Improving Multi-turn Dialogue Consistency with Self-Recall Thinking

TL;DR

This paper introduces Self-Recall Thinking (SRT), a novel framework for enhancing multi-turn dialogue systems by improving long-range dependency tracking and response consistency without external memory modules.

Contribution

SRT enables models to selectively recall relevant historical turns and reason over context, improving dialogue coherence and efficiency compared to existing methods.

Findings

SRT improves F1 score by 4.7% over prior methods.

SRT reduces end-to-end latency by 14.7%.

SRT outperforms state-of-the-art baselines in multi-turn dialogue tasks.

Abstract

Large language model (LLM) based multi-turn dialogue systems often struggle to track dependencies across non-adjacent turns, undermining both consistency and scalability. As conversations lengthen, essential information becomes sparse and is buried in irrelevant context, while processing the entire dialogue history incurs severe efficiency bottlenecks. Existing solutions either rely on high latency external memory or lose fine-grained details through iterative summarization. In this paper, we propose Self-Recall Thinking (SRT), a framework designed to address long-range contextual dependency and sparse informative signals in multi-turn dialogue. SRT identifies helpful historical turns and uses them to generate contextually appropriate responses, enabling the model to selectively recall and reason over context during inference. This process yields an endogenous reasoning process that integrates interpretable recall steps without external modules. SRT incorporates: (1) Dependency Construction: Generating and converting it into self-recall chains; (2)Capability Initialization: Training to enable reasoning chains with recall tokens capability; (3)Reasoning Improvement: Refining accuracy via verifiable rewards to optimize recall and reasoning for correct answers. Experiments on multiple datasets demonstrate that SRT improves F1 score by 4.7% and reduces end-to-end latency by 14.7% over prior methods, achieving a balance between reasoning latency and accuracy, and outperforming state-of-the-art baselines.