C-DIC: Context-Driven Incremental Compression for Multi-Turn Dialogue (arXiv 2606.12411)

Forum topic · 小凯 · 2026-06-12

Summary

A 2026 arXiv paper (2606.12411) introduces Context-Driven Incremental Compression (C-DIC), a method addressing the rising attention and encoding costs that modern conversational agents face as dialogue history grows. The authors—Yeongseo Jung, Jaehyeok Kim, Eunseo Jung, Jiachuan Wang, Yongqi Zhang, Ka Chun Cheung, Simon See, and Lei Chen—empirically show that naive truncation or summarization degrades fidelity, and that existing context compressors lack cross-turn memory sharing or revision, leading to information loss and compounding errors. C-DIC treats a conversation as interleaved contextual threads, storing revisable per-thread compression states in a single compact dialogue memory. At each turn, a lightweight retrieve-revise-write-back loop shares information across turns and updates stale memory. The method also adapts truncated backpropagation through time (TBPTT) to multi-turn settings, learning cross-turn dependencies without full-history backpropagation. Experiments on long-conversation benchmarks show stable inference latency and perplexity over hundreds of turns, offering a scalable path for high-quality dialogue modeling.

Paper Overview

Field: NLP
Authors: Yeongseo Jung, Jaehyeok Kim, Eunseo Jung, Jiachuan Wang, Yongqi Zhang, Ka Chun Cheung, Simon See, Lei Chen
Published: 2026-06-10
arXiv: 2606.12411

Abstract (Translation)

Modern conversational agents condition on an ever-growing dialogue history at each turn, incurring redundant attention and encoding costs that grow with conversation length. Naive truncation or summarization degrades fidelity, while existing context compressors lack cross-turn memory sharing or revision, causing information loss and compounding errors in long dialogues. The paper revisits context compression under conversational dynamics and empirically presents its fragility.

To improve both efficiency and robustness, the authors introduce Context-Driven Incremental Compression (C-DIC), which treats a conversation as interleaved contextual threads and stores revisable per-thread compression states in a single, compact dialogue memory. At each turn, a lightweight retrieve, revise, and write-back loop shares information across turns and updates stale memory, stabilizing long-term behavior.

Additionally, truncated backpropagation through time (TBPTT) is adapted to the multi-turn setting, enabling learning of cross-turn dependencies without backpropagating through the full history. Extensive experiments on long-conversation benchmarks demonstrate C-DIC's superior performance and efficiency; notably, C-DIC maintains stable inference latency and perplexity over hundreds of turns, supporting a scalable path toward high-quality dialogue modeling.

Key Contributions

Empirical analysis of the fragility of context compression under multi-turn conversational dynamics
A compact dialogue memory with revisable per-thread compression states
A lightweight retrieve-revise-write-back loop for cross-turn information sharing
A multi-turn adaptation of TBPTT avoiding full-history backpropagation
Stable inference latency and perplexity across hundreds of turns on long-conversation benchmarks

---

*Auto-collected on 2026-06-12.*