We propose Continual Knowledge Adaptation for Reinforcement Learning (CKA-RL), which enables the accumulation and effective utilization of historical knowledge.

We propose Pseudo-Zeroth-Order (PseuZO), a framework that estimates Jacobian matrices via model output differentiation and applies EMA for variance reduction, with theoretical convergence guarantees and empirical validation.

We adaptes Mamba2's structured mask to 2D scaning and integrates it into the self-attention mechanism of ViTs as an explicit positional encoding.

We propose CFPA for 3D part assembly that integrates semantic super-parts, dual-range feature propagation, and symmetry-aware supervision, achieving SOTA accuracy, consistency, and diversity on standard benchmarks.

We propose a novel online functional tensor decomposition (OFTD) framework for streaming data completion.

We discover commonalities across different language models and leverage them to achieve training-free cross-model adapter transfer.