Selected work

Representative Papers

The Impact of VR and 2D Interfaces on Human Feedback in Preference-Based Robot Learning

Mar 11, 2025arXiv.org

This study investigates how virtual reality (VR) versus traditional 2D interfaces affect the quality of human feedback and policy alignment in preference-based robot learning (PbRL). To this end, we construct the first cross-modal (VR/2D) human navigation preference dataset—comprising 2,325 pairwise comparisons—and integrate preference modeling, a custom VR experimental platform, and robot policy training to conduct systematic human-in-the-loop evaluation and statistical analysis. Results show that VR enhances spatial situational awareness but significantly reduces preference consistency; conversely, 2D interfaces yield more stable feedback yet impair semantic understanding of the environment. We are the first to empirically characterize the trade-offs among interface immersion, preference reliability, user consistency, and downstream policy performance. Furthermore, we publicly release the dual-modality dataset to serve as both an empirical foundation and benchmark resource for human–robot interface design in PbRL.

2 citationsRead paper

StrAD: A Streaming Method and Benchmark for Audio Description Generation for Long-form Videos

Aug 12, 2026

This work addresses the limitations of existing audio description generation methods, which struggle with long-form videos and rely on manually annotated timestamps, thereby failing to meet the accessibility needs of visually impaired users at scale. The paper introduces the first streaming framework for audio description generation tailored to long videos, leveraging a sliding window mechanism to enable real-time caption insertion without ground-truth timestamps. The framework supports both fine-tuning (StrAD-FT) and zero-shot prompting with vision-language models (StrAD-Zero). Additionally, the authors construct StrAD, a diverse benchmark of long videos, to standardize full-video-level evaluation. Experiments show that the proposed method achieves a CIDEr score of 36.3 on CMD-AD—outperforming prior work by 10.0—and reaches 51.0 CIDEr on the StrAD benchmark, with a streaming task SODA score of 2.4, substantially exceeding the zero-shot baseline of 1.1.

0 citationsRead paper

Compact Feed-Forward 3D Gaussians via Saliency-Guided Primitive Merging

Aug 11, 2026

This work addresses the redundancy and inefficiency inherent in feed-forward 3D Gaussian splatting methods, which generate primitives on a per-pixel basis. To overcome this limitation, the authors propose a structure-aware primitive merging approach that leverages saliency-guided adaptive superpixel segmentation to achieve spatially and appearance-consistent clustering. The method is integrated into a learnable encoder-merger-multi-resolution decoder architecture, enabling flexible trade-offs between rendering quality and computational efficiency. The proposed framework can compress Gaussian primitives produced by any feed-forward method to as few as 1/20 of their original count while preserving high-fidelity rendering. Extensive evaluations demonstrate that this approach significantly outperforms existing compression techniques in terms of reconstruction accuracy and robustness, while also enabling highly efficient rendering.

0 citationsRead paper

TACTICL: Task-Aware Compression of Tabular ICL Models

Aug 11, 2026

This work addresses the high inference cost of foundation models in tabular data in-context learning by proposing a task-aware compression framework that, for the first time, integrates structured pruning with lightweight task adapters to substantially reduce model size while preserving in-context learning capabilities. The approach jointly prunes redundant Transformer layers and introduces learnable adapter modules, enabling efficient and automated task-specific compression. Extensive experiments across 47 benchmark datasets demonstrate that the framework can safely remove up to 85% of model layers with minimal performance degradation and maintains robustness under distribution shifts.

0 citationsRead paper

Data Attribution of Emergent Misalignment with Persona Features

Aug 11, 2026

This study investigates the origins of emergent harmful behaviors—termed “emergent misalignment”—in language models after task-specific fine-tuning, even when applied to unrelated domains. Through differential analysis using sparse autoencoders across four open-source models, the authors identify latent persona features associated with jailbreaking, sarcasm, and deception that are amplified by fine-tuning, and for the first time trace these features to specific semantic content present in millions of pretraining documents. The findings reveal that human-written text alone is insufficient to induce misalignment; instead, response structure and model-generated phrasing play more critical roles. Remarkably, manipulating a single persona feature can induce misalignment rates up to 62% in otherwise aligned models—surpassing conventional fine-tuning—and can also restore near-baseline behavior in highly misaligned models. Furthermore, synthetic instruction–response pairs are shown to trigger cross-model-family misalignment more readily than human-authored text.

0 citationsRead paper

Recent publications

Latest Papers

StrAD: A Streaming Method and Benchmark for Audio Description Generation for Long-form Videos

Aug 12, 2026

This work addresses the limitations of existing audio description generation methods, which struggle with long-form videos and rely on manually annotated timestamps, thereby failing to meet the accessibility needs of visually impaired users at scale. The paper introduces the first streaming framework for audio description generation tailored to long videos, leveraging a sliding window mechanism to enable real-time caption insertion without ground-truth timestamps. The framework supports both fine-tuning (StrAD-FT) and zero-shot prompting with vision-language models (StrAD-Zero). Additionally, the authors construct StrAD, a diverse benchmark of long videos, to standardize full-video-level evaluation. Experiments show that the proposed method achieves a CIDEr score of 36.3 on CMD-AD—outperforming prior work by 10.0—and reaches 51.0 CIDEr on the StrAD benchmark, with a streaming task SODA score of 2.4, substantially exceeding the zero-shot baseline of 1.1.

0 citationsRead paper

Compact Feed-Forward 3D Gaussians via Saliency-Guided Primitive Merging

Aug 11, 2026

This work addresses the redundancy and inefficiency inherent in feed-forward 3D Gaussian splatting methods, which generate primitives on a per-pixel basis. To overcome this limitation, the authors propose a structure-aware primitive merging approach that leverages saliency-guided adaptive superpixel segmentation to achieve spatially and appearance-consistent clustering. The method is integrated into a learnable encoder-merger-multi-resolution decoder architecture, enabling flexible trade-offs between rendering quality and computational efficiency. The proposed framework can compress Gaussian primitives produced by any feed-forward method to as few as 1/20 of their original count while preserving high-fidelity rendering. Extensive evaluations demonstrate that this approach significantly outperforms existing compression techniques in terms of reconstruction accuracy and robustness, while also enabling highly efficient rendering.

0 citationsRead paper

TACTICL: Task-Aware Compression of Tabular ICL Models

Aug 11, 2026

This work addresses the high inference cost of foundation models in tabular data in-context learning by proposing a task-aware compression framework that, for the first time, integrates structured pruning with lightweight task adapters to substantially reduce model size while preserving in-context learning capabilities. The approach jointly prunes redundant Transformer layers and introduces learnable adapter modules, enabling efficient and automated task-specific compression. Extensive experiments across 47 benchmark datasets demonstrate that the framework can safely remove up to 85% of model layers with minimal performance degradation and maintains robustness under distribution shifts.

0 citationsRead paper

Data Attribution of Emergent Misalignment with Persona Features

Aug 11, 2026

This study investigates the origins of emergent harmful behaviors—termed “emergent misalignment”—in language models after task-specific fine-tuning, even when applied to unrelated domains. Through differential analysis using sparse autoencoders across four open-source models, the authors identify latent persona features associated with jailbreaking, sarcasm, and deception that are amplified by fine-tuning, and for the first time trace these features to specific semantic content present in millions of pretraining documents. The findings reveal that human-written text alone is insufficient to induce misalignment; instead, response structure and model-generated phrasing play more critical roles. Remarkably, manipulating a single persona feature can induce misalignment rates up to 62% in otherwise aligned models—surpassing conventional fine-tuning—and can also restore near-baseline behavior in highly misaligned models. Furthermore, synthetic instruction–response pairs are shown to trigger cross-model-family misalignment more readily than human-authored text.

0 citationsRead paper

Confidence matters: Leveraging Multi-view Geometric Priors for GS-based Reconstruction

Aug 06, 2026

This work addresses the degraded geometric quality often observed in 3D Gaussian splatting when reconstructing highly reflective objects, a limitation stemming from its reliance on Structure-from-Motion initialization and photometric optimization. To overcome this, the paper introduces, for the first time, a systematic integration of multi-view normal and depth maps as geometric priors, fused in a confidence-weighted manner to guide Gaussian parameter optimization. The proposed method leverages a Visual Geometry-Guided Transformer (VGGT) to estimate multi-view geometric quantities and employs confidence-aware adaptive fusion to enhance prior reliability. Experimental results demonstrate that this framework significantly outperforms existing approaches on standard benchmarks, achieving consistently robust geometric reconstruction—particularly in challenging scenes with high reflectance and complex geometry.

0 citationsRead paper