Internal Coherence Maximization for Unsupervised Persona-Conditioned Value Specification in Pluralistic Alignment

Aligning AI systems with human values requires more than broad principles alone. In pluralistic settings, different groups may reasonably prefer different answers, so models often need concrete examples that show how values apply in practice. This thesis studies whether such group-specific examples can be constructed automatically and then reused for downstream prediction. It uses Internal Coherence Maximization (ICM), an unsupervised method that infers labels for logically related claims and refines them to improve internal coherence while preserving support from model scores. On OpinionQA, the few-shot condition built from ICM-inferred labels reaches 81.93% accuracy, slightly above the gold-label few-shot reference at 81.03% and well above zero-shot baselines. The same pattern transfers to Persona-Tailoring, where the ICM-label few-shot condition reaches 77.37% and remains close to the gold-label reference at 77.97%. Perturbation and stability analyses further indicate that coherence improves generalization and logical stability. These results suggest that coherent unsupervised examples can serve as practical value-specification artifacts. A separate local OpinionQA intervention analysis also suggests that Human-Guided Correction may support bounded post hoc revision, though it is evaluated on a different local slice from the thesis’s main aggregate benchmark.

Author

• Pei, Yiding

Language

• en

Date created

• 2026

Type of Work

• Master's Thesis

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Computer Science

Degree

• M.S.

Advisor

• Feng, Shi

Committee Member(s)

• Feng, Shi
• Youssef, Abdou
• Zirikly, Ayah

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/ff3656491