DZ-TDPO: Non-Destructive Temporal Alignment for Mutable State Tracking in Long-Context Dialogue

TL;DR

DZ-TDPO introduces a non-destructive alignment method for long-context dialogue systems, effectively resolving state inertia issues and achieving state-of-the-art performance without compromising model capabilities.

Contribution

It proposes a novel non-destructive temporal alignment framework combining conflict-aware constraints and attention bias, improving dialogue state tracking across model scales.

Findings

Achieves 55.4% win rate on MSC dataset with Phi-3.5.

Larger models like Qwen2.5-7B reach 50.8% win rate with minimal perplexity increase.

Demonstrates that attention regulation can mitigate historical inertia without destructive training.

Abstract

Long-context dialogue systems suffer from State Inertia, where static constraints prevent models from resolving conflicts between evolving user intents and established historical context. To address this, we propose DZ-TDPO, a non-destructive alignment framework that synergizes conflict-aware dynamic KL constraints with a calibrated temporal attention bias. Experiments on the Multi-Session Chat (MSC) dataset demonstrate that DZ-TDPO achieves state-of-the-art win rates (55.4% on Phi-3.5) while maintaining robust zero-shot generalization. Our scaling analysis reveals a "Capacity-Stability Trade-off": while smaller models incur an "alignment tax" (perplexity surge) to overcome historical inertia, the larger Qwen2.5-7B model achieves 50.8% win rate with negligible perplexity overhead. This confirms that TAI can be alleviated via precise attention regulation rather than destructive weight updates, preserving general capabilities (MMLU) across model scales. Code and data are available: https://github.com/lyj20071013/DZ-TDPO