Empathetic response generation must decide not only what to say but also how to react to the interlocutor's affective state. We formalize this as response‑side affective‑orientation control and employ multi‑annotator emoji distributions as weak affective evidence rather than as output symbols or gold labels, thereby inducing a latent control space that approximates the listener's stance. Building on EmpatheticDialogues, we create EmojiDialogue, an utterance‑level extension enriched with emoji votes and confidence scores. We then propose EmoStance: it first models source‑side affective expression, predicts a soft response‑side orientation vector from dialogue context and speaker roles, and finally steers a frozen instruction‑tuned large language model (LLM) via continuous prefix embeddings.
In a blind pairwise evaluation with 20 annotators and 800 judgments, EmoStance achieved a decisive win rate of 62.2%, showing the clearest gains in contextual specificity and perceived responsiveness, while remaining complementary to external‑knowledge augmentation methods. The code, annotation metadata, and reconstruction scripts are released on GitHub: https://github.com/18277390221/EmoStance.
Review: EmoStance demonstrates that weak supervision from emojis can provide fine‑grained affective‑orientation control, offering a practical route to generate empathetic responses that balance naturalness and emotional alignment.