MetaX Weekly AI Paper Review Week 28 2025 -- LLM Memory Management and Parameter-Efficient Learning Through MemOS and LoRA Optimization, Practical Applications of Reinforcement Learning-Based Vision-Language Models and Diffusion Generation Systems: MemOS LLM Memory Operating System integrating text, activation, and parameter memory for unified management enabling long-term reasoning and personalization. SingLoRA simplifies LoRA through product of single matrix and its transpose, reducing parameters by 50% while achieving more stable and higher performance. Scaling RL to Long Videos: 52K long video QA dataset and MR-SP training infrastructure enables LongVILA-R1 to achieve Gemini-1.5-Pro level long video reasoning performance. T-LoRA applies updates of different intensity by diffusion time step, solving overfitting in single-image diffusion model customization. MLM vs CLM: Large-scale experiment with 30 models shows CLM to MLM sequential training is optimal under fixed budget. Additional papers covered: vision-language model advances combining visual perception with language reasoning for complex scene understanding; diffusion model applications for controllable image and video generation; and evaluation benchmarks for measuring AI progress on multimodal reasoning tasks.