Selected work

Representative Papers

Cardiologent: Multi-Agent Clinical Decision Support for Patient-Level Arrhythmia Assessment, Urgency, and Management

Jul 28, 2026

This study addresses the limitation of existing arrhythmia analysis methods, which typically identify only rhythm types without providing personalized clinical decision support. The authors propose the first end-to-end multi-agent system that integrates single-lead electrocardiogram and photoplethysmography signals to construct patient-level rhythm profiles. By incorporating clinical data and retrieval-augmented guideline reasoning, the system generates auditable diagnostic and therapeutic recommendations. It uniquely unifies rhythm assessment, clinical significance interpretation, and urgency-based management advice within a collaborative multi-agent framework, further validated through expert alignment mechanisms. The approach achieves state-of-the-art performance across tasks including comprehensive diagnosis, clinical relevance determination, urgency stratification, and management suggestion, with inter-rater agreement (ICC = 0.74 with cardiologists and 0.66 with large language models) comparable to inter-expert consensus (ICC = 0.67).

0 citationsRead paper

A Failure-Mode Benchmark for Polymorphic Sybil Poisoning in RAG

Jul 04, 2026

This work addresses the vulnerability of Retrieval-Augmented Generation (RAG) systems to polymorphic Sybil poisoning attacks—coordinated, diverse misinformation injections—and introduces the first evaluation framework tailored to this threat model. By enforcing protocol-level isolation to disentangle retrieval effects and categorizing model outputs into four failure modes (gold, hijacked, refusal, and semantic drift), the authors establish a quad-class assessment scheme along with clean-to-poisoned transition matrices. Experiments spanning five readers (7B–120B parameters), two retrievers, 3,145 questions, and 2,982 Sybil groups reveal that polymorphic attacks achieve a 22.8% hijacking rate, an 18.8-percentage-point increase over monomorphic attacks. Moreover, conventional Attack Success Rate (ASR) metrics critically overlook refusal and drift risks: models with identical ASR exhibit performance gaps exceeding 16 percentage points, exposing a significant blind spot in current RAG evaluations.

0 citationsRead paper

See Selectively, Act Adaptively: Dual-Level Structural Decomposition for Bimanual Robot Manipulation

Jun 11, 2026

This work addresses the challenges in dual-arm robotic manipulation arising from dynamic shifts in visual attention across task phases and the alternating nature of arm interactions between independent and coordinated modes, which complicate policy learning. To tackle this, the authors propose a novel two-level vision–language–action framework that jointly integrates view-selective visual routing and an interaction-aware Mixture-of-Experts (MoE) action policy. This architecture explicitly decouples visual processing from bimanual coordination structure, thereby introducing effective inductive bias. Evaluated on the RoboTwin 2.0 simulation platform, the method achieves a 27.7% average improvement in success rate in simulation and a 43.3% gain on long-horizon real-world tasks, significantly outperforming single-module baselines.

0 citationsRead paper

SlotGCG: Exploiting the Positional Vulnerability in LLMs for Jailbreak Attacks

Jun 03, 2026

This work addresses a critical limitation in existing optimization-based jailbreak attacks, such as GCG, which typically insert adversarial tokens at fixed positions and overlook the pivotal role of position selection in attack efficacy. To this end, we present SlotGCG—a plug-and-play, attack-agnostic position search mechanism that systematically quantifies the vulnerability of each prompt slot for the first time. SlotGCG introduces a Vulnerable Slot Score (VSS) to evaluate positional susceptibility and integrates it with optimization strategies like Greedy Coordinate Gradient to apply targeted perturbations at the most vulnerable locations. Extensive experiments demonstrate that SlotGCG improves attack success rates by 14% over GCG across multiple large language models, converges faster, and outperforms baselines by up to 42% under defensive settings, with only approximately 200ms of additional preprocessing overhead.

0 citationsRead paper

SurgicalMamba: Dual-Path SSD with State Regramming for Online Surgical Phase Recognition

May 14, 2026

This work addresses the challenge of online surgical phase recognition, which requires frame-by-frame prediction using only historical context and faces difficulties due to extremely long videos, non-uniform temporal dynamics, and highly correlated visual features from a narrow visual domain. To tackle these issues, the authors propose a causal model built upon Mamba2’s Structured State Space Duality (SSD), featuring a dual-path SSD architecture that disentangles short- and long-term temporal dependencies. They further introduce an intensity-modulated temporal reparameterization mechanism to adaptively adjust to varying information density and incorporate an unsupervised state reprogramming module based on the Cayley transform to achieve cross-channel phase alignment. The method achieves state-of-the-art performance across seven public surgical phase recognition benchmarks, attaining online accuracy and Jaccard scores of 94.6%/82.7% on Cholec80 and 89.5%/68.9% on AutoLaparo, with a real-time inference speed of 119 fps on a single GPU.

0 citationsRead paper

Recent publications

Latest Papers

Cardiologent: Multi-Agent Clinical Decision Support for Patient-Level Arrhythmia Assessment, Urgency, and Management

Jul 28, 2026

This study addresses the limitation of existing arrhythmia analysis methods, which typically identify only rhythm types without providing personalized clinical decision support. The authors propose the first end-to-end multi-agent system that integrates single-lead electrocardiogram and photoplethysmography signals to construct patient-level rhythm profiles. By incorporating clinical data and retrieval-augmented guideline reasoning, the system generates auditable diagnostic and therapeutic recommendations. It uniquely unifies rhythm assessment, clinical significance interpretation, and urgency-based management advice within a collaborative multi-agent framework, further validated through expert alignment mechanisms. The approach achieves state-of-the-art performance across tasks including comprehensive diagnosis, clinical relevance determination, urgency stratification, and management suggestion, with inter-rater agreement (ICC = 0.74 with cardiologists and 0.66 with large language models) comparable to inter-expert consensus (ICC = 0.67).

0 citationsRead paper

A Failure-Mode Benchmark for Polymorphic Sybil Poisoning in RAG

Jul 04, 2026

This work addresses the vulnerability of Retrieval-Augmented Generation (RAG) systems to polymorphic Sybil poisoning attacks—coordinated, diverse misinformation injections—and introduces the first evaluation framework tailored to this threat model. By enforcing protocol-level isolation to disentangle retrieval effects and categorizing model outputs into four failure modes (gold, hijacked, refusal, and semantic drift), the authors establish a quad-class assessment scheme along with clean-to-poisoned transition matrices. Experiments spanning five readers (7B–120B parameters), two retrievers, 3,145 questions, and 2,982 Sybil groups reveal that polymorphic attacks achieve a 22.8% hijacking rate, an 18.8-percentage-point increase over monomorphic attacks. Moreover, conventional Attack Success Rate (ASR) metrics critically overlook refusal and drift risks: models with identical ASR exhibit performance gaps exceeding 16 percentage points, exposing a significant blind spot in current RAG evaluations.

0 citationsRead paper

See Selectively, Act Adaptively: Dual-Level Structural Decomposition for Bimanual Robot Manipulation

Jun 11, 2026

This work addresses the challenges in dual-arm robotic manipulation arising from dynamic shifts in visual attention across task phases and the alternating nature of arm interactions between independent and coordinated modes, which complicate policy learning. To tackle this, the authors propose a novel two-level vision–language–action framework that jointly integrates view-selective visual routing and an interaction-aware Mixture-of-Experts (MoE) action policy. This architecture explicitly decouples visual processing from bimanual coordination structure, thereby introducing effective inductive bias. Evaluated on the RoboTwin 2.0 simulation platform, the method achieves a 27.7% average improvement in success rate in simulation and a 43.3% gain on long-horizon real-world tasks, significantly outperforming single-module baselines.

0 citationsRead paper

SlotGCG: Exploiting the Positional Vulnerability in LLMs for Jailbreak Attacks

Jun 03, 2026

This work addresses a critical limitation in existing optimization-based jailbreak attacks, such as GCG, which typically insert adversarial tokens at fixed positions and overlook the pivotal role of position selection in attack efficacy. To this end, we present SlotGCG—a plug-and-play, attack-agnostic position search mechanism that systematically quantifies the vulnerability of each prompt slot for the first time. SlotGCG introduces a Vulnerable Slot Score (VSS) to evaluate positional susceptibility and integrates it with optimization strategies like Greedy Coordinate Gradient to apply targeted perturbations at the most vulnerable locations. Extensive experiments demonstrate that SlotGCG improves attack success rates by 14% over GCG across multiple large language models, converges faster, and outperforms baselines by up to 42% under defensive settings, with only approximately 200ms of additional preprocessing overhead.

0 citationsRead paper

SurgicalMamba: Dual-Path SSD with State Regramming for Online Surgical Phase Recognition

May 14, 2026

This work addresses the challenge of online surgical phase recognition, which requires frame-by-frame prediction using only historical context and faces difficulties due to extremely long videos, non-uniform temporal dynamics, and highly correlated visual features from a narrow visual domain. To tackle these issues, the authors propose a causal model built upon Mamba2’s Structured State Space Duality (SSD), featuring a dual-path SSD architecture that disentangles short- and long-term temporal dependencies. They further introduce an intensity-modulated temporal reparameterization mechanism to adaptively adjust to varying information density and incorporate an unsupervised state reprogramming module based on the Cayley transform to achieve cross-channel phase alignment. The method achieves state-of-the-art performance across seven public surgical phase recognition benchmarks, attaining online accuracy and Jaccard scores of 94.6%/82.7% on Cholec80 and 89.5%/68.9% on AutoLaparo, with a real-time inference speed of 119 fps on a single GPU.

0 citationsRead paper