Can imaginary listeners reduce dialogue agent contradictions?

Persona-based dialogue agents routinely contradict their own stated attributes. Previous solutions either require Natural Language Inference (NLI) labels for training or attach extra trained modules. The "Will I Sound Like Me?" approach (2020) takes a different path: it endows existing agents with public self-consciousness at inference time through an imaginary listener, inspired by social cognition and pragmatics.

The mechanism uses the Rational Speech Acts (RSA) framework. Before generating an utterance, the agent simulates how a listener would interpret it — specifically, whether the listener could distinguish this speaker's persona from a distractor persona based on the utterance. Utterances that would not help identify the speaker (because they are generic or contradictory) are suppressed. The agent learns to ask: "Would I sound like me if I said this?"

The framework extends beyond persona to context consistency in general dialogue. The distractor selection — which alternative persona to contrast against — can be learned rather than manual or random.

This connects to Why does supervised learning fail to enforce persona consistency? as a complementary approach: offline RL punishes contradiction during training while RSA prevents it during inference. The RSA approach requires no additional training data but operates at generation time, adding computational cost per utterance. The trade-off is training-time correction vs. inference-time self-monitoring — both address the same root cause that generative models are never explicitly rewarded for consistency.

The deeper insight is that persona consistency is fundamentally a pragmatic property, not a semantic one. It is about how utterances function in identifying a speaker, not just about logical compatibility of facts.