Unexplained Tool-Use Spike in claude-3.6-sonnet

Investigate and explain the cause of the anomalous spike in bail-tool usage observed for the model claude-3.6-sonnet during the reported neutral-prompts and BailBench experiments with the bail tool method.

Background

When comparing bail behavior across models on both neutral prompts and BailBench, the authors note an unexpected large increase in tool use for claude-3.6-sonnet under the bail tool method.

This spike deviates from other models’ behavior and remains unexplained, suggesting a model- or setup-specific factor affecting tool invocation that has not yet been identified.

References

The spike in claude-3.6-sonnet tool use is odd and still unexplained.

— The LLM Has Left The Chat: Evidence of Bail Preferences in Large Language Models (2509.04781 - Ensign et al., 5 Sep 2025) in Appendix O.2: Bails Georg: Models that have high bail rates on all prompts