Causal impact of role-header and chat-template conventions on user-turn generation

Determine the extent to which differences in training-time role-header and chat-template conventions (e.g., whether a user-role header signals a new independent query versus a continuation) causally affect user-turn generation and measured interaction awareness across model families.

Background

The paper considers an alternative explanation for low genuine-followup rates: that models may have been trained with chat-template conventions which bias user-role headers toward starting new exchanges rather than following up on prior assistant turns.

Because the authors do not have access to the post-training data for the evaluated models, they cannot rule out this explanation, leaving the causal impact of such conventions on user-turn generation behavior unresolved.

References

Different model families may have been trained with different conventions for what follows a user-role header (a new independent query versus a continuation of the same conversation). We do not have access to the post-training data for any of the evaluated models, so we cannot rule this out entirely.

— Beyond the Assistant Turn: User Turn Generation as a Probe of Interaction Awareness in Language Models  (2604.02315 - Shekkizhar et al., 2 Apr 2026) in Appendix — Continued Discussion (Alternative explanations): Role-header semantics