We identify a monotonic entropy descent scenario in the block decoding for dLLMs's correct generation. We further develop a RL post-training method to achieve this descent with a dynamic size block mechanism.

We introduce TRN-R1-Zero, a RL-based LLM post-training method for effective reasoning over text-rich networks, such as citation, hyperlink, social and co-purchase domains.

We introduce a new dataset on block size for dLLMs and a new RL post-training frameworks for dLLMs.

We formally define future-aware and no-future criteria for self-contained blocks and a novel variable-size self-contained blocks methods for dLLMs.

We propose TIDE, an information decomposition framework that separates feature, structure, and joint signals in graph neural networks, using an information bottleneck objective to improve OOD detection by filtering spurious information.

We propose GFMate, a pre-training-agnostic test-time prompt tuning framework that uses centroid and layer prompts to adapt GFMs by leveraging both labelled and unlabelled target data.

We introduce TextTopoOOD, a framework for modeling diverse OOD scenarios on text-rich networks, and propose TNT-OOD, a novel detection method that captures the intricate interplay between text and topology.

We propose the GOLD framework for graph OOD detection, an implicit adversarial learning pipeline with synthetic OOD exposure without pre-trained models.

We propose the GrapHoST framework for graph learning to conduct test-time graph transformation based on homophily to enhance the robustness of graph models.

We introduce a benchmark for graph condensation with a thorough methodology development method and an extensive evaluation protocol.

We extend the Condense-and-Train (CaT) continual graph learning algorithm with a more efficient and effective, psudo-label guided memory bank (PUMA🐆) framework.

We introduce a Condense-and-Train (CaT🐱) memory-based continual graph learning algorithm using graph condensation to construct a more representative memory bank. And a Train-in-Memory continual learning scheme can further alleviate the imbalanced training issue in Class Incremental Learning.

We develop a distillation strategy for case-to-case graph-based method (see our CaseLink below) acceleration for legal case retrieval.

We introduce an inductive graph learning paradigm for legal case retrieval to tackle the challenge of unseen testing query and candidate cases.

We introduce an extended CaseGNN++ method with graph augmentations based on the CaseGNN framework.

We introduce a structural modelling of law case for effective retrieval with the aid of summarisation from LLM and grpah neural networks.