Looped Language Models Enhance Complex Tool Use in AI Agents.

This paper investigates the application of looped language models (LLMs) to enhance the capabilities of AI agents in complex tool-calling scenarios. Unlike traditional models, looped LLMs can perform recurrent computations, allowing them to better coordinate multiple API calls, manage intermediate states, and track dependencies across various tool interactions. The study evaluated these models on several benchmarks, demonstrating that recurrent processing significantly improves performance in compositional and dependency-aware tool use. The findings suggest that increasing recurrent depth generally boosts accuracy for multi-step tasks. However, the research also highlights that adaptive inference, which allocates computational resources only when necessary, provides a more efficient balance between performance and computational cost. This indicates a promising architectural direction for developing more reliable and sophisticated agentic systems capable of intricate planning and execution of tool-based workflows.