratio principal components to W^Q-LoRA to enhance task adaptation, while retaining low-rank auxiliary components for W^V-LoRA to preserve representational generality. Our method operates within standard LoRA architecture and enables efficient deployment via singular value decomposition (SVD). Experiments demonstrate that, with minimal parameter overhead (<0.1%), our approach accelerates convergence by 1.8× on average and improves downstream task accuracy by +2.3% on average, consistently outperforming vanilla LoRA.

0 citationsRead paper

Recent publications

Latest Papers

MetaGDPO: Alleviating Catastrophic Forgetting with Metacognitive Knowledge through Group Direct Preference Optimization

Nov 15, 2025

Small language models (<8B) suffer from catastrophic forgetting during knowledge distillation due to (i) misalignment between training data and the model’s intrinsic capabilities, and (ii) the absence of explicit constraints preserving previously acquired knowledge in conventional objectives. Method: We propose a dual-path solution: (1) a novel 5K-sample multitask reasoning dataset—first to incorporate metacognitive knowledge annotations—paired with a task-capability matching data selection mechanism; and (2) GDPO (Group Direct Preference Optimization), a preference-based optimization framework that leverages a reference model to implicitly guide parameter updates, jointly optimizing knowledge retention and transfer. Results: Experiments demonstrate significant mitigation of catastrophic forgetting; under resource-constrained settings, our approach approaches the performance of GRPO while substantially enhancing both reasoning capability and knowledge retention in small models.

0 citationsRead paper

Value-Aligned Prompt Moderation via Zero-Shot Agentic Rewriting for Safe Image Generation

Nov 12, 2025

Generative vision-language models (e.g., Stable Diffusion) excel in creative image synthesis but remain vulnerable to adversarial prompts that elicit unsafe, offensive, or culturally inappropriate outputs; existing defenses often compromise image quality or incur substantial computational overhead. This paper introduces VALOR—a modular, zero-shot proxy framework that enhances safety and utility in text-to-image generation via hierarchical prompt analysis and value-aligned reasoning. Its core contributions include: (i) integrated multi-level NSFW detection, cultural-value alignment, and intent disambiguation; (ii) LLM-driven selective prompt rewriting and optional stylistic regeneration; and (iii) semantics-preserving safe regeneration with dynamic role-instruction adaptation. Experiments demonstrate that VALOR achieves up to 100% suppression of unsafe outputs across diverse adversarial and culturally sensitive prompts, while preserving prompt fidelity, creativity, and functional utility.

0 citationsRead paper

Neutral Agent-based Adversarial Policy Learning against Deep Reinforcement Learning in Multi-party Open Systems

Oct 12, 2025

In multi-agent open systems, deep reinforcement learning (DRL) agents are vulnerable to adversarial attacks; however, existing methods require either full environmental control or direct interaction with the victim agent, severely limiting practical applicability. Method: This paper proposes an indirect adversarial strategy learning framework leveraging a *neutral agent*—an autonomous entity deployed in the shared environment that neither interacts with the victim nor controls the environment, yet exerts implicit interference to mislead the victim’s decision-making. The approach integrates DRL with adversarial policy optimization to generate robust, generalizable deceptive behaviors. Results: Experiments on SMAC and Highway-env demonstrate cross-scenario effectiveness: the method consistently misleads high-performance victim agents while significantly improving attack stealthiness, transferability, and real-world feasibility compared to conventional approaches reliant on strong assumptions.

0 citationsRead paper

LPCVAE: A Conditional VAE with Long-Term Dependency and Probabilistic Time-Frequency Fusion for Time Series Anomaly Detection

Oct 12, 2025

Existing VAE-based time-series anomaly detection methods suffer from three key limitations: reliance on single-window modeling, insufficient capture of long-term temporal dependencies, and inadequate fusion of time-frequency information. To address these issues, this paper proposes a conditional variational autoencoder (CVAE) framework. The framework incorporates LSTM networks to model cross-window long-term dependencies, employs short-time Fourier transform (STFT) to extract time-frequency features, and introduces a distribution-level adaptive fusion mechanism based on Product of Experts (PoE) to enable probabilistic collaborative modeling of time-domain and frequency-domain latent representations. Extensive experiments on four public benchmark datasets demonstrate that the proposed method consistently outperforms state-of-the-art baselines. Ablation studies further confirm that both long-term dependency modeling and adaptive time-frequency fusion are critical for enhancing detection accuracy and robustness.

0 citationsRead paper

AILoRA: Function-Aware Asymmetric Initialization for Low-Rank Adaptation of Large Language Models

Oct 09, 2025

To address suboptimal performance and slow convergence of Low-Rank Adaptation (LoRA) in fine-tuning large language models, this paper proposes a function-aware asymmetric LoRA initialization. We first identify the functional asymmetry between query (W^Q) and value (W^V) projection matrices in self-attention: W^Q primarily governs task-specific modeling, whereas W^V emphasizes general feature extraction. Leveraging this insight, we introduce a principal–auxiliary component co-initialization strategy—assigning high signal-to-noise-ratio principal components to W^Q-LoRA to enhance task adaptation, while retaining low-rank auxiliary components for W^V-LoRA to preserve representational generality. Our method operates within standard LoRA architecture and enables efficient deployment via singular value decomposition (SVD). Experiments demonstrate that, with minimal parameter overhead (<0.1%), our approach accelerates convergence by 1.8× on average and improves downstream task accuracy by +2.3% on average, consistently outperforming vanilla LoRA.

0 citationsRead paper