FF-BPSN achieves SOTA in proactive dialogue path planning

Target-oriented proactive dialogue systems aim to steer conversations toward specific goals while providing helpful suggestions. A key challenge is planning a coherent dialogue path—a sequence of topics or actions that naturally lead to the target. Existing methods struggle with this novel problem, often relying on simple heuristics or single-direction planning that misses contextual cues.

To address this, researchers from (affiliations not specified) introduce the Forward-Focused Bidirectional Pseudo-Siamese Network (FF-BPSN). This architecture employs two identical transformer-based decoders: one plans the path forward (from start to target) and one backward (from target to start). A forward-focused module then integrates both directional plans, prioritizing forward information to produce the final path. This bidirectional approach captures both future constraints and past context, improving coherence and goal adherence. Evaluated on the DuRecDial and DuRecDial 2.0 benchmarks, FF-BPSN achieves state-of-the-art results in path planning accuracy, and when used to prompt large language models, it significantly improves response relevance and proactivity. The paper is accepted at ICASSP 2026.