Researchers have developed an online learning algorithm that significantly improves the data efficiency of reinforcement learning from human feedback. The algorithm updates reward and language models as choice data is received, allowing it to match the performance of offline training on much smaller datasets. For example, with large language models, the algorithm can achieve similar results using just 20,000 labels instead of 200,000, representing a 10x gain in data efficiency. The team claims that their approach could potentially lead to a 1,000x improvement in data efficiency. --- Why it matters: This matters because it has the potential to greatly reduce the amount of labeled data required for training AI models, making them more practical and efficient to develop. Source: https://arxiv.org/abs/2603.17378