GEM: Graph-Enhanced Mixture-of-Experts for Dialogue State Tracking
A novel framework named GEM (Graph-Enhanced Mixture-of-Experts) has been introduced to enhance Dialogue State Tracking (DST) in multi-domain interactions. This system merges a Graph Neural Network, which effectively captures dialogue structure and turn-level relationships, with a fine-tuned T5-Small encoder-decoder for sequence modeling, all managed by an advanced router. To facilitate complex value generation, GEM incorporates ReAct agents that execute structured reasoning based on dialogue context. When evaluated on the MultiWOZ 2.2 benchmark, GEM achieves a Joint Goal Accuracy of 65.19%, significantly exceeding the performance of end-to-end LLM methods (which peak at 38.43%) and surpassing the previous leading results. The research can be found on arXiv under reference 2605.04449.
Key facts
- GEM stands for Graph-Enhanced Mixture-of-Experts
- Framework combines language models and graph-structured dialogue understanding
- Uses ReAct agent-based reasoning for DST
- Dynamic routing between Graph Neural Network and T5-Small
- Achieves 65.19% Joint Goal Accuracy on MultiWOZ 2.2
- Outperforms end-to-end LLM approaches (best 38.43%)
- Surpasses previous state-of-the-art
- Paper available on arXiv: 2605.04449
Entities
Institutions
- arXiv