AI Learns to Think Before Acting, Speeding Up Drug Discovery

Finding the right molecule for a new drug is like looking for a needle in a haystack — except the needle can be shaped in countless ways, and every test costs time and money. A new system called WMLLM, developed by researchers in China, dramatically speeds up this hunt by teaching AI to predict outcomes before running experiments.

WMLLM is an AI agent that uses what researchers call "predict-then-act" thinking. Instead of randomly guessing molecules and testing them, it first leans on knowledge learned from past scientific data to predict which directions look most promising. Only then does it generate candidates to test. This is similar to how an experienced chef decides what to add to a recipe by imagining how ingredients will blend before tasting — rather than throwing everything in and adjusting afterward.

What makes WMLLM special is that it improves itself. As it performs searches, it learns from each result and refines both its predictions and its strategy. Like a chess player who studies their own games to get better, WMLLM evolves its internal model of the problem through trial and feedback, guided by reinforcement learning and multi-turn reasoning. This self-evolving loop was applied to complex, multi-objective molecular optimization — where you have to balance several goals like potency and safety at once.

The results are promising: WMLLM found better molecules using fewer tests than existing methods, achieving state-of-the-art performance on a standard benchmark. While this is early research, the broader implication is that AI systems don't just have to process information — they can learn to think strategically. That could shorten the painfully slow path from lab experiment to life-saving medicine, potentially saving billions in drug development costs and getting treatments to patients sooner.