Guglielmo Camporese (gool-yell-moe)

Audio emotion steering in Stable Audio 3: a mood direction lives in the DiT residual stream, linearly decodable at the middle layers, and shifts generated audio from grief to euphoria across genres — same prompt, same weights, only α changes.

The emotion map inside LLMs is real, geometric, and causal — sharply localised at certain layers. Positive and negative affect have asymmetric causal profiles.