Emotion Flip Reasoning via Stacked Instruction Finetuning of LLMs

To find what caused an emotion flip, the model has to answer a few smaller questions first: did the emotion change, whose emotion changed, and which earlier utterance caused it?

If everything is predicted at once, those steps are hidden inside one answer. When the model gets the final utterance wrong, it is hard to tell where the reasoning failed.

I separate the reasoning into steps.

With stacked instruction finetuning, the model learns to identify the flip, track whose emotion changed, and then find the utterance that caused it.