Focusing Slot: Inference-Time Slot Steering for Zero-Shot Dialogue State Tracking

Zero-shot dialogue state tracking (DST) aims to generalize to unseen domains without retraining or target-domain annotations. However, existing approaches that either append slot- or schema-level descriptions or add lightweight adapters, face two inference-time limitations that cause negative transfer: (i) domain bias, where shared slot names across domains bias the model toward the source domain; and (ii) slot-understanding bias, where target-only slots fail to adequately capture their relationships with and distinctions from existing slots, leading the model to mistakenly reuse source domain slot patterns. To address this, we present StDST, a training-free and plug-and-play activation steering method. Our approach constructs steering vectors for each slot using source-domain slot descriptions. Specifically, we combine a positive prototype and a negative prototype to form the vector for shared slots, and employ a global negative vector for target-only slots. During inference, we inject a small vector offset along these vectors to explicitly align slot semantics, which mitigates domain bias and slot-understanding bias, thereby reducing negative transfer. Through extensive experiments on the MultiWOZ and SGD datasets, StDST demonstrates the effectiveness across various domains, and the consistent gains achieved when pairing it with diverse architectures further attest to its utility.