Selected work

Representative Papers

Tellimation: Making Narrative Gaps Visible in Children's Storytelling with Just-in-Time Animation

Oct 04, 2026

This study addresses the narrative incoherence exhibited by children with language impairments, whose storytelling frequently omits essential details and logical relations. To mitigate this, we introduce the concept of visualizing narrative gaps and construct a design space for real-time detection and animation generation spanning eight narrative dimensions. The system comprises twenty parametric animations grounded in classical animation principles, dynamically matched through voice-scenario discrepancy detection and historical context algorithms to provide visual scaffolding rather than verbal prompting. This work demonstrates the efficacy of non-verbal intervention: adults comprehended most animations without instruction, while children successfully filled 46% of narrative gaps with visual assistance, representing a substantial improvement over the unassisted baseline of 13%.

0 citationsRead paper

SyntaxBench: A Statistical Diagnostic Framework for Character-Level Reasoning in Large Language Models

Oct 02, 2026

This study addresses the reliance of existing character-level reasoning evaluations for large language models on isolated probes and aggregated metrics, which lack fine-grained diagnostics. We construct a statistical diagnostic benchmark comprising six task categories. Through controlled input design and a paired statistical testing framework integrating McNemar’s test, bootstrap confidence intervals, and tokenization visualization, we systematically evaluate models’ character-processing capabilities under zero- to four-shot settings. Our findings reveal differential effects of tokenization mechanisms and reasoning paradigms on character-level accuracy: random strings elicit stronger character awareness than natural English text, chain-of-thought prompting does not necessarily improve performance, and substring extraction tasks yield notably low accuracy. These insights offer a novel perspective for understanding the underlying character-processing mechanisms of large language models.

0 citationsRead paper

Privacy Foundations for Multi-Institutional Scientific Artificial Intelligence

Sep 30, 2026

This study addresses the challenge of composing privacy guarantees for multi-institutional scientific AI operating in mixed-trust environments by proposing a unified privacy framework. Methodologically, it reformulates privacy as a six-element assurance problem and establishes a declarative registry to evaluate risks across the model lifecycle. The framework orchestrates synergistic protections by integrating differential privacy, federated learning, secure multi-party computation, and trusted execution environments. Key contributions include identifying critical research gaps specific to leadership-class facilities, such as metadata leakage and instrument side channels, and translating privacy assurances into a universal format that is declarable, comparable, and auditable. Furthermore, this work delineates six priority directions for privacy research, ultimately providing a systematic governance paradigm for cross-institutional scientific collaboration.

0 citationsRead paper

Motion Concept Unlearning in Video Diffusion Models

Sep 29, 2026

This study addresses the challenge of safely erasing action concepts in video diffusion models by proposing MUTE, a training-free method. To our knowledge, this work presents the first systematic investigation into action erasure within video Diffusion Transformers (DiTs), establishing three essential criteria: specificity, spatial selectivity, and temporal naturalness. By conducting causal intervention analysis on attention mechanisms, MUTE extracts concept directions via token neutralization and achieves precise removal through spatial gating combined with velocity correction prior to classifier-free guidance. Experiments on Wan2.1 and CogVideoX demonstrate that MUTE significantly outperforms existing baselines, effectively suppressing target actions while preserving the natural dynamics of non-target motions.

0 citationsRead paper

Agentic Relative Camera Pose Estimation via Learned Ranking and Verification

Sep 29, 2026

This study addresses the challenge that single camera pose estimators struggle with complex scenarios involving wide baselines and textureless regions. To overcome this limitation, we propose PoseAgent, a multi-agent framework that introduces learnable ranking and verification mechanisms. By employing profiling, ranking, and verification agents, PoseAgent dynamically orchestrates multiple candidate estimators to adaptively schedule the optimal solution. The core innovation lies in constructing a pose error verification network that surpasses existing models, enabling precise estimator evaluation and efficient collaboration. Experimental results demonstrate that PoseAgent achieves up to a 4.2% improvement in AUC@5° across multiple benchmarks, significantly outperforming both the strongest individual estimators and recent vision-language model-based approaches.

0 citationsRead paper

Recent publications

Latest Papers

Tellimation: Making Narrative Gaps Visible in Children's Storytelling with Just-in-Time Animation

Oct 04, 2026

This study addresses the narrative incoherence exhibited by children with language impairments, whose storytelling frequently omits essential details and logical relations. To mitigate this, we introduce the concept of visualizing narrative gaps and construct a design space for real-time detection and animation generation spanning eight narrative dimensions. The system comprises twenty parametric animations grounded in classical animation principles, dynamically matched through voice-scenario discrepancy detection and historical context algorithms to provide visual scaffolding rather than verbal prompting. This work demonstrates the efficacy of non-verbal intervention: adults comprehended most animations without instruction, while children successfully filled 46% of narrative gaps with visual assistance, representing a substantial improvement over the unassisted baseline of 13%.

0 citationsRead paper

SyntaxBench: A Statistical Diagnostic Framework for Character-Level Reasoning in Large Language Models

Oct 02, 2026

This study addresses the reliance of existing character-level reasoning evaluations for large language models on isolated probes and aggregated metrics, which lack fine-grained diagnostics. We construct a statistical diagnostic benchmark comprising six task categories. Through controlled input design and a paired statistical testing framework integrating McNemar’s test, bootstrap confidence intervals, and tokenization visualization, we systematically evaluate models’ character-processing capabilities under zero- to four-shot settings. Our findings reveal differential effects of tokenization mechanisms and reasoning paradigms on character-level accuracy: random strings elicit stronger character awareness than natural English text, chain-of-thought prompting does not necessarily improve performance, and substring extraction tasks yield notably low accuracy. These insights offer a novel perspective for understanding the underlying character-processing mechanisms of large language models.

0 citationsRead paper

Privacy Foundations for Multi-Institutional Scientific Artificial Intelligence

Sep 30, 2026

This study addresses the challenge of composing privacy guarantees for multi-institutional scientific AI operating in mixed-trust environments by proposing a unified privacy framework. Methodologically, it reformulates privacy as a six-element assurance problem and establishes a declarative registry to evaluate risks across the model lifecycle. The framework orchestrates synergistic protections by integrating differential privacy, federated learning, secure multi-party computation, and trusted execution environments. Key contributions include identifying critical research gaps specific to leadership-class facilities, such as metadata leakage and instrument side channels, and translating privacy assurances into a universal format that is declarable, comparable, and auditable. Furthermore, this work delineates six priority directions for privacy research, ultimately providing a systematic governance paradigm for cross-institutional scientific collaboration.

0 citationsRead paper

Motion Concept Unlearning in Video Diffusion Models

Sep 29, 2026

This study addresses the challenge of safely erasing action concepts in video diffusion models by proposing MUTE, a training-free method. To our knowledge, this work presents the first systematic investigation into action erasure within video Diffusion Transformers (DiTs), establishing three essential criteria: specificity, spatial selectivity, and temporal naturalness. By conducting causal intervention analysis on attention mechanisms, MUTE extracts concept directions via token neutralization and achieves precise removal through spatial gating combined with velocity correction prior to classifier-free guidance. Experiments on Wan2.1 and CogVideoX demonstrate that MUTE significantly outperforms existing baselines, effectively suppressing target actions while preserving the natural dynamics of non-target motions.

0 citationsRead paper

Agentic Relative Camera Pose Estimation via Learned Ranking and Verification

Sep 29, 2026

This study addresses the challenge that single camera pose estimators struggle with complex scenarios involving wide baselines and textureless regions. To overcome this limitation, we propose PoseAgent, a multi-agent framework that introduces learnable ranking and verification mechanisms. By employing profiling, ranking, and verification agents, PoseAgent dynamically orchestrates multiple candidate estimators to adaptively schedule the optimal solution. The core innovation lies in constructing a pose error verification network that surpasses existing models, enabling precise estimator evaluation and efficient collaboration. Experimental results demonstrate that PoseAgent achieves up to a 4.2% improvement in AUC@5° across multiple benchmarks, significantly outperforming both the strongest individual estimators and recent vision-language model-based approaches.

0 citationsRead paper