Looped Language Models for Compositional Tool Calling

Research shows that recurrent computation in looped language models improves performance in multi-step, dependency-aware tool calling. While gains are smaller for isolated API calls, recurrent depth significantly benefits complex, compositional tool-use scenarios.