Selected work

Representative Papers

Q-Steer: Action-Value Guidance for Molecular Policy Optimization

Jul 28, 2026

This work addresses the myopic optimization problem in molecular design, where sparse rewards are only available upon complete molecule generation, making it difficult for policy-based methods to evaluate the contribution of intermediate actions. To overcome this, the authors propose Q-Steer, a general framework that guides token selection during rollout by leveraging an offline-trained and frozen Prefix-Action Value Scorer (PAVS-Q). This scorer estimates the expected future reward for each candidate SMILES token given the current prefix, and the normalized value advantage is incorporated into the sampling logits to steer generation. Without modifying the online optimizer or increasing the oracle query budget, Q-Steer consistently improves performance across diverse policy and language model backbones. On the PMO23 benchmark with a fixed budget of 10,000 oracle calls, it achieves gains in mean valid unique score ranging from +0.033 to +0.049 across all eight backbone–optimizer combinations, outperforming baselines in 18–20 out of 20 tasks per setting.

0 citationsRead paper

IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer

Jul 21, 2026

This work addresses the challenge of achieving temporally consistent modeling of both geometric structure and object identity in dynamic scenes with existing streaming methods. The authors propose a streaming 4D Instance-Anchored Geometry Transformer that processes video frames online through causal spatiotemporal modeling, incrementally updating a unified representation of camera poses, geometry, and instance identities. Key contributions include the first unified online learning framework for geometry and instance semantics over long-term dynamic sequences, the construction of InsScene4D-147K—a large-scale temporally consistent 4D dataset—and the design of a causal spatiotemporal Transformer, an incremental geometry-instance joint representation, and a geometry-guided automatic annotation pipeline. The method significantly outperforms current streaming approaches in 3D reconstruction, pose estimation, instance tracking, and open-vocabulary segmentation, enabling scalable long-term online inference.

0 citationsRead paper

Towards Accurate Emotion-Attributed Video Captioning via Fine-grained Emotion-Cause Pair Extraction

Jun 07, 2026

Existing methods for affective video captioning rely on holistic visual features, which often fail to precisely localize the specific segments responsible for eliciting emotions, resulting in redundant descriptions and inaccurate emotion attribution. To address this limitation, this work proposes a two-stage fine-grained emotion-cause pair extraction framework that introduces, for the first time, a concept-aware visual semantic decomposition module and a vision-guided explainable emotion learning mechanism. By incorporating Valence-Arousal-Dominance (VAD) vector constraints and cross-modal contrastive alignment, the model achieves precise coupling between emotions and their visual causes. Evaluated on three benchmarks including EVC-MSVD, the proposed approach outperforms state-of-the-art methods, yielding absolute improvements of 4.4% in BLEU-2 and 5.4% in ROUGE-L scores.

0 citationsRead paper

Structure-Guided Mixed Masked Pretraining and Spatial Continuity Regularization for Printed Circuit Board Defect Detection

Jun 02, 2026

This study addresses the challenge of detecting subtle, low-contrast defects embedded within dense backgrounds in printed circuit board (PCB) inspection. To this end, the authors propose a two-stage detection framework: first, a structure-guided hybrid masked sparse convolution is employed for pretraining to learn structural priors inherent to PCB layouts; second, during fine-tuning, a spatial continuity regularization is introduced to enhance the compactness and coherence of predictions for elongated defect regions. By innovatively integrating structural prior modeling with spatial constraints, the method achieves state-of-the-art performance on the DsPCBSD+ dataset, attaining 85.5% mAP₀.₅ and 52.3% mAP₀.₅:₀.₉₅, significantly outperforming existing strong baselines.

0 citationsRead paper

Partially Observable Adversarial Patch Attacks on Vision-Language-Action Models in Robotics

Jun 02, 2026

This work addresses the challenge of deploying adversarial attacks against vision-language-action (VLA) robotic systems in real-world settings, where full execution trajectories are typically unavailable. The authors propose a novel threat model based on partial observability—requiring only a prefix of the trajectory—and introduce a two-stage framework to generate fixed adversarial patches that disrupt both semantic understanding and action planning in VLA models. Key innovations include attention-map-guided patch placement and a joint perturbation strategy that simultaneously corrupts semantic grounding and amplifies trajectory curvature. To the best of the authors’ knowledge, this is the first method to achieve long-horizon adversarial attacks on VLA agents under partial observability, demonstrating significant reductions in task success rates and strong robustness across both simulated and real-world environments.

0 citationsRead paper

Recent publications

Latest Papers

Q-Steer: Action-Value Guidance for Molecular Policy Optimization

Jul 28, 2026

This work addresses the myopic optimization problem in molecular design, where sparse rewards are only available upon complete molecule generation, making it difficult for policy-based methods to evaluate the contribution of intermediate actions. To overcome this, the authors propose Q-Steer, a general framework that guides token selection during rollout by leveraging an offline-trained and frozen Prefix-Action Value Scorer (PAVS-Q). This scorer estimates the expected future reward for each candidate SMILES token given the current prefix, and the normalized value advantage is incorporated into the sampling logits to steer generation. Without modifying the online optimizer or increasing the oracle query budget, Q-Steer consistently improves performance across diverse policy and language model backbones. On the PMO23 benchmark with a fixed budget of 10,000 oracle calls, it achieves gains in mean valid unique score ranging from +0.033 to +0.049 across all eight backbone–optimizer combinations, outperforming baselines in 18–20 out of 20 tasks per setting.

0 citationsRead paper

IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer

Jul 21, 2026

This work addresses the challenge of achieving temporally consistent modeling of both geometric structure and object identity in dynamic scenes with existing streaming methods. The authors propose a streaming 4D Instance-Anchored Geometry Transformer that processes video frames online through causal spatiotemporal modeling, incrementally updating a unified representation of camera poses, geometry, and instance identities. Key contributions include the first unified online learning framework for geometry and instance semantics over long-term dynamic sequences, the construction of InsScene4D-147K—a large-scale temporally consistent 4D dataset—and the design of a causal spatiotemporal Transformer, an incremental geometry-instance joint representation, and a geometry-guided automatic annotation pipeline. The method significantly outperforms current streaming approaches in 3D reconstruction, pose estimation, instance tracking, and open-vocabulary segmentation, enabling scalable long-term online inference.

0 citationsRead paper

Towards Accurate Emotion-Attributed Video Captioning via Fine-grained Emotion-Cause Pair Extraction

Jun 07, 2026

Existing methods for affective video captioning rely on holistic visual features, which often fail to precisely localize the specific segments responsible for eliciting emotions, resulting in redundant descriptions and inaccurate emotion attribution. To address this limitation, this work proposes a two-stage fine-grained emotion-cause pair extraction framework that introduces, for the first time, a concept-aware visual semantic decomposition module and a vision-guided explainable emotion learning mechanism. By incorporating Valence-Arousal-Dominance (VAD) vector constraints and cross-modal contrastive alignment, the model achieves precise coupling between emotions and their visual causes. Evaluated on three benchmarks including EVC-MSVD, the proposed approach outperforms state-of-the-art methods, yielding absolute improvements of 4.4% in BLEU-2 and 5.4% in ROUGE-L scores.

0 citationsRead paper

Structure-Guided Mixed Masked Pretraining and Spatial Continuity Regularization for Printed Circuit Board Defect Detection

Jun 02, 2026

This study addresses the challenge of detecting subtle, low-contrast defects embedded within dense backgrounds in printed circuit board (PCB) inspection. To this end, the authors propose a two-stage detection framework: first, a structure-guided hybrid masked sparse convolution is employed for pretraining to learn structural priors inherent to PCB layouts; second, during fine-tuning, a spatial continuity regularization is introduced to enhance the compactness and coherence of predictions for elongated defect regions. By innovatively integrating structural prior modeling with spatial constraints, the method achieves state-of-the-art performance on the DsPCBSD+ dataset, attaining 85.5% mAP₀.₅ and 52.3% mAP₀.₅:₀.₉₅, significantly outperforming existing strong baselines.

0 citationsRead paper

Partially Observable Adversarial Patch Attacks on Vision-Language-Action Models in Robotics

Jun 02, 2026

This work addresses the challenge of deploying adversarial attacks against vision-language-action (VLA) robotic systems in real-world settings, where full execution trajectories are typically unavailable. The authors propose a novel threat model based on partial observability—requiring only a prefix of the trajectory—and introduce a two-stage framework to generate fixed adversarial patches that disrupt both semantic understanding and action planning in VLA models. Key innovations include attention-map-guided patch placement and a joint perturbation strategy that simultaneously corrupts semantic grounding and amplifies trajectory curvature. To the best of the authors’ knowledge, this is the first method to achieve long-horizon adversarial attacks on VLA agents under partial observability, demonstrating significant reductions in task success rates and strong robustness across both simulated and real-world environments.

0 citationsRead paper