DZ-TDPO: Non-Destructive Temporal Alignment for Mutable State Tracking in Long-Context Dialogue
This addresses a key bottleneck for long-context dialogue systems by enabling mutable state tracking without degrading general capabilities, though it appears incremental as it builds on existing alignment and attention methods.
The paper tackled the problem of State Inertia in long-context dialogue systems, where static constraints hinder conflict resolution between evolving user intents and historical context, and proposed DZ-TDPO, a non-destructive alignment framework that achieved state-of-the-art win rates (e.g., 86.2% on Phi-3.5 and 99.4% on Qwen2.5-7B) with minimal perplexity overhead.
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 learnable temporal attention bias. Experiments on the Multi-Session Chat (MSC) dataset demonstrate that DZ-TDPO achieves state-of-the-art win rates (86.2% on Phi-3.5) while maintaining robust zero-shot generalization. Crucially, 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 near-perfect alignment (99.4% 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