The field of multimodal reasoning for long-horizon video understanding is moving towards more efficient and effective methods for fusing and aligning multiple modalities. Recent developments have focused on addressing the limitations of existing methods, such as data inefficiency and vanishing advantages, through the use of off-policy training architectures and novel credit assignment strategies. Additionally, there is a growing interest in using reinforcement learning to optimize temporal sampling policies and improve long-form video-language understanding. Noteworthy papers in this area include: AVATAR, which introduces a framework that addresses the limitations of existing methods through off-policy training and temporal advantage shaping. TSPO, which proposes a trainable event-aware temporal agent and a reinforcement learning paradigm to advance MLLMs' long-form video-language understanding. Thinking With Videos, which proposes a novel end-to-end agentic video reasoning framework that uses a visual toolbox to densely sample new video frames on demand. ReasoningTrack, which proposes a novel reasoning-based vision-language tracking framework that uses a pre-trained vision-language model and reinforcement learning to optimize reasoning and language generation.