New Framework COSPLAY Boosts Decision-Making in Long-Horizon Tasks

Published on April 24, 2026

Recent advancements in large language models (LLMs) offered significant promise for interactive environments, primarily through their ability to engage in multi-step reasoning. Traditionally, these models struggled with long-horizon tasks, often faltering in decision-making due to their inability to retain and apply structured skills consistently. However, a breakthrough has emerged with the introduction of COSPLAY, a co-evolving framework designed to enhance agent performance across various game environments.

COSPLAY addresses the core issues faced . It enables a decision agent to efficiently retrieve skills from a dynamically updated skill bank. Concurrently, the skill bank agent works to discover and refine reusable skills from the agent’s experiences, forming a robust system that improves both action generation and skill retrieval mechanisms.

Initial experiments demonstrate the effectiveness of this approach. In tests across six distinct gaming environments, COSPLAY achieved an average reward improvement of over 25.1 percent compared to established LLM baselines. This enhancement was particularly notable in single-player benchmarks, illustrating its superior skill management capabilities.

The implications are substantial for the future of AI in gaming and complex decision-making tasks. consistency of long-horizon decision-making, COSPLAY not only enhances gameplay but also sets the stage for more complex applications of AI in various fields. These advances could redefine how agents learn and adapt in real-time scenarios, unlocking new potentials for AI technologies.

New Framework COSPLAY Boosts Decision-Making in Long-Horizon Tasks

Related News

Related Articles