Generalization of the Alignment Tax to Closed-Source GPT-Class Models and Other Domains

Ascertain whether the alignment tax—response homogenization that degrades sampling-based uncertainty estimation—generalizes to closed-source GPT-class language models and to additional domains such as code and dialogue.

Background

The study’s empirical evidence covers open-source models between 3B and 14B parameters across several datasets, showing DPO-driven response homogenization and its impact on sampling-based uncertainty.

The authors explicitly note that generalization to closed-source GPT-class systems and other domains (e.g., code, dialogue) remains unconfirmed, making this a critical open question for broader applicability and deployment.

References

Generalization to closed-source GPT-class models and other domains (code, dialogue) unconfirmed.

— The Alignment Tax: Response Homogenization in Aligned LLMs and Its Implications for Uncertainty Estimation  (2603.24124 - Liu, 25 Mar 2026) in Limitations (4)

Our study has three main limitations. First, we evaluate four open-source code LLMs in the 7--8B range because our analyses require token probabilities and hidden representations and whether the findings extend to larger proprietary models remains open.

— Characterizing Overconfident Failure in LLM-Based Code Generation  (2610.11300 - Rathnasuriya et al., 8 Oct 2026) in Section 7, Limitations