Behavior of CBPO under offline distribution shifts

Characterize the behavior of Cross-Behavior Policy Optimization under offline distribution shifts across domains or model families.

Background

Cross-Behavior Policy Optimization relies on offline demonstrations, reference responses, and replayed rollouts that may differ from the distributions encountered during optimization or evaluation. The paper does not assess how the method behaves when offline data distributions shift across domains or model families, leaving its robustness under such shifts unresolved.

References

CBPO also incurs an upfront generation cost to estimate the frozen prompt normalizer, and its behavior under offline distribution shifts across domains or model families remains untested.

— Cross-Behavior Policy Optimization: Unified Learning from Offline Demonstrations and On- and Off-Policy Rollouts via a Shared Regression Target  (2610.12830 - Cheng et al., 9 Oct 2026) in Section Limitations