Sailaja Rajanala

Fine-grained text-to-audio retrieval improves significantly when structured logical reasoning meets efficient cross-modal alignment.

Toxic audio retrieval can be effectively mitigated without sacrificing accuracy, thanks to a novel debiasing framework that preserves context and intent.

LLM-distilled procedural summaries, when combined with traditional dense retrieval, unlock significant gains in materials science question answering, outperforming paragraph-only retrieval.