Selected work

Representative Papers

MagicGUI-RMS: A Multi-Agent Reward Model System for Self-Evolving GUI Agents via Automated Feedback Reflux

Jan 19, 2026

This work addresses the lack of efficient, scalable automated evaluation and continual learning mechanisms for GUI agents by proposing a multi-agent reward framework that integrates a domain-specific reward model (DS-RM) with a general-purpose reward model (GP-RM). The approach enables fine-grained behavioral scoring, error correction, and self-evolutionary learning through collaborative assessment, coupled with automatic construction of structured reward data and a feedback reflux mechanism that eliminates the need for manual annotation. Experimental results demonstrate that the framework significantly improves task accuracy and behavioral robustness, establishing an efficient and scalable reward-driven paradigm for self-evolving GUI agents.

1 citationsRead paper

When Context Gets Root: Privilege Escalation in LLM Harnesses

Aug 27, 2026

研究解决了LLM中由于上下文构建导致的指令特权提升问题，通过多代理机制在六个编码代理框架上实现了13个攻击目标。

0 citationsRead paper

URNet: A Unified Reparameterized Network for Efficient RGB-D Semantic Segmentation

Aug 06, 2026

This work addresses the limitations of existing RGB-D semantic segmentation methods, which typically employ dual-encoder architectures that suffer from inadequate depth representation, restricted cross-modal interaction, and computational redundancy. To overcome these issues, we propose URNet, a unified framework that processes both RGB and depth inputs through a single encoder to enable efficient multi-scale feature fusion. The core innovations include integrating reparameterized blocks (RepBlocks) with Linear Gated Attention (LGA) modules, allowing simultaneous feature extraction and cross-modal interaction within a unified architecture, as well as designing a lightweight, general-purpose Pyramid Merging Decoder (PMD). Extensive experiments demonstrate that URNet achieves state-of-the-art performance across multiple benchmarks while significantly improving inference efficiency.

0 citationsRead paper

IACM-RL: Intent-Aware Context Management and Reinforcement Learning for Complex Tool Invocation under Dynamic Intent Fluctuations

Aug 03, 2026

This work addresses the challenges of intent drift and infinite API loops in long-horizon tool-augmented interactions, where dynamic user intentions frequently shift. To tackle these issues, the authors propose the IACM-RL framework, which leverages a newly constructed DynamicIntent dataset and introduces a belief-state-based context manager equipped with a structured staleness tagging mechanism to identify outdated constraints. The framework further incorporates hierarchical intent-aware rewards, multiple auxiliary losses, and reinforcement learning to enable autonomous, intent-sensitive context updates. Experimental results demonstrate that IACM-RL significantly outperforms existing baselines on DynamicIntent, BFCL-V3, and 𝜏²-Bench, effectively mitigating infinite loops and stale context problems while enhancing out-of-domain generalization.

0 citationsRead paper

TurboClear: One-Step Object-Effect Removal via Region-Calibrated Distribution Matching and Fusion

Aug 02, 2026

This work addresses the inefficiency of existing diffusion models in object and effect removal tasks, which typically rely on multi-step denoising, as well as the inability of single-step distillation methods to preserve the asymmetric behavior between edited and retained regions. To overcome these limitations, the authors propose TurboClear—a single-step removal model based on SDXL—that introduces Region-aware Distribution Matching (RDM) to enable region-sensitive distillation, effectively retaining the teacher model’s asymmetric editing characteristics. Additionally, a lightweight Learnable Spatial Fusion (LSF) mechanism is incorporated to enhance inference efficiency. Experimental results demonstrate that TurboClear achieves 40.04× and 665× speedups over ObjectClear and OmniPaint, respectively, while delivering comparable or superior visual removal quality.

0 citationsRead paper

Recent publications

Latest Papers

When Context Gets Root: Privilege Escalation in LLM Harnesses

Aug 27, 2026

研究解决了LLM中由于上下文构建导致的指令特权提升问题，通过多代理机制在六个编码代理框架上实现了13个攻击目标。

0 citationsRead paper

URNet: A Unified Reparameterized Network for Efficient RGB-D Semantic Segmentation

Aug 06, 2026

This work addresses the limitations of existing RGB-D semantic segmentation methods, which typically employ dual-encoder architectures that suffer from inadequate depth representation, restricted cross-modal interaction, and computational redundancy. To overcome these issues, we propose URNet, a unified framework that processes both RGB and depth inputs through a single encoder to enable efficient multi-scale feature fusion. The core innovations include integrating reparameterized blocks (RepBlocks) with Linear Gated Attention (LGA) modules, allowing simultaneous feature extraction and cross-modal interaction within a unified architecture, as well as designing a lightweight, general-purpose Pyramid Merging Decoder (PMD). Extensive experiments demonstrate that URNet achieves state-of-the-art performance across multiple benchmarks while significantly improving inference efficiency.

0 citationsRead paper

IACM-RL: Intent-Aware Context Management and Reinforcement Learning for Complex Tool Invocation under Dynamic Intent Fluctuations

Aug 03, 2026

This work addresses the challenges of intent drift and infinite API loops in long-horizon tool-augmented interactions, where dynamic user intentions frequently shift. To tackle these issues, the authors propose the IACM-RL framework, which leverages a newly constructed DynamicIntent dataset and introduces a belief-state-based context manager equipped with a structured staleness tagging mechanism to identify outdated constraints. The framework further incorporates hierarchical intent-aware rewards, multiple auxiliary losses, and reinforcement learning to enable autonomous, intent-sensitive context updates. Experimental results demonstrate that IACM-RL significantly outperforms existing baselines on DynamicIntent, BFCL-V3, and 𝜏²-Bench, effectively mitigating infinite loops and stale context problems while enhancing out-of-domain generalization.

0 citationsRead paper

TurboClear: One-Step Object-Effect Removal via Region-Calibrated Distribution Matching and Fusion

Aug 02, 2026

This work addresses the inefficiency of existing diffusion models in object and effect removal tasks, which typically rely on multi-step denoising, as well as the inability of single-step distillation methods to preserve the asymmetric behavior between edited and retained regions. To overcome these limitations, the authors propose TurboClear—a single-step removal model based on SDXL—that introduces Region-aware Distribution Matching (RDM) to enable region-sensitive distillation, effectively retaining the teacher model’s asymmetric editing characteristics. Additionally, a lightweight Learnable Spatial Fusion (LSF) mechanism is incorporated to enhance inference efficiency. Experimental results demonstrate that TurboClear achieves 40.04× and 665× speedups over ObjectClear and OmniPaint, respectively, while delivering comparable or superior visual removal quality.

0 citationsRead paper

TriShield: Zero-Utility-Loss Defense Against Privacy Backdoors in Federated Language Model Fine-Tuning via Orthogonal Gradient Projection and Optimizer State Entanglement

Jul 30, 2026

This work addresses the vulnerability of federated fine-tuning of language models to NeuroImprint attacks, wherein a malicious parameter server implants privacy backdoors to reconstruct client data—a threat against which existing defenses either fail or severely degrade model utility. To overcome this limitation, we propose TriShield, a three-layer deterministic defense mechanism that achieves, for the first time, privacy-backdoor resistance with zero utility loss. TriShield detects parameter artifacts via memory neuron signatures, constructs virtual iterations through entanglement of Adam/AdamW optimizer states, and enforces semantic-subspace orthogonality via SVD-based gradient projection, theoretically guaranteeing zero mutual information between uploaded gradients and individual training samples. Evaluated on GPT-2 and Llama-Guard-3-1B, our method reduces attack reconstruction success to 0% while preserving or even improving model accuracy, with less than 5% additional GPU overhead.

0 citationsRead paper