Latent chain-of-thought frequently has the answer before it starts
On logical tasks, predictions stay almost the same if the researchers remove the chain.
A study of 21 model-task configurations found that, on logical tasks, COCONUT frequently has an answer that is readable before the latent chain starts. Predictions stay almost the same if the researchers remove the chain. On arithmetic tasks, the answer changes in all training recipes in the test if the chain content changes.
Sources
Posted
