Selected work

Representative Papers

PILOT-Bench: A Benchmark for Legal Reasoning in the Patent Domain with IRAC-Aligned Classification Tasks

Jan 08, 2026Proceedings of the Natural Legal Language Processing Workshop 2025

This study addresses the absence of a systematic evaluation benchmark for large language models (LLMs) in the domain of patent law reasoning. The authors construct the first benchmark centered on decisions from the U.S. Patent Trial and Appeal Board (PTAB), aligning PTAB rulings with USPTO patent data to formulate three structured classification tasks grounded in the IRAC legal analysis framework: issue type, cited authority, and sub-decision. The benchmark enables multidimensional evaluation across input variations, model families, and error analyses, offering a comprehensive assessment of both open- and closed-source LLMs. Experimental results reveal a substantial performance gap: the best closed-source model achieves a Micro-F1 score of 0.75 on the issue-type task, whereas the strongest open-source model, Qwen-8B, attains only 0.56, highlighting significant limitations in current models’ capacity for patent-related legal reasoning.

1 citationsRead paper

Geometry-Aware Style Transfer in 3D Gaussian Splatting

Jun 23, 2026

Existing 3D Gaussian splatting-based style transfer methods are largely confined to color stylization and often neglect geometric adaptation, leading to inconsistencies in the overall scene structure. This work proposes a geometry-aware joint transfer framework that, for the first time, simultaneously optimizes appearance and geometric features within 3D Gaussian splatting. The approach employs a decoupled optimization strategy that alternately updates color and geometry parameters, complemented by a Geometry-aware Contrastive Feature Matching (GCFM) mechanism that integrates RGB, depth, and edge information for contrastive learning. By explicitly aligning geometric structure with visual style, the method effectively mitigates interference between color and geometry updates. Extensive experiments demonstrate that our approach significantly outperforms existing techniques both qualitatively and quantitatively, achieving high-fidelity style transfer with consistent 3D structural integrity.

0 citationsRead paper

Semantic Zone-Based Map Management for Stable AI-Integrated Mobile Robots

Mar 31, 2026

This work addresses the challenge of unstable localization and degraded performance in memory-constrained edge mobile robots when concurrently deploying large language models and dense maps, primarily caused by keyframe loading latency. The study introduces indoor semantic regions—such as rooms and corridors—into dense map management and proposes a semantic-aware keyframe scheduling strategy. By integrating SLAM, vision-language models, and semantic segmentation, the method prioritizes the loading of spatially relevant map content according to semantic regions on an NVIDIA Jetson Orin Nano platform. This approach not only maintains robust localization stability but also significantly enhances large model inference efficiency, achieving a 3.3 tokens/s increase in throughput and a 21.7% reduction in latency compared to purely geometric map management, while completely eliminating memory overflow and execution stuttering.

0 citationsRead paper

Soft Filtering: Guiding Zero-shot Composed Image Retrieval with Prescriptive and Proscriptive Constraints

Dec 23, 2025

Zero-shot compositional image retrieval (ZS-CIR) faces three key challenges: ambiguous user intent, entanglement of positive and negative semantics, and the mismatch between the single-target assumption and real-world query ambiguity. To address these, we propose SoFT—a training-free soft filtering module that introduces dual-track text constraint modeling for the first time, explicitly distinguishing prescriptive (“must include”) from proscriptive (“must avoid”) semantics. Leveraging multimodal large language models (MLLMs), SoFT dynamically parses both constraint types from reference images and modification texts to enable zero-shot re-ranking of retrieval results. Additionally, we design a pipeline for generating multi-objective CIR benchmarks supporting fine-grained, ambiguity-robust evaluation. On CIRR, CIRCO, and FashionIQ, SoFT achieves improvements of +12.94 in R@5, +6.13 in mAP@50, and +4.59 in R@50, significantly enhancing both robustness and accuracy of ZS-CIR.

0 citationsRead paper

Semantic Zone based 3D Map Management for Mobile Robot

Dec 13, 2025

For large-scale indoor environments (e.g., hospitals, logistics centers), mobile robots face excessive memory consumption in 3D mapping and inefficient cross-region retrieval due to conventional geometry- or time-driven SLAM memory management. This paper proposes a semantic-region-driven 3D map memory management method. It parses the environment into functionally explicit semantic units (e.g., corridors, lobbies), constructs a region graph at the semantic-region granularity, and designs a two-level memory scheduling mechanism integrated into RTAB-Map—combining real-time semantic segmentation with strict memory-threshold control. To our knowledge, this is the first work to shift the memory management paradigm from geometry/time-driven to semantic-driven. The approach significantly reduces redundant signature loading/unloading frequency and cumulative memory footprint while preserving navigational map usability, thereby achieving stable, predictable, and task-oriented memory utilization.

0 citationsRead paper

Recent publications

Latest Papers

Geometry-Aware Style Transfer in 3D Gaussian Splatting

Jun 23, 2026

Existing 3D Gaussian splatting-based style transfer methods are largely confined to color stylization and often neglect geometric adaptation, leading to inconsistencies in the overall scene structure. This work proposes a geometry-aware joint transfer framework that, for the first time, simultaneously optimizes appearance and geometric features within 3D Gaussian splatting. The approach employs a decoupled optimization strategy that alternately updates color and geometry parameters, complemented by a Geometry-aware Contrastive Feature Matching (GCFM) mechanism that integrates RGB, depth, and edge information for contrastive learning. By explicitly aligning geometric structure with visual style, the method effectively mitigates interference between color and geometry updates. Extensive experiments demonstrate that our approach significantly outperforms existing techniques both qualitatively and quantitatively, achieving high-fidelity style transfer with consistent 3D structural integrity.

0 citationsRead paper

Semantic Zone-Based Map Management for Stable AI-Integrated Mobile Robots

Mar 31, 2026

This work addresses the challenge of unstable localization and degraded performance in memory-constrained edge mobile robots when concurrently deploying large language models and dense maps, primarily caused by keyframe loading latency. The study introduces indoor semantic regions—such as rooms and corridors—into dense map management and proposes a semantic-aware keyframe scheduling strategy. By integrating SLAM, vision-language models, and semantic segmentation, the method prioritizes the loading of spatially relevant map content according to semantic regions on an NVIDIA Jetson Orin Nano platform. This approach not only maintains robust localization stability but also significantly enhances large model inference efficiency, achieving a 3.3 tokens/s increase in throughput and a 21.7% reduction in latency compared to purely geometric map management, while completely eliminating memory overflow and execution stuttering.

0 citationsRead paper

PILOT-Bench: A Benchmark for Legal Reasoning in the Patent Domain with IRAC-Aligned Classification Tasks

Jan 08, 2026Proceedings of the Natural Legal Language Processing Workshop 2025

This study addresses the absence of a systematic evaluation benchmark for large language models (LLMs) in the domain of patent law reasoning. The authors construct the first benchmark centered on decisions from the U.S. Patent Trial and Appeal Board (PTAB), aligning PTAB rulings with USPTO patent data to formulate three structured classification tasks grounded in the IRAC legal analysis framework: issue type, cited authority, and sub-decision. The benchmark enables multidimensional evaluation across input variations, model families, and error analyses, offering a comprehensive assessment of both open- and closed-source LLMs. Experimental results reveal a substantial performance gap: the best closed-source model achieves a Micro-F1 score of 0.75 on the issue-type task, whereas the strongest open-source model, Qwen-8B, attains only 0.56, highlighting significant limitations in current models’ capacity for patent-related legal reasoning.

1 citationsRead paper

Soft Filtering: Guiding Zero-shot Composed Image Retrieval with Prescriptive and Proscriptive Constraints

Dec 23, 2025

Zero-shot compositional image retrieval (ZS-CIR) faces three key challenges: ambiguous user intent, entanglement of positive and negative semantics, and the mismatch between the single-target assumption and real-world query ambiguity. To address these, we propose SoFT—a training-free soft filtering module that introduces dual-track text constraint modeling for the first time, explicitly distinguishing prescriptive (“must include”) from proscriptive (“must avoid”) semantics. Leveraging multimodal large language models (MLLMs), SoFT dynamically parses both constraint types from reference images and modification texts to enable zero-shot re-ranking of retrieval results. Additionally, we design a pipeline for generating multi-objective CIR benchmarks supporting fine-grained, ambiguity-robust evaluation. On CIRR, CIRCO, and FashionIQ, SoFT achieves improvements of +12.94 in R@5, +6.13 in mAP@50, and +4.59 in R@50, significantly enhancing both robustness and accuracy of ZS-CIR.

0 citationsRead paper

Semantic Zone based 3D Map Management for Mobile Robot

Dec 13, 2025

For large-scale indoor environments (e.g., hospitals, logistics centers), mobile robots face excessive memory consumption in 3D mapping and inefficient cross-region retrieval due to conventional geometry- or time-driven SLAM memory management. This paper proposes a semantic-region-driven 3D map memory management method. It parses the environment into functionally explicit semantic units (e.g., corridors, lobbies), constructs a region graph at the semantic-region granularity, and designs a two-level memory scheduling mechanism integrated into RTAB-Map—combining real-time semantic segmentation with strict memory-threshold control. To our knowledge, this is the first work to shift the memory management paradigm from geometry/time-driven to semantic-driven. The approach significantly reduces redundant signature loading/unloading frequency and cumulative memory footprint while preserving navigational map usability, thereby achieving stable, predictable, and task-oriented memory utilization.

0 citationsRead paper