Represented but Ignored: A Causal Account of Prosodic Underuse in Audio-Language Models
Source
arxiv.org
Author
Linkai Peng, Baorian Nuchged
Date
Context
An audio-LLMA large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.Full definition → can retain information about how something was said yet fail to use it in its answer. In their paper abstract, researchers report this pattern across four models built for audio understanding. Information about speech expression remained detectable inside later processing stages even when the final response only partly reflected it.
The researchers also changed selected internal states and observed answers shift in the predicted direction. That supports a role for the underused information in producing responses. The intervention did not selectively restore the correct answer, so the finding identifies a possible bottleneck in using speech cues rather than establishing a reliable fix.
Terms in this piece · Glossary
LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.