Researchers have developed looped language models to improve the ability of AI systems to perform complex tasks that require multiple steps and interactions with external tools. In a study, they evaluated these models in compositional tool-calling settings, where models must coordinate API calls, maintain state, and preserve dependencies across interactions. The results show that looped language models generally outperform non-looped models in these settings, particularly when using adaptive inference to allocate computation only as needed. --- Why it matters: This research matters because it could lead to more effective AI systems that can perform complex tasks with multiple steps and tool interactions, such as planning, coordination, and execution of workflows. This has implications for the development of agentic systems that require reliable planning and execution of tasks. Source: https://arxiv.org/abs/2608.18171