PolyAlign: Conditional Human-Distribution Alignment

PolyAlign is a newly introduced framework for conditional human-distribution alignment in language models, addressing the limitations of traditional supervised fine-tuning (SFT) by aligning models to context-specific human response distributions rather than a single global behavior. It employs a combination of Bucket-Aware SFT and Human-Distribution Preference Optimization (HDPO) to optimize performance across varied interaction contexts, as demonstrated in a bilingual evaluation suite involving English and Chinese. This approach enhances conditional naturalness and distributional faithfulness, indicating a shift towards interaction-aware alignment strategies that better reflect human variability in responses.