Can AI truly learn and apply rules like humans do? This question lies at the heart of building more intelligent and adaptable AI systems. A new research paper explores this challenge, introducing RULEARN, a benchmark designed to test how well AI agents can learn rules in interactive environments. Imagine an AI placed in an escape room, tasked with deciphering a hidden code. Or imagine an AI needing to synthesize complex materials based on unknown chemical reactions. RULEARN presents AI agents with similar puzzles, requiring them to gather information, form hypotheses, and test their theories to uncover hidden rules. This benchmark consists of three types of puzzles—the “Function Operator,” the “Escape Room,” and the “Reactor”—each designed to assess different facets of rule learning. To help AI agents navigate these challenges, researchers developed IDEA, an innovative reasoning framework that combines induction, deduction, and abduction, mirroring human problem-solving strategies. An AI using IDEA starts by making an educated guess (abduction) based on initial observations. It then creates a plan (deduction) to test this hypothesis and interacts with the environment. Finally, it refines its initial guess (induction) based on the new information it gathers. This loop continues until the AI cracks the code, synthesizes the material, or solves the puzzle. Tests using five different large language models (LLMs) show that IDEA significantly improved their ability to learn rules, boosting their success rates by about 10%. However, even with IDEA, AI still lags behind humans. Humans, it turns out, are remarkably efficient at learning from their interactions, refining their hypotheses with fewer attempts and adapting their strategies more flexibly. One key finding is that current LLMs often repeat actions unnecessarily, even when the results are predictable. IDEA helps reduce this repetition by guiding the AI to choose more informative actions. Another challenge is that LLMs sometimes struggle to adjust their initial guesses even when faced with contradictory evidence, highlighting a limitation in their ability to learn from experience. The research behind RULEARN and IDEA offers crucial insights into how we can bridge the gap between human and AI rule learning. It emphasizes the importance of developing AI systems that can not only process information but also learn from interactions, adapt to new situations, and refine their strategies based on feedback. As AI continues to evolve, research like this will pave the way for more versatile and capable AI agents that can effectively navigate the complexities of the real world.