Bootstrap Wayfinding Questions to Elicit Emotion Shift Reasoning with Large Language Models

When someone’s emotion changes in a conversation, the important question is usually not “what emotion was each sentence?” but “what caused the change?”

The usual approach labels every turn first, then searches for the line that triggered the shift. That creates an unnecessary dependency: if one of those intermediate emotion labels is wrong, the final explanation can be wrong too.

Imagine two people are talking normally. Then one person says something hurtful, and the next reply turns angry. The useful signal is the boundary between those two moments. We already know where the change happened. Labelling every earlier turn as calm, neutral or happy does not help us find the cause, and a mistaken label can actually send the reasoning in the wrong direction.

The shift itself gives us the best place to start.

I use wayfinding questions to guide the model across that boundary.

Instead of asking it to classify every utterance first, the method asks targeted questions about what changed between the conversation before the shift and the conversation after it. That helps the model reason backwards from the observed change to the utterance that caused it.

Start from the change you can see, then reason towards its cause.