Selected work

Representative Papers

ODYSSEY: Reestablishing Confidentiality in Confidential Blockchain via Delegated Execution

Jun 03, 2026

This work addresses execution inference and replay attacks in consortium blockchains arising from side-channel leaks and state continuity within Trusted Execution Environments (TEEs). To mitigate these threats, the authors propose a delegated execution model wherein clients entrust transaction execution to a designated node while other nodes merely synchronize the resulting outputs. This architecture integrates location-aware concurrent execution with a robust delegation failure handling mechanism, significantly reducing the attack surface and preserving transaction confidentiality without compromising performance. A prototype implementation built on FISCO BCOS achieves approximately 4,000 transactions per second (TPS) throughput with latencies as low as 0.4–0.5 seconds in a three-node wide-area network deployment.

0 citationsRead paper

Evolvable Embodied Agent for Robotic Manipulation via Long Short-Term Reflection and Optimization

Apr 15, 2026

Traditional robotic approaches are limited in generalization, training efficiency, and interpretability, hindering their ability to continuously self-adapt through environmental feedback. This work proposes the Evolvable Embodied Agent (EEAgent) framework, which integrates large vision-language models (VLMs) for environmental perception and task planning, and introduces a Long- and Short-Term Reflection Optimization (LSTRO) mechanism. LSTRO dynamically fuses historical experiences with newly acquired knowledge to iteratively refine prompting strategies, enabling continual self-evolution of the agent. Evaluated on six tasks in the VIMA-Bench benchmark, the proposed method achieves a new state-of-the-art performance and significantly outperforms existing baselines in complex scenarios, demonstrating its effectiveness and advancement in enabling self-evolving embodied intelligence.

0 citationsRead paper

SACRED: A Faithful Annotated Multimedia Multimodal Multilingual Dataset for Classifying Connectedness Types in Online Spirituality

Mar 28, 2026

This study addresses the lack of high-quality, publicly available multimodal datasets for classifying connection types in online spiritual contexts. To bridge this gap, the authors collaborated with social scientists to construct SACRED—the first multilingual, multimedia, multimodal annotated dataset dedicated to this domain—and identified a novel connection type that extends communication theory. The annotation process employed an expert-in-the-loop framework, and the dataset was systematically evaluated using rule-based methods, fine-tuned models, and 13 state-of-the-art large language models. Experimental results demonstrate the effectiveness of both the dataset and the proposed approach: DeepSeek-V3 achieved 79.19% accuracy on the Quora text test set, while GPT-4o-mini attained an F1 score of 63.99% on visual tasks.

0 citationsRead paper

Anchoring Emotions in Text: Robust Multimodal Fusion for Mimicry Intensity Estimation

Mar 16, 2026

This work addresses the challenge of estimating Emotional Mimicry Intensity (EMI) in naturalistic settings, where noise and missing modalities severely degrade performance. To this end, we propose TAEMI, a novel framework that leverages text transcriptions as semantic anchors to align and denoise noisy visual and auditory signals via a text-anchored dual cross-attention mechanism. TAEMI further enhances robustness through learnable missing-modality tokens and a modality dropout strategy during training. Departing from conventional symmetric fusion paradigms, our approach uniquely positions textual information as the central guide for multimodal temporal modeling. Evaluated on the Hume-VidMimic2 dataset, TAEMI achieves state-of-the-art performance across six continuous emotion dimensions, significantly outperforming existing baselines with a higher average Pearson correlation coefficient.

0 citationsRead paper

Astra: Activation-Space Tail-Eigenvector Low-Rank Adaptation of Large Language Models

Feb 22, 2026

This work proposes Astra, a novel parameter-efficient fine-tuning (PEFT) method that addresses a key limitation of existing low-rank adaptation approaches such as LoRA: their neglect of the subspace spanned by tail eigenvectors in the activation space, which constrains fine-tuning performance. Astra is the first to systematically leverage a task-specific calibration set to estimate these tail eigenvectors of output activations, constructing a task-adaptive low-rank adapter whose parameter updates are constrained within this informative subspace. By identifying critical activation directions via eigendecomposition, Astra achieves significantly faster convergence and improved model performance with only a marginal increase in trainable parameters. Evaluated across 16 NLU and NLG benchmarks, Astra consistently outperforms state-of-the-art PEFT methods and, in several tasks, even surpasses full fine-tuning.

0 citationsRead paper

Recent publications

Latest Papers

ODYSSEY: Reestablishing Confidentiality in Confidential Blockchain via Delegated Execution

Jun 03, 2026

This work addresses execution inference and replay attacks in consortium blockchains arising from side-channel leaks and state continuity within Trusted Execution Environments (TEEs). To mitigate these threats, the authors propose a delegated execution model wherein clients entrust transaction execution to a designated node while other nodes merely synchronize the resulting outputs. This architecture integrates location-aware concurrent execution with a robust delegation failure handling mechanism, significantly reducing the attack surface and preserving transaction confidentiality without compromising performance. A prototype implementation built on FISCO BCOS achieves approximately 4,000 transactions per second (TPS) throughput with latencies as low as 0.4–0.5 seconds in a three-node wide-area network deployment.

0 citationsRead paper

Evolvable Embodied Agent for Robotic Manipulation via Long Short-Term Reflection and Optimization

Apr 15, 2026

Traditional robotic approaches are limited in generalization, training efficiency, and interpretability, hindering their ability to continuously self-adapt through environmental feedback. This work proposes the Evolvable Embodied Agent (EEAgent) framework, which integrates large vision-language models (VLMs) for environmental perception and task planning, and introduces a Long- and Short-Term Reflection Optimization (LSTRO) mechanism. LSTRO dynamically fuses historical experiences with newly acquired knowledge to iteratively refine prompting strategies, enabling continual self-evolution of the agent. Evaluated on six tasks in the VIMA-Bench benchmark, the proposed method achieves a new state-of-the-art performance and significantly outperforms existing baselines in complex scenarios, demonstrating its effectiveness and advancement in enabling self-evolving embodied intelligence.

0 citationsRead paper

SACRED: A Faithful Annotated Multimedia Multimodal Multilingual Dataset for Classifying Connectedness Types in Online Spirituality

Mar 28, 2026

This study addresses the lack of high-quality, publicly available multimodal datasets for classifying connection types in online spiritual contexts. To bridge this gap, the authors collaborated with social scientists to construct SACRED—the first multilingual, multimedia, multimodal annotated dataset dedicated to this domain—and identified a novel connection type that extends communication theory. The annotation process employed an expert-in-the-loop framework, and the dataset was systematically evaluated using rule-based methods, fine-tuned models, and 13 state-of-the-art large language models. Experimental results demonstrate the effectiveness of both the dataset and the proposed approach: DeepSeek-V3 achieved 79.19% accuracy on the Quora text test set, while GPT-4o-mini attained an F1 score of 63.99% on visual tasks.

0 citationsRead paper

Anchoring Emotions in Text: Robust Multimodal Fusion for Mimicry Intensity Estimation

Mar 16, 2026

This work addresses the challenge of estimating Emotional Mimicry Intensity (EMI) in naturalistic settings, where noise and missing modalities severely degrade performance. To this end, we propose TAEMI, a novel framework that leverages text transcriptions as semantic anchors to align and denoise noisy visual and auditory signals via a text-anchored dual cross-attention mechanism. TAEMI further enhances robustness through learnable missing-modality tokens and a modality dropout strategy during training. Departing from conventional symmetric fusion paradigms, our approach uniquely positions textual information as the central guide for multimodal temporal modeling. Evaluated on the Hume-VidMimic2 dataset, TAEMI achieves state-of-the-art performance across six continuous emotion dimensions, significantly outperforming existing baselines with a higher average Pearson correlation coefficient.

0 citationsRead paper

Astra: Activation-Space Tail-Eigenvector Low-Rank Adaptation of Large Language Models

Feb 22, 2026

This work proposes Astra, a novel parameter-efficient fine-tuning (PEFT) method that addresses a key limitation of existing low-rank adaptation approaches such as LoRA: their neglect of the subspace spanned by tail eigenvectors in the activation space, which constrains fine-tuning performance. Astra is the first to systematically leverage a task-specific calibration set to estimate these tail eigenvectors of output activations, constructing a task-adaptive low-rank adapter whose parameter updates are constrained within this informative subspace. By identifying critical activation directions via eigendecomposition, Astra achieves significantly faster convergence and improved model performance with only a marginal increase in trainable parameters. Evaluated across 16 NLU and NLG benchmarks, Astra consistently outperforms state-of-the-art PEFT methods and, in several tasks, even surpasses full fine-tuning.

0 citationsRead paper