Selected work

Representative Papers

Mr3D-VL: A generalist vision language foundation model for Multiparametric 3D Magnetic Resonance Imaging

Aug 12, 2026

Current AI models struggle to support natural language interaction, interpretability, and cross-modal spatial reasoning in multiparametric 3D MRI, limiting their performance in tasks such as glioma grading. This work proposes the first general-purpose vision–language foundation model tailored for multiparametric 3D MRI, which jointly models modality and spatial information through a shared 3D encoder, 4D rotary positional embeddings, and a multi-resolution feature injection mechanism to enable cross-scale perception. Pretrained unsupervised with 4 billion parameters, the model significantly outperforms existing general and domain-specific large models on report generation (BERTScore: 0.856), visual question answering (accuracy: 0.713), and multiple-choice tasks (accuracy: 0.912).

0 citationsRead paper

XClipGS: Exact Half-Space Clipping for Medical Volume Gaussian Splatting

Aug 07, 2026

This work addresses the challenge of accurately clipping Gaussian splats against arbitrary planes in medical volume rendering, where conventional methods often suffer from structural distortion or information leakage due to imprecise intersection handling. The authors propose a differentiable, closed-form clipping operator that requires no additional parameters, deriving an analytical solution for half-space constrained Gaussian integrals based on a local affine model. By integrating conditional Gaussian cumulative distribution functions within the EWA rendering framework, the method achieves precise clipping. Evaluated on eight CT and MRI datasets using novel metrics—Clipping Distortion Error (CDE) and Leakage (Leak)—alongside multi-distance and multi-angle supervision, the approach attains a PSNR of 33.56 dB (outperforming ClipGS), a cross-sectional SSIM of 0.860, and approximately 40× lower leakage. It consistently achieves state-of-the-art results across all four metrics under arbitrary clipping normals while maintaining a rendering speed exceeding 650 FPS.

0 citationsRead paper

See Only When Needed: Context-Aware Attention Intervention for Mitigating Hallucinations in LVLMs

Jun 29, 2026

This work addresses object hallucination in large vision-language models (LVLMs), a prevalent issue during text generation that existing training-free methods often exacerbate by over-relying on visual inputs at the expense of linguistic fluency. The authors propose Context-Aware Intervention (CAI), a training-free, inference-time mechanism guided by the principle “attend only when necessary.” CAI leverages early-layer representations to identify semantically aligned regions and dynamically triggers fine-grained attention intervention during deep decoding by jointly assessing visual relevance and prediction uncertainty—applying conservative attention shifts precisely when visual grounding degrades and entropy is high. Theoretically, CAI corresponds to the minimal-KL reweighting of attention distributions and optionally integrates contrastive decoding for further debiasing. Experiments demonstrate that CAI consistently outperforms current training-free approaches across multiple LVLM architectures and benchmarks, effectively suppressing hallucinations while preserving language fluency.

0 citationsRead paper

Mask to Concept: Auto-Promptable SAM3 via Efficient Test-Time Concept Embedding Search for Few-Shot Annotation

Jun 25, 2026

This work addresses the limitations of foundation models in medical image segmentation—namely their reliance on manual prompts, insufficient integration of fine-grained clinical knowledge, and low annotation efficiency—by proposing the M2C framework. Operating within a frozen SAM3 architecture, M2C automatically optimizes transferable visual concept embeddings from only a few annotated samples, enabling few-shot automatic annotation without additional modules or parameter fine-tuning. It further introduces a Hybrid Uncertainty Estimation (HUE) module that integrates active learning with a consistency check between concept-based and geometric prompts, establishing a human-in-the-loop self-reinforcing annotation cycle. Experiments demonstrate that M2C achieves state-of-the-art few-shot performance on medical segmentation benchmarks, substantially improving both annotation efficiency and accuracy while requiring minimal expert intervention and offering high scalability.

0 citationsRead paper

Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models

May 08, 2026

This work addresses a critical limitation in current vision-language-action (VLA) models, which rely solely on behavior cloning from successful demonstrations and thus struggle to recover from execution errors, often failing due to minor deviations. To overcome this, the authors propose an adaptive failure-aware learning framework that, for the first time, incorporates online-generated failure trajectories as adaptive negative supervision signals into end-to-end VLA training. The approach employs a dual-action generator architecture sharing a common vision-language backbone to jointly model successful and failed behaviors, coupled with a distribution-distance-driven sampling mechanism that dynamically avoids error-prone regions. Requiring no manually designed failure modes or external recovery interventions, the method achieves substantially improved policy robustness with minimal parameter overhead, outperforming existing VLA baselines in both in-domain and out-of-domain settings across short- and long-horizon manipulation tasks.

0 citationsRead paper

Recent publications

Latest Papers

Mr3D-VL: A generalist vision language foundation model for Multiparametric 3D Magnetic Resonance Imaging

Aug 12, 2026

Current AI models struggle to support natural language interaction, interpretability, and cross-modal spatial reasoning in multiparametric 3D MRI, limiting their performance in tasks such as glioma grading. This work proposes the first general-purpose vision–language foundation model tailored for multiparametric 3D MRI, which jointly models modality and spatial information through a shared 3D encoder, 4D rotary positional embeddings, and a multi-resolution feature injection mechanism to enable cross-scale perception. Pretrained unsupervised with 4 billion parameters, the model significantly outperforms existing general and domain-specific large models on report generation (BERTScore: 0.856), visual question answering (accuracy: 0.713), and multiple-choice tasks (accuracy: 0.912).

0 citationsRead paper

XClipGS: Exact Half-Space Clipping for Medical Volume Gaussian Splatting

Aug 07, 2026

This work addresses the challenge of accurately clipping Gaussian splats against arbitrary planes in medical volume rendering, where conventional methods often suffer from structural distortion or information leakage due to imprecise intersection handling. The authors propose a differentiable, closed-form clipping operator that requires no additional parameters, deriving an analytical solution for half-space constrained Gaussian integrals based on a local affine model. By integrating conditional Gaussian cumulative distribution functions within the EWA rendering framework, the method achieves precise clipping. Evaluated on eight CT and MRI datasets using novel metrics—Clipping Distortion Error (CDE) and Leakage (Leak)—alongside multi-distance and multi-angle supervision, the approach attains a PSNR of 33.56 dB (outperforming ClipGS), a cross-sectional SSIM of 0.860, and approximately 40× lower leakage. It consistently achieves state-of-the-art results across all four metrics under arbitrary clipping normals while maintaining a rendering speed exceeding 650 FPS.

0 citationsRead paper

See Only When Needed: Context-Aware Attention Intervention for Mitigating Hallucinations in LVLMs

Jun 29, 2026

This work addresses object hallucination in large vision-language models (LVLMs), a prevalent issue during text generation that existing training-free methods often exacerbate by over-relying on visual inputs at the expense of linguistic fluency. The authors propose Context-Aware Intervention (CAI), a training-free, inference-time mechanism guided by the principle “attend only when necessary.” CAI leverages early-layer representations to identify semantically aligned regions and dynamically triggers fine-grained attention intervention during deep decoding by jointly assessing visual relevance and prediction uncertainty—applying conservative attention shifts precisely when visual grounding degrades and entropy is high. Theoretically, CAI corresponds to the minimal-KL reweighting of attention distributions and optionally integrates contrastive decoding for further debiasing. Experiments demonstrate that CAI consistently outperforms current training-free approaches across multiple LVLM architectures and benchmarks, effectively suppressing hallucinations while preserving language fluency.

0 citationsRead paper

Mask to Concept: Auto-Promptable SAM3 via Efficient Test-Time Concept Embedding Search for Few-Shot Annotation

Jun 25, 2026

This work addresses the limitations of foundation models in medical image segmentation—namely their reliance on manual prompts, insufficient integration of fine-grained clinical knowledge, and low annotation efficiency—by proposing the M2C framework. Operating within a frozen SAM3 architecture, M2C automatically optimizes transferable visual concept embeddings from only a few annotated samples, enabling few-shot automatic annotation without additional modules or parameter fine-tuning. It further introduces a Hybrid Uncertainty Estimation (HUE) module that integrates active learning with a consistency check between concept-based and geometric prompts, establishing a human-in-the-loop self-reinforcing annotation cycle. Experiments demonstrate that M2C achieves state-of-the-art few-shot performance on medical segmentation benchmarks, substantially improving both annotation efficiency and accuracy while requiring minimal expert intervention and offering high scalability.

0 citationsRead paper

Failing Forward: Adaptive Failure-Informed Learning for Vision-Language-Action Models

May 08, 2026

This work addresses a critical limitation in current vision-language-action (VLA) models, which rely solely on behavior cloning from successful demonstrations and thus struggle to recover from execution errors, often failing due to minor deviations. To overcome this, the authors propose an adaptive failure-aware learning framework that, for the first time, incorporates online-generated failure trajectories as adaptive negative supervision signals into end-to-end VLA training. The approach employs a dual-action generator architecture sharing a common vision-language backbone to jointly model successful and failed behaviors, coupled with a distribution-distance-driven sampling mechanism that dynamically avoids error-prone regions. Requiring no manually designed failure modes or external recovery interventions, the method achieves substantially improved policy robustness with minimal parameter overhead, outperforming existing VLA baselines in both in-domain and out-of-domain settings across short- and long-horizon manipulation tasks.

0 citationsRead paper