Developed a weighted Direct Preference Optimization approach that reduces semantic drift during automated prompt optimization, and established a theoretical bound on drift in preference-tuned prompt generators.
Improved CLIP similarity by 8–12% and human-preference scores by 5–9% over standard DPO across three benchmarks and three language-model backbones.
Sem-DPO is a weighted Direct Preference Optimization approach that reduces semantic drift during automated prompt optimization and establishes a theoretical bound on drift in preference-tuned prompt generators. Across three benchmarks and three language-model backbones, it improves CLIP similarity by 8–12% and human-preference scores by 5–9% over standard DPO, and is preferred in 82.2% of pairwise human evaluations.
@inproceedings{khan2025semdpo,title={Sem-DPO: Mitigating Semantic Inconsistency in Preference Optimization for Prompt Engineering},author={Mohamed, Anas and Khan, Azal Ahmad and Wang, Xinran and Khan, Ahmad Faraz and Ge, Shuwen and Khan, Saman Bahzad and Ahmad, Ayaan and Anwar, Ali},booktitle={Findings of the Association for Computational Linguistics: ACL 2026},year={2026},pages={23656--23674},doi={10.18653/v1/2026.findings-acl.1184},url={https://aclanthology.org/2026.findings-acl.1184/},}