EVA: Efficient Video Agent with RL — Access Video AI Capabilities via NexaAPI

EVA: Efficient Video Agent with RL — Access Video AI Capabilities via NexaAPI A new paper from SenseTime Research just landed on HuggingFace: EVA (Efficient Reinforcement Learning for End-to-End Video Agent) ( arXiv 2603.22918 ). This research introduces a novel approach to video understanding that could reshape how AI processes long videos. What is EVA? EVA tackles a fundamental challenge in AI video understanding: long token sequences with extensive temporal dependencies and redundant frames . Traditional approaches process entire videos or uniformly sampled frames — EVA does something smarter. Key innovations: Planning-before-perception : EVA decides what to watch, when to watch, and how to watch Iterative reasoning : summary → plan → action → reflection loop Three-stage training : SFT → KTO (Kahneman-Tversky Optimization) → GRPO 6-12% improvement over general MLLM baselines on 6 video benchmarks 1-3% gain over prior adaptive agent methods The code and model are available at github.