Retrieval-Feedback-Driven Distillation and Preference Alignment for Efficient LLM-based Query Expansion

Researchers developed a framework to make large language model-based query expansion more efficient by distilling knowledge from powerful teacher models into compact student models. The approach uses retrieval feedback and preference alignment to maintain 97% of the original performance while dramatically reducing inference costs.

• →A new distillation framework transfers query expansion capabilities from large teacher models to smaller, more efficient student models.
• →The method uses retrieval-metric-driven strategy to automatically create training pairs based on nDCG@10 performance differences.
• →The distilled Qwen3-4B model achieves 97% of DeepSeek-685B's performance on TREC DL19 benchmark with much lower inference cost.
• →Direct Preference Optimization is applied to align model generation with retrieval objectives rather than relying on few-shot examples.
• →The approach demonstrates effectiveness across both English and Chinese retrieval tasks, showing cross-language applicability.