Selected work

Representative Papers

MRBench: A Comprehensive Benchmark for Human Motion-Text Retrieval

Aug 08, 2026

Existing action–text retrieval benchmarks are limited by action homogeneity, imbalanced category distributions, and oversimplified textual descriptions, hindering the evaluation of cross-domain and cross-granularity alignment capabilities. To address these limitations, this work proposes MRBench—a comprehensive benchmark featuring heterogeneous actions, balanced category distribution, and multi-granularity textual descriptions—alongside a lightweight granularity-aware model. Leveraging multi-source action data and large language models, the authors construct 3,390 action instances paired with 10,170 multi-granularity text descriptions. The model employs a frozen backbone with branch-specific adapters and a comparability-preserving score fusion strategy. Experiments reveal that existing methods exhibit significant generalization gaps and granularity sensitivity on MRBench, whereas the proposed approach substantially improves fine-grained retrieval performance while maintaining competitive standard retrieval accuracy.

0 citationsRead paper

Improving Auto-Design of Neural PDE Solvers with a Domain-Specific Language

Aug 04, 2026

This work addresses the challenge that effective neural PDE solvers are extremely sparse in the design space, rendering large language models inefficient for their automated discovery. To overcome this, the authors propose ADSL-PDE, a framework that introduces structured search states to decouple high-level design decisions—such as architecture, physical constraints, and optimization objectives—from low-level code implementation. By formulating a domain-specific language to restructure the search space, the method substantially increases the density of valid candidate solvers. Building upon this representation, the framework integrates deterministic compiler-based mapping with an evolutionary algorithm guided by empirical feedback and large language model suggestions. Experiments across multiple PDE benchmarks demonstrate that the approach improves solver performance by over 52% within the first ten iterations, while significantly enhancing search efficiency and optimization stability.

0 citationsRead paper

Recent publications

Latest Papers

MRBench: A Comprehensive Benchmark for Human Motion-Text Retrieval

Aug 08, 2026

Existing action–text retrieval benchmarks are limited by action homogeneity, imbalanced category distributions, and oversimplified textual descriptions, hindering the evaluation of cross-domain and cross-granularity alignment capabilities. To address these limitations, this work proposes MRBench—a comprehensive benchmark featuring heterogeneous actions, balanced category distribution, and multi-granularity textual descriptions—alongside a lightweight granularity-aware model. Leveraging multi-source action data and large language models, the authors construct 3,390 action instances paired with 10,170 multi-granularity text descriptions. The model employs a frozen backbone with branch-specific adapters and a comparability-preserving score fusion strategy. Experiments reveal that existing methods exhibit significant generalization gaps and granularity sensitivity on MRBench, whereas the proposed approach substantially improves fine-grained retrieval performance while maintaining competitive standard retrieval accuracy.

0 citationsRead paper

Improving Auto-Design of Neural PDE Solvers with a Domain-Specific Language

Aug 04, 2026

This work addresses the challenge that effective neural PDE solvers are extremely sparse in the design space, rendering large language models inefficient for their automated discovery. To overcome this, the authors propose ADSL-PDE, a framework that introduces structured search states to decouple high-level design decisions—such as architecture, physical constraints, and optimization objectives—from low-level code implementation. By formulating a domain-specific language to restructure the search space, the method substantially increases the density of valid candidate solvers. Building upon this representation, the framework integrates deterministic compiler-based mapping with an evolutionary algorithm guided by empirical feedback and large language model suggestions. Experiments across multiple PDE benchmarks demonstrate that the approach improves solver performance by over 52% within the first ten iterations, while significantly enhancing search efficiency and optimization stability.

0 citationsRead paper