Context-Driven Incremental Compression for Multi-Turn Dialogue Generation
English summary
The paper proposes Context-Driven Incremental Compression (C-DIC), which structures dialogue history as interleaved contextual threads and maintains revisable per-thread compression states in a compact dialogue memory. At each turn, a retrieve-revise-write-back loop shares information across turns and updates stale memories. It also adapts truncated backpropagation-through-time (TBPTT) to learn cross-turn dependencies without full-history backpropagation. Experiments on long-form dialogue benchmarks show C-DIC achieves stable inference latency and perplexity over hundreds of turns, outperforming existing context compression methods.
Chinese summary
本文提出上下文驱动增量压缩方法(C-DIC)，将对话历史组织为交错上下文线程，并在紧凑对话记忆中维护可修正的每线程压缩状态。每轮通过检索-修正-回写循环跨轮共享信息并更新过时记忆。同时适配截断时间反向传播(TBPTT)以学习跨轮依赖，无需全历史反向传播。在长对话基准测试中，C-DIC在数百轮对话上保持稳定的推理延迟和困惑度，优于现有上下文压缩方法。
Key points
Multi-turn dialogue history grows rapidly, causing redundant attention and encoding costs; naive truncation or summarization harms fidelity, and existing compressors lack cross-turn memory sharing.
多轮对话历史快速增长导致冗余的注意力和编码成本；简单截断或摘要损害保真度，现有压缩器缺乏跨轮记忆共享。
C-DIC models conversations as interleaved threads with per-thread compression states, enabling a lightweight retrieve-revise-write-back loop that updates a compact dialogue memory each turn.
C-DIC将会话建模为交错线程并维护每线程压缩状态，通过轻量级检索-修正-回写循环在每轮更新紧凑对话记忆。
Training uses an adapted TBPTT that learns cross-turn dependencies without backpropagating through the entire history, making long-context training feasible.
训练采用适配的截断时间反向传播，学习跨轮依赖而无需对整个历史反向传播，使长上下文训练可行。
On long-form dialogue benchmarks, C-DIC maintains stable inference latency and perplexity over hundreds of turns, outperforming baselines in both efficiency and quality.
在长对话基准上，C-DIC在数百轮内保持稳定的推理延迟和困惑度，在效率和质量上均优于基线方法。