Researchers introduce COSPLAY, an AI framework that lets language models learn and reuse skills to complete multi-step tasks—solving a core weakness in AI decision-making for long, complex jobs

Researchers at multiple institutions published COSPLAY, a framework pairing an LLM (AI text-understanding system) decision agent with a learnable skill bank that discovers reusable strategies from trial-and-error attempts. The LLM learns which stored skills to retrieve and chain together across many steps—a capability that existing LLMs struggle with because they typically forget patterns between separate tasks.

Unlike standard LLMs that start fresh on each problem, COSPLAY lets the skill bank extract and store generalizable moves from unlabeled experience, so the decision agent can pull proven strategies instead of reinventing solutions. In long-horizon games (tasks with 50+ decision steps), this means the AI can handle delayed rewards and incomplete information—conditions where current LLMs often fail or stall.

Game developers and robotics engineers building AI agents for complex environments now have a concrete path to make those agents learn faster and perform reliably over dozens of steps. Anyone deploying LLMs for multi-step reasoning (customer support workflows, autonomous task planning, game AI) can test whether COSPLAY's skill-reuse approach cuts training time and improves consistency compared to training from scratch each time.