Selected work

Representative Papers

Rethinking the Flow-Based Gradual Domain Adaption: A Semi-Dual Optimal Transport Perspective

Feb 01, 2026

This work addresses the information loss and performance degradation in progressive domain adaptation caused by reliance on sample-based log-likelihood estimation. To overcome this limitation, the authors propose an Entropy-Regularized Semi-dual Unbalanced Optimal Transport framework (E-SUOT). By constructing a sample-based intermediate domain, E-SUOT reformulates the flow-model-driven adaptation process as a Lagrangian dual problem and derives an equivalent semi-dual objective that circumvents explicit likelihood estimation. This formulation transforms the unstable minimax training paradigm into a stable alternating optimization procedure, for which the authors provide theoretical guarantees on stability and generalization. Experimental results demonstrate that the proposed framework significantly improves performance across multiple benchmarks in progressive domain adaptation.

0 citationsRead paper

Bridging Semantic Understanding and Popularity Bias with LLMs

Jan 14, 2026

This work addresses the limitations of existing recommendation systems in mitigating popularity bias, which often focus on superficial metrics such as diversity or long-tail coverage while neglecting the semantic origins of bias, thereby compromising both debiasing efficacy and recommendation accuracy. To overcome this, we propose FairLRM, a novel framework that, for the first time, decomposes popularity bias into dual perspectives—item-side and user-side—and leverages a large language model (RecLLM) with structured instructional prompting to model the causal mechanisms of bias at the semantic level. By capturing the underlying semantics of biased interactions, FairLRM transcends conventional debiasing strategies, significantly enhancing both fairness and accuracy in recommendations, and yielding more trustworthy and semantically aware recommendation outcomes.

0 citationsRead paper

FairGU: Fairness-aware Graph Unlearning in Social Networks

Jan 14, 2026

This work addresses the critical issue that existing graph unlearning methods often inadvertently leak or amplify sensitive attributes when deleting nodes, thereby compromising algorithmic fairness. To mitigate this, we propose FairGU—the first framework that explicitly integrates fairness guarantees into the graph unlearning process. FairGU employs a fairness-aware module in conjunction with a structure-preserving strategy to jointly optimize the removal of node influence while effectively suppressing the exposure and amplification of sensitive attributes. Extensive experiments on multiple real-world graph datasets demonstrate that FairGU significantly outperforms current graph unlearning and fairness-enhancing baselines, achieving substantial improvements in fairness metrics without sacrificing model utility.

0 citationsRead paper

FairGE: Fairness-Aware Graph Encoding in Incomplete Social Networks

Jan 14, 2026

This work addresses the challenge of fairness in graph representation learning when sensitive attributes are missing in social networks. Existing graph Transformer methods often reconstruct sensitive attributes, inadvertently introducing bias and risking privacy leakage. To overcome this, the authors propose FairGE, a novel framework that encodes fairness directly through spectral graph theory without generating sensitive attributes. FairGE leverages the leading eigenvectors of the graph Laplacian to capture structural information and employs zero-padding for missing attributes to preserve independence. This approach effectively mitigates bias amplification and privacy concerns. Extensive experiments on seven real-world datasets demonstrate that FairGE improves both statistical parity and equal opportunity by at least 16% on average compared to state-of-the-art baselines.

0 citationsRead paper

SUPERChem: A Multimodal Reasoning Benchmark in Chemistry

Nov 30, 2025

Existing chemical reasoning benchmarks suffer from task oversimplification, inadequate process evaluation, and misalignment with expert-level capabilities. To address these limitations, we introduce SUPERChem, a multimodal benchmark comprising 500 expert-crafted, cross-subfield challenging problems. SUPERChem introduces Reasoning Path Fidelity (RPF), a novel scoring metric that quantifies reasoning quality by comparing model-generated solution paths against expert-annotated ground-truth traces. It employs an original content generation and iterative curation pipeline to ensure zero data contamination. Integrating both textual and visual problem formulations, SUPERChem establishes a human–machine comparative evaluation framework enabling analysis of visual modality’s impact on chemical reasoning. Human experts achieve a baseline accuracy of 40.3%, while the strongest evaluated model—GPT-5 (High)—scores only 38.5%, confirming the benchmark’s rigor and discriminative power. SUPERChem is the first benchmark to enable systematic, quantitative assessment of expert-level chemical reasoning processes.

0 citationsRead paper

Recent publications

Latest Papers

Rethinking the Flow-Based Gradual Domain Adaption: A Semi-Dual Optimal Transport Perspective

Feb 01, 2026

This work addresses the information loss and performance degradation in progressive domain adaptation caused by reliance on sample-based log-likelihood estimation. To overcome this limitation, the authors propose an Entropy-Regularized Semi-dual Unbalanced Optimal Transport framework (E-SUOT). By constructing a sample-based intermediate domain, E-SUOT reformulates the flow-model-driven adaptation process as a Lagrangian dual problem and derives an equivalent semi-dual objective that circumvents explicit likelihood estimation. This formulation transforms the unstable minimax training paradigm into a stable alternating optimization procedure, for which the authors provide theoretical guarantees on stability and generalization. Experimental results demonstrate that the proposed framework significantly improves performance across multiple benchmarks in progressive domain adaptation.

0 citationsRead paper

Bridging Semantic Understanding and Popularity Bias with LLMs

Jan 14, 2026

This work addresses the limitations of existing recommendation systems in mitigating popularity bias, which often focus on superficial metrics such as diversity or long-tail coverage while neglecting the semantic origins of bias, thereby compromising both debiasing efficacy and recommendation accuracy. To overcome this, we propose FairLRM, a novel framework that, for the first time, decomposes popularity bias into dual perspectives—item-side and user-side—and leverages a large language model (RecLLM) with structured instructional prompting to model the causal mechanisms of bias at the semantic level. By capturing the underlying semantics of biased interactions, FairLRM transcends conventional debiasing strategies, significantly enhancing both fairness and accuracy in recommendations, and yielding more trustworthy and semantically aware recommendation outcomes.

0 citationsRead paper

FairGU: Fairness-aware Graph Unlearning in Social Networks

Jan 14, 2026

This work addresses the critical issue that existing graph unlearning methods often inadvertently leak or amplify sensitive attributes when deleting nodes, thereby compromising algorithmic fairness. To mitigate this, we propose FairGU—the first framework that explicitly integrates fairness guarantees into the graph unlearning process. FairGU employs a fairness-aware module in conjunction with a structure-preserving strategy to jointly optimize the removal of node influence while effectively suppressing the exposure and amplification of sensitive attributes. Extensive experiments on multiple real-world graph datasets demonstrate that FairGU significantly outperforms current graph unlearning and fairness-enhancing baselines, achieving substantial improvements in fairness metrics without sacrificing model utility.

0 citationsRead paper

FairGE: Fairness-Aware Graph Encoding in Incomplete Social Networks

Jan 14, 2026

This work addresses the challenge of fairness in graph representation learning when sensitive attributes are missing in social networks. Existing graph Transformer methods often reconstruct sensitive attributes, inadvertently introducing bias and risking privacy leakage. To overcome this, the authors propose FairGE, a novel framework that encodes fairness directly through spectral graph theory without generating sensitive attributes. FairGE leverages the leading eigenvectors of the graph Laplacian to capture structural information and employs zero-padding for missing attributes to preserve independence. This approach effectively mitigates bias amplification and privacy concerns. Extensive experiments on seven real-world datasets demonstrate that FairGE improves both statistical parity and equal opportunity by at least 16% on average compared to state-of-the-art baselines.

0 citationsRead paper

SUPERChem: A Multimodal Reasoning Benchmark in Chemistry

Nov 30, 2025

Existing chemical reasoning benchmarks suffer from task oversimplification, inadequate process evaluation, and misalignment with expert-level capabilities. To address these limitations, we introduce SUPERChem, a multimodal benchmark comprising 500 expert-crafted, cross-subfield challenging problems. SUPERChem introduces Reasoning Path Fidelity (RPF), a novel scoring metric that quantifies reasoning quality by comparing model-generated solution paths against expert-annotated ground-truth traces. It employs an original content generation and iterative curation pipeline to ensure zero data contamination. Integrating both textual and visual problem formulations, SUPERChem establishes a human–machine comparative evaluation framework enabling analysis of visual modality’s impact on chemical reasoning. Human experts achieve a baseline accuracy of 40.3%, while the strongest evaluated model—GPT-5 (High)—scores only 38.5%, confirming the benchmark’s rigor and discriminative power. SUPERChem is the first benchmark to enable systematic, quantitative assessment of expert-level chemical reasoning processes.

0 citationsRead paper