Kuo Yang

Meta-review auto-generation (MRAG) has emerged as a promising solution to accelerate academic publishing by automatically generating structured, concise meta-reviews with minimal human intervention. Although existing MRAG methods have shown feasibility, they are limited by data constraints, suboptimal modeling strategies, and inefficient training paradigms. In this paper, we propose a novel approach to MRAG that overcomes these limitations from three key perspectives. First, we introduce MetaSum, the largest and most up-to-date meta-review dataset, covering $31$ conferences, $225$ venues, and $52,993$ submissions across various research domains. Second, unlike prior methods that truncate lengthy reviews, we treat MRAG as a long-context modeling task to retain complete textual information. Third, we develop a multi-role, dialogue-style training framework with efficient end-to-end optimization, fully leveraging LLMs for meta-review generation. Extensive experiments demonstrate that our approach significantly outperforms existing MRAG baselines and state-of-the-art LLMs. Our dataset and code will be made open-source.

Investigated the phenomenon of "thinking collapse" in MLLMs when performing spatial reasoning tasks using Grounded-Relative Policy Optimization (GRPO) reinforcement learning. Proposed Reinforce with Spatial Reward (RSR), a novel method integrating spatially-grounded rewards and carefully designed prompts to maintain robust reasoning. Developed "Super-Ref," a new benchmark specifically designed to rigorously evaluate spatial reasoning abilities, demonstrating substantial improvements over baseline methods.

Proposed a novel data-centric framework, Degree-Centric Structure Adaptation (DCSA), to improve Graph Neural Network (GNN) performance on nodes with various degrees. Introduced supervision-guided attachment for isolated nodes and influence-directed detachment for high-degree super nodes, optimizing graph structures without altering model architectures or training parameters. Demonstrated through comprehensive analyses and experiments that this approach significantly enhances classification performance on both synthetic and real-world datasets.