Search papers, labs, and topics across Lattice.
2
0
5
Explicit reasoning steps ("thinking mode") boost spatial audio question answering accuracy by 5.1%, especially when combined with source separation.
By grounding LLMs in timestamped acoustic events instead of raw audio, LongAudio-RAG enables accurate question answering over multi-hour audio, outperforming standard RAG and text-to-SQL baselines.