Selected work

Representative Papers

From LLMs to LRMs: Rethinking Pruning for Reasoning-Centric Models

Jan 26, 2026

This work addresses the limitations of existing model pruning methods, which are primarily designed for instruction-following large language models and struggle to effectively adapt to reasoning-augmented models that explicitly generate long reasoning chains. The study systematically investigates pruning strategies tailored for both model types, proposing a calibration and recovery pipeline aligned with the original training distribution. It evaluates static depth pruning, static width pruning, and dynamic pruning across 17 tasks, revealing—for the first time—the dependence of pruning efficacy on the underlying reasoning paradigm. The findings indicate that depth pruning is better suited for classification tasks, while width pruning demonstrates greater robustness in generative and reasoning tasks. Moreover, static pruning better preserves reasoning capabilities, whereas dynamic pruning remains challenging for long-chain reasoning scenarios.

1 citationsRead paper

Educational Short Videos: Bibliometric Trends, Thematic Structure, and Operationalisation

Aug 07, 2026

This study addresses the lack of comparability and integration in existing research on educational short videos, which stems from overly generalized labeling that leads to inconsistencies in functions, activities, contexts, and evaluation criteria. To systematically map the field’s developmental trajectory and thematic structure, the authors combine bibliometric analysis, non-negative matrix factorization (NMF) topic modeling, and structured content analysis. Innovatively redefining “short videos” by educational purpose rather than duration alone, the study identifies 16 top-level themes and their evolutionary relationships. Among these, Skill Development emerges as the core theme, while Video-Based Health Interventions and related areas represent high-impact directions. The analysis further reveals that learning outcomes primarily cluster around two dimensions: knowledge/achievement and engagement/motivation.

0 citationsRead paper

InstanceSplat: Instance-Aware Feed-Forward 3D Gaussian Splatting for Scene Understanding

Aug 07, 2026

This work addresses the limitations of existing feed-forward 3D Gaussian splatting methods, which lack instance awareness, and current instance-aware approaches that rely on per-scene optimization and decouple reconstruction from semantics, hindering joint learning. To overcome these issues, we propose InstanceSplat, the first end-to-end, unified feed-forward 3D Gaussian splatting framework that jointly achieves 3D reconstruction, instance segmentation, and open-vocabulary semantic understanding from pose-free multi-view images in a single forward pass. The key innovation lies in leveraging shared 3D Gaussians to establish cross-view consistent instance features and introducing an instance-guided mechanism for reconstruction and semantic aggregation. InstanceSplat achieves state-of-the-art performance across novel view synthesis, instance segmentation, and semantic understanding, while offering efficient inference, strong generalization, and robustness to unseen scenes and input configurations.

0 citationsRead paper

Towards the Harness of Embodied Agents

Aug 03, 2026

This work addresses the challenge of effectively extending the successful “harness” paradigm from code-based agents to embodied agents in the physical world, where environmental states are not directly observable and action outcomes are difficult to automatically evaluate. To overcome these limitations, the authors propose Thea, a framework that encapsulates robotic capabilities as modular tools, constructs a symbolic scene graph to serve as contextual representation, and introduces an exit-code–based mechanism for action evaluation and diagnosis, thereby establishing a closed-loop control architecture for embodied intelligence. Thea represents the first successful adaptation of the harness paradigm to real-world physical environments, demonstrating autonomous execution of complex, long-horizon tasks and validating the approach’s feasibility and effectiveness. Furthermore, the compositional use of tools enables emergent, rich behavioral repertoires.

0 citationsRead paper

WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning

Jul 30, 2026

Existing large language model pruning methods struggle to balance accuracy and inference efficiency under high sparsity: static pruning lacks input adaptability, while dynamic pruning suffers from coarse granularity and limited acceleration. To address this, this work proposes WIDE, the first framework to advance dynamic pruning to the neuron block level, enabling end-to-end differentiable, token-wise dynamic width pruning during both prefill and decoding stages. WIDE adaptively selects groups of attention heads and feed-forward network channels for each token. Through a co-design of pruning and kernels, along with a two-stage training pipeline and hardware-agnostic block-skipping techniques—both inter-block and intra-block—WIDE achieves a 55.1% speedup over the state-of-the-art dynamic depth pruning at 50% sparsity, delivering kernel-level speedups of 1.98× and 4.95× in prefill and decoding phases, respectively, and an overall end-to-end acceleration of 1.68×–1.55×.

0 citationsRead paper

Recent publications

Latest Papers

Educational Short Videos: Bibliometric Trends, Thematic Structure, and Operationalisation

Aug 07, 2026

This study addresses the lack of comparability and integration in existing research on educational short videos, which stems from overly generalized labeling that leads to inconsistencies in functions, activities, contexts, and evaluation criteria. To systematically map the field’s developmental trajectory and thematic structure, the authors combine bibliometric analysis, non-negative matrix factorization (NMF) topic modeling, and structured content analysis. Innovatively redefining “short videos” by educational purpose rather than duration alone, the study identifies 16 top-level themes and their evolutionary relationships. Among these, Skill Development emerges as the core theme, while Video-Based Health Interventions and related areas represent high-impact directions. The analysis further reveals that learning outcomes primarily cluster around two dimensions: knowledge/achievement and engagement/motivation.

0 citationsRead paper

InstanceSplat: Instance-Aware Feed-Forward 3D Gaussian Splatting for Scene Understanding

Aug 07, 2026

This work addresses the limitations of existing feed-forward 3D Gaussian splatting methods, which lack instance awareness, and current instance-aware approaches that rely on per-scene optimization and decouple reconstruction from semantics, hindering joint learning. To overcome these issues, we propose InstanceSplat, the first end-to-end, unified feed-forward 3D Gaussian splatting framework that jointly achieves 3D reconstruction, instance segmentation, and open-vocabulary semantic understanding from pose-free multi-view images in a single forward pass. The key innovation lies in leveraging shared 3D Gaussians to establish cross-view consistent instance features and introducing an instance-guided mechanism for reconstruction and semantic aggregation. InstanceSplat achieves state-of-the-art performance across novel view synthesis, instance segmentation, and semantic understanding, while offering efficient inference, strong generalization, and robustness to unseen scenes and input configurations.

0 citationsRead paper

Towards the Harness of Embodied Agents

Aug 03, 2026

This work addresses the challenge of effectively extending the successful “harness” paradigm from code-based agents to embodied agents in the physical world, where environmental states are not directly observable and action outcomes are difficult to automatically evaluate. To overcome these limitations, the authors propose Thea, a framework that encapsulates robotic capabilities as modular tools, constructs a symbolic scene graph to serve as contextual representation, and introduces an exit-code–based mechanism for action evaluation and diagnosis, thereby establishing a closed-loop control architecture for embodied intelligence. Thea represents the first successful adaptation of the harness paradigm to real-world physical environments, demonstrating autonomous execution of complex, long-horizon tasks and validating the approach’s feasibility and effectiveness. Furthermore, the compositional use of tools enables emergent, rich behavioral repertoires.

0 citationsRead paper

WIDE: Boosting Adaptive LLM Inference via Token-level Dynamic Width Pruning

Jul 30, 2026

Existing large language model pruning methods struggle to balance accuracy and inference efficiency under high sparsity: static pruning lacks input adaptability, while dynamic pruning suffers from coarse granularity and limited acceleration. To address this, this work proposes WIDE, the first framework to advance dynamic pruning to the neuron block level, enabling end-to-end differentiable, token-wise dynamic width pruning during both prefill and decoding stages. WIDE adaptively selects groups of attention heads and feed-forward network channels for each token. Through a co-design of pruning and kernels, along with a two-stage training pipeline and hardware-agnostic block-skipping techniques—both inter-block and intra-block—WIDE achieves a 55.1% speedup over the state-of-the-art dynamic depth pruning at 50% sparsity, delivering kernel-level speedups of 1.98× and 4.95× in prefill and decoding phases, respectively, and an overall end-to-end acceleration of 1.68×–1.55×.

0 citationsRead paper

MMOOC: A Comprehensive Benchmark for Out-of-Context Evaluation in Multimodal Large Language Models

Jul 29, 2026

This work addresses the challenge that multimodal large language models (MLLMs) struggle to distinguish between out-of-context questions—requiring rejection—and in-context shifted questions—amenable to answering—under contextual shift scenarios, a gap inadequately covered by existing benchmarks. To this end, we introduce MMOOC, a large-scale evaluation benchmark that systematically encompasses eight types of contextual shifts, three question formats, and diverse visual settings. Rigorous multi-stage data cleaning and human verification ensure high data quality. We further propose an LLM-as-a-Judge-based metric to assess reasoning correctness. Experimental results reveal a significant imbalance in current models’ ability to appropriately reject or answer such questions, while post-training strategies demonstrate notable improvements in robustness. The benchmark will be publicly released.

0 citationsRead paper