Establish CIPR’s scalability and long-horizon effectiveness

Establish the effect of Cross-Interface Policy Regularization on models larger than 9 billion parameters and on long multi-turn tasks.

Background

Cross-Interface Policy Regularization (CIPR) is evaluated in experiments using models of up to 9 billion parameters and predominantly single-turn function-calling tasks. The paper therefore does not resolve whether its reported robustness benefits persist as model scale increases or when agents must maintain consistency across extended multi-turn interactions. The authors explicitly identify both dimensions as requiring further empirical establishment.

References

Our experiments use models of up to $9$B parameters, one training run per configuration, and mostly single-turn function calling, so the effect at larger scale and in long multi-turn tasks remains to be established.

— One Task, Many Interfaces: Simple Cross-Interface Consistency Training Improves LLM Agent Robustness  (2610.20576 - Tajdini et al., 14 Oct 2026) in Limitations paragraph, Section 6 (Conclusion)