Learning to Understand and Correct: Evolving Curricular Training for Long-Context Dialogue State Tracking

ACL ARR 2026 January Submission9241 Authors

Keywords: dialogue state tracking

Abstract:

Dialogue State Tracking (DST) is a core component of task-oriented dialogue systems, responsible for maintaining structured user goals across dialogue turns.
However, existing DST models suffer notable performance degradation in long multi-turn conversations due to limited long-context understanding and error propagation from earlier predictions. Through empirical analysis, we characterize these challenges and show how they hinder DST performance in long-context settings.
To address them, we propose \textbf{\model}, a novel recursive modeling framework that enhances long-context tracking and self-correction. \model conditions each prediction on the previously predicted dialogue state, alleviating the burden of long dialogue histories. It further incorporates two complementary training strategies: \emph{Progressive State Exposure}, which mitigates error propagation by gradually replacing ground-truth states with model predictions during training, and a \emph{Traceability-Aware Curriculum Learning}, which facilitates state revision through staged learning.
Extensive experiments on two benchmark datasets across multiple backbones demonstrate that \model~consistently outperforms strong baselines, particularly in long-context and cross-domain scenarios.

Paper Type: Long

Research Area: Dialogue and Interactive Systems

Research Area Keywords: dialogue state tracking

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 9241