Looped language models, which apply recurrent computation within a single model, have already shown promise on reasoning benchmarks. Now, a new study from Hugging Face explores their potential for agentic tool use—a domain that remains largely uncharted.
The research focuses on compositional tool-calling scenarios, where models must coordinate multiple API calls, maintain intermediate state, and preserve dependencies across interactions. The team evaluated both native and retrofitted looped models on three benchmarks: API-Bank, BFCL, and NESTful.
In controlled experiments, recurrent computation generally benefited compositional and dependency-aware tool use, while gains on isolated API invocation were smaller and more model-dependent. Accuracy on multi-step tool use typically increased with recurrent depth, but adaptive inference—which allocates extra computation only when needed—delivered a more favorable compute-performance trade-off.
The findings suggest that looped language models are a promising architecture for agentic systems requiring reliable planning, coordination, and execution of complex tool workflows. The paper also lists several related works, including looped state-space models and latent reasoning approaches.