SpiritLM

Multimodal AI models show partial carryover of factual recall from text to speech

Researchers have investigated how multimodal language models handle factual recall across text and speech modalities. Using causal mediation analysis on the SpiritLM model, they found that the mechanisms for factual rec…