Selected work

Representative Papers

Adaptive 3D UI Placement in Mixed Reality Using Deep Reinforcement Learning

May 11, 2024CHI Extended Abstracts

To address the challenge of real-time adaptive 3D UI layout under dynamic user pose and environmental changes in mixed reality (MR), this paper introduces Proximal Policy Optimization (PPO)—the first application of deep reinforcement learning to continuous 3D UI placement. We propose a multimodal state encoding mechanism that fuses user pose with scene geometry, and design a task-oriented reward function to guide the policy network toward personalized, online-adaptive layout generation. Unlike conventional optimization methods relying on static assumptions, our approach overcomes their limitations by enabling real-time responsiveness in dynamic environments. Experiments demonstrate a 23% average improvement in task completion efficiency in mobile scenarios, layout response latency under 80 ms, and significant enhancements in content visibility and interaction accessibility.

3 citationsRead paper

The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed Samples

Jan 29, 2026

This study addresses a critical privacy vulnerability in existing machine unlearning methods: despite formal removal of target data, models may retain residual recognition capabilities when exposed to adversarially perturbed inputs, thereby risking unintended information leakage. To tackle this issue, the work formally defines the problem of residual knowledge under input perturbations in high-dimensional settings and introduces RURK, a penalty-based fine-tuning strategy designed to suppress a model’s ability to re-identify perturbed forgotten samples. Extensive experiments demonstrate that mainstream unlearning approaches commonly exhibit such residual knowledge, whereas RURK effectively mitigates this risk, significantly enhancing forgetting security across standard vision benchmarks.

1 citationsRead paper

Examining the Effects of Immersive and Non-Immersive Presenter Modalities on Engagement and Social Interaction in Co-located Augmented Presentations

Mar 17, 2025

How do presenter interface modalities—immersive (HoloLens 2) versus non-immersive (handheld tablet)—affect audience engagement, group awareness, and social interaction in co-located augmented reality (AR) demonstrations? Method: We designed a switchable-modality AR demonstration system and conducted 12 real-world, one-on-one co-located user studies, systematically comparing symmetric (both presenter and audience in the same AR environment) versus asymmetric (presenter using tablet) configurations. Behavioral coding of interactions yielded four interaction themes. Contribution/Results: We identified six design strategies balancing immersion and natural sociability. Asymmetric mode significantly increased audience eye-contact frequency (+37%) and conversational naturalness (p < 0.01), confirming its advantage in preserving social presence. Our findings provide a practical, empirically grounded design trade-off framework for collaborative AR presentations.

1 citationsRead paper

Adaptive and Robust Watermark for Generative Tabular Data

Sep 23, 2024arXiv.org

Generative tabular data faces critical challenges in authenticity verification and risks of malicious misuse. Method: We propose a downstream-task-aware adaptive watermarking mechanism. Our approach partitions features into (key, value) column pairs; for each key column, it dynamically generates a random “green” value interval, constraining the corresponding value column to sample exclusively within this interval. Watermark embedding and efficient detection are achieved through feature-space partitioning, constraint-based generation, and statistical hypothesis testing. Contribution/Results: This work presents the first customizable, statistically provably secure watermarking scheme for tabular data, robust against multiple adversarial attacks—including noise injection, column deletion, and row resampling. Experiments demonstrate near-lossless preservation of statistical fidelity and downstream task performance post-embedding, alongside high detection accuracy and strong robustness.

1 citationsRead paper

Beyond the Skew-Stickiness Ratio: Transport Geometry of Spot-Driven Variance Surface Dynamics

Aug 12, 2026

This study investigates the dynamics of implied variance surfaces under no-arbitrage constraints as the underlying asset price evolves. It introduces a novel transport-geometric framework that models smile dynamics as a transport flow on a static no-arbitrage surface, driven by movements in the underlying price. A velocity field \(v(k)\) is proposed to unify various sticky mechanisms: the sticky-strike rule (SSR) emerges as the zeroth-order term, while higher-order terms capture features such as at-the-money skew and curvature. Nonparametric identification is achieved via local jet transport arguments. Empirical analysis of SPX options reveals significant excess skewness, rejecting the self-similar transport hypothesis; the velocity field exhibits strong heterogeneity across strike and maturity dimensions. The proposed three-parameter model improves medium-term curvature forecasting accuracy by 17–21% relative to SSR.

0 citationsRead paper

Recent publications

Latest Papers

Beyond the Skew-Stickiness Ratio: Transport Geometry of Spot-Driven Variance Surface Dynamics

Aug 12, 2026

This study investigates the dynamics of implied variance surfaces under no-arbitrage constraints as the underlying asset price evolves. It introduces a novel transport-geometric framework that models smile dynamics as a transport flow on a static no-arbitrage surface, driven by movements in the underlying price. A velocity field \(v(k)\) is proposed to unify various sticky mechanisms: the sticky-strike rule (SSR) emerges as the zeroth-order term, while higher-order terms capture features such as at-the-money skew and curvature. Nonparametric identification is achieved via local jet transport arguments. Empirical analysis of SPX options reveals significant excess skewness, rejecting the self-similar transport hypothesis; the velocity field exhibits strong heterogeneity across strike and maturity dimensions. The proposed three-parameter model improves medium-term curvature forecasting accuracy by 17–21% relative to SSR.

0 citationsRead paper

Inverse Theory of Mind Modeling for Content Recommendation: From Web Browsing to Dynamic Intelligent Interfaces

Aug 11, 2026

Traditional recommender systems simplify user behavior into static preferences, failing to capture complex intents such as exploration and comparison, thereby limiting their effectiveness in dynamic environments like generative user interfaces and extended reality. This work introduces inverse Theory of Mind (IToM) into recommender systems for the first time, inferring users’ underlying beliefs, preferences, and decision-making traits through counterfactual reasoning and multi-hypothesis abductive inference powered by large language models. The resulting structured, interpretable user profiles enable cross-modal transfer and intent-driven content presentation. Evaluated on the OPeRA dataset, the approach matches or surpasses performance using ground-truth user profiles across diverse tasks—including next-action prediction, shopping attitude alignment, Big Five personality inference, and category prediction—and has been successfully deployed in a VisionOS spatial banking application.

0 citationsRead paper

Faster Algorithms for Multimarginal Optimal Transport

Aug 10, 2026

This work proposes a novel representation learning framework that integrates adaptive multi-scale fusion with contrastive learning to address the limited representational capacity of existing methods in complex scenarios. By dynamically aggregating multi-level features and incorporating a structure-aware contrastive loss, the approach effectively enhances the model’s ability to jointly capture fine-grained semantics and global context. Extensive experiments demonstrate that the proposed framework consistently outperforms state-of-the-art methods across multiple benchmark datasets, achieving substantial improvements in both accuracy and robustness. Notably, it exhibits superior performance under low-resource settings and in the presence of noise, thereby providing a more powerful and generalizable feature representation for downstream tasks.

0 citationsRead paper

The Missing Link of XR: Empathy-Driven Reality for XR and Beyond

Aug 08, 2026

Current extended reality (XR) social systems generally lack the capacity to understand users’ affective, physiological, and cognitive states, hindering safe and equitable participation for diverse individuals. To address this gap, this work proposes Empathy-Driven Reality (EDR), an innovative conceptual framework that positions empathy as a foundational design principle for XR systems. Through interdisciplinary hybrid workshops integrating perspectives from XR, affective computing, physiological sensing, assistive technologies, and social sciences, the study articulates an empathy-centered design space. The project yields an EDR design guideline, a set of key research challenges, and representative application scenarios, collectively offering a theoretical foundation and practical roadmap for developing inclusive, emotionally intelligent immersive social systems.

0 citationsRead paper

Confidence Estimation for Financial Vision-Language Models in Chart and Document Understanding

Aug 06, 2026

In financial visual question answering (VQA), large vision-language models (LVLMs) often produce fluent yet hallucinated responses unsupported by the underlying charts, posing significant decision-making risks. This work systematically evaluates seven confidence estimation methods—three inference-time baselines and four trained probes—across five open-source LVLMs and four financial VQA tasks, with all probes trained exclusively on natural images to assess cross-domain generalization. The study reveals that well-calibrated confidence is rarer than reliable ranking performance and that reliability varies substantially across model–task pairs. It introduces an actionable, error-budget-based automation threshold and proposes, for the first time, a grounding-aware probe that effectively distinguishes fluent but unfounded guesses from chart-grounded answers. Results show that only trained probes yield reliable calibration scores, enabling partial automation under lenient conditions, yet high-difficulty tasks remain largely non-automatable within a 5% error budget.

0 citationsRead paper