Video understanding has long been a bottleneck for multimodal models because processing every single frame of a lengthy clip is computationally expensive and often redundant. Most current systems act as passive observers, scanning through data without a strategic approach to finding information.

The new EVA (Efficient Video Agent) framework changes this by introducing a "planning-before-perception" mindset. Instead of blindly watching entire sequences, the agent uses an iterative cycle of summarizing, planning, acting, and reflecting to decide exactly which moments in a video deserve its attention. This approach mimics how humans might skim a long film to find a specific scene, drastically reducing the "visual budget," or the total amount of visual data the model needs to process at once.

To make this complex behavior possible, the researchers developed a three-stage learning pipeline. This starts with supervised fine-tuning to teach basic imitation, followed by advanced reinforcement learning techniques like Kahneman-Tversky Optimization (KTO) and Generalized Reward Policy Optimization (GRPO). These methods help bridge the gap between simple pattern matching and actual strategic reasoning.

In testing across six major benchmarks, EVA outperformed standard models by up to 12%. By autonomously deciding what, when, and how to watch, the system proves that being selective is often more effective than being exhaustive in the world of computer vision.

Video understanding has long been a bottleneck for multimodal models because processing every single frame of a lengthy clip is computationally expensive and often redundant. Most current systems act as passive observers, scanning through data without a strategic approach to finding information.

The new EVA (Efficient Video Agent) framework changes this by introducing a "planning-before-perception" mindset. Instead of blindly watching entire sequences, the agent uses an iterative cycle of summarizing, planning, acting, and reflecting to decide exactly which moments in a video deserve its attention. This approach mimics how humans might skim a long film to find a specific scene, drastically reducing the "visual budget," or the total amount of visual data the model needs to process at once.

To make this complex behavior possible, the researchers developed a three-stage learning pipeline. This starts with supervised fine-tuning to teach basic imitation, followed by advanced reinforcement learning techniques like Kahneman-Tversky Optimization (KTO) and Generalized Reward Policy Optimization (GRPO). These methods help bridge the gap between simple pattern matching and actual strategic reasoning.

In testing across six major benchmarks, EVA outperformed standard models by up to 12%. By autonomously deciding what, when, and how to watch, the system proves that being selective is often more effective than being exhaustive in the world of computer vision.