Imagine training a video AI to understand motion and context without forcing it to waste time thinking out loud. Researchers have unveiled a new method called OraRL that cuts training costs in half while significantly boosting performance. This breakthrough challenges the assumption that complex reasoning steps are necessary for high-quality video understanding, offering a faster, leaner path to smarter multimodal models.

The team focused on reinforcement learning for video multimodal large language models. Traditionally, these models struggle with sample efficiency because generating high-quality reasoning chains is expensive and slow. OraRL flips the script by treating existing annotations not just as scores, but as direct positive examples, or oracles, for the model to learn from. This approach integrates known good answers directly into the training loop, allowing the model to optimize against proven successes rather than guessing blindly.

However, simply adding these perfect examples creates a mathematical trap known as advantage inversion. When a high-quality oracle is included, it can artificially raise the baseline, causing the model to mistakenly penalize its own good work. To solve this, the researchers designed a decoupled advantage estimator. This system separates the baseline from the oracle’s influence, ensuring the model learns the right lessons without getting confused by conflicting signals.

The results are striking. OraRL requires only 2.2 times the step time of standard supervised fine-tuning, compared to nearly five times for existing methods. It scales effectively from small to large models and delivers superior accuracy on video tasks. Most notably, the model achieves state-of-the-art results without using chain-of-thought reasoning, decoding in just 130 milliseconds instead of nearly five seconds. This proves that efficient, high-performance video AI is possible without the heavy computational burden of complex reasoning steps.