Institution profile

Honor Device Co., Ltd

Industry researchasia · cn
Official website
Research library50linked papers
Opportunities0open roles
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

Mingbird: A Local-First Agent Harness Enabling Small Open Models to Complete Real Tasks

Oct 01, 2026

This study addresses the persistent failure of small open-source models in local agent frameworks, where they frequently struggle with real-world tasks due to context overflow, divergent self-correction, and tool-calling loops. To overcome these limitations, this work proposes a local-first agent framework targeting Windows and Ollama environments. It introduces ten compensatory mechanisms—including byte-level net-zero prefill budgeting, task-rereading completion gating, and signature-level loop detection—integrated with refined context management and deterministic artifact scoring to systematically mitigate small-model failure modes. Evaluated on the LRAB benchmark, the proposed framework achieves a score of 0.886, significantly outperforming mainstream alternatives. Furthermore, it attains 0.856 on τ²-bench, demonstrating its effectiveness in substantially enhancing the reliability and task completion rates of small language models executing real-world tasks.

0 citationsRead paper

Spackle: Completing Large View Single Image NVS with Adaptive Gaussians

Sep 25, 2026

This study addresses the capacity competition between visible and occluded regions in single-image novel view synthesis, which arises from using a fixed number of Gaussians. To mitigate this issue, we propose a lightweight residual learning framework based on 3D Gaussian Splatting. The method operates in three stages: predicting baseline attributes, identifying reconstruction deficiencies to optimize residual Gaussians, and merging both components during inference for efficient rendering. By introducing an adaptive residual mechanism, our approach effectively alleviates capacity competition without increasing computational overhead. It significantly enhances scene fidelity under large viewpoint deviations, achieving state-of-the-art performance.

0 citationsRead paper

ClearGS: Reliability-Aware Gaussian Splatting from Handheld Videos

Sep 25, 2026

This study addresses the challenges of 3D Gaussian Splatting reconstruction from handheld videos, where uneven view coverage and mixed image quality degrade performance. To overcome these issues, this work proposes a reliability-aware view allocation framework. Departing from conventional binary frame selection, it introduces hierarchical weight-based supervision. Furthermore, by integrating rendering-guided reference-free image restoration with a full-trajectory integration mechanism, the method recovers fine details while preserving complete trajectory coverage in the absence of clear references. Experimental results demonstrate that the proposed approach achieves state-of-the-art performance on both the GS2E and GSOTM datasets. It significantly improves CLIP-IQA and MUSIQ scores while reducing LPIPS error, thereby enabling high-quality 3D reconstruction.

0 citationsRead paper

Outcome-Sensitive Motion Search for Impact-Aware Dexterous Catching

Sep 24, 2026

This study addresses the challenges in reinforcement learning for dexterous grasping, where coordinating interception with impact mitigation is difficult and teacher demonstrations are often imperfect. We propose an outcome-sensitive motion search method for constructing demonstration manifolds. By introducing the concept of an outcome-sensitive window, our approach integrates local geodesic search to optimize successful trajectories and rectify failure cases. Furthermore, a calibrated action error model is employed to validate action quality, thereby overcoming demonstration data bottlenecks. Experimental results demonstrate that the proposed method effectively mitigates the limitations of teacher policies, enabling the resulting policy to surpass privileged reinforcement learning teachers in both grasping success rate and impact mitigation performance.

0 citationsRead paper
Recent publications

Latest Papers

Mingbird: A Local-First Agent Harness Enabling Small Open Models to Complete Real Tasks

Oct 01, 2026

This study addresses the persistent failure of small open-source models in local agent frameworks, where they frequently struggle with real-world tasks due to context overflow, divergent self-correction, and tool-calling loops. To overcome these limitations, this work proposes a local-first agent framework targeting Windows and Ollama environments. It introduces ten compensatory mechanisms—including byte-level net-zero prefill budgeting, task-rereading completion gating, and signature-level loop detection—integrated with refined context management and deterministic artifact scoring to systematically mitigate small-model failure modes. Evaluated on the LRAB benchmark, the proposed framework achieves a score of 0.886, significantly outperforming mainstream alternatives. Furthermore, it attains 0.856 on τ²-bench, demonstrating its effectiveness in substantially enhancing the reliability and task completion rates of small language models executing real-world tasks.

0 citationsRead paper

Spackle: Completing Large View Single Image NVS with Adaptive Gaussians

Sep 25, 2026

This study addresses the capacity competition between visible and occluded regions in single-image novel view synthesis, which arises from using a fixed number of Gaussians. To mitigate this issue, we propose a lightweight residual learning framework based on 3D Gaussian Splatting. The method operates in three stages: predicting baseline attributes, identifying reconstruction deficiencies to optimize residual Gaussians, and merging both components during inference for efficient rendering. By introducing an adaptive residual mechanism, our approach effectively alleviates capacity competition without increasing computational overhead. It significantly enhances scene fidelity under large viewpoint deviations, achieving state-of-the-art performance.

0 citationsRead paper

ClearGS: Reliability-Aware Gaussian Splatting from Handheld Videos

Sep 25, 2026

This study addresses the challenges of 3D Gaussian Splatting reconstruction from handheld videos, where uneven view coverage and mixed image quality degrade performance. To overcome these issues, this work proposes a reliability-aware view allocation framework. Departing from conventional binary frame selection, it introduces hierarchical weight-based supervision. Furthermore, by integrating rendering-guided reference-free image restoration with a full-trajectory integration mechanism, the method recovers fine details while preserving complete trajectory coverage in the absence of clear references. Experimental results demonstrate that the proposed approach achieves state-of-the-art performance on both the GS2E and GSOTM datasets. It significantly improves CLIP-IQA and MUSIQ scores while reducing LPIPS error, thereby enabling high-quality 3D reconstruction.

0 citationsRead paper

Outcome-Sensitive Motion Search for Impact-Aware Dexterous Catching

Sep 24, 2026

This study addresses the challenges in reinforcement learning for dexterous grasping, where coordinating interception with impact mitigation is difficult and teacher demonstrations are often imperfect. We propose an outcome-sensitive motion search method for constructing demonstration manifolds. By introducing the concept of an outcome-sensitive window, our approach integrates local geodesic search to optimize successful trajectories and rectify failure cases. Furthermore, a calibrated action error model is employed to validate action quality, thereby overcoming demonstration data bottlenecks. Experimental results demonstrate that the proposed method effectively mitigates the limitations of teacher policies, enabling the resulting policy to surpass privileged reinforcement learning teachers in both grasping success rate and impact mitigation performance.

0 citationsRead paper