StreamingDialogue: Prolonged Dialogue Learning via Long Context Compression with Minimal Losses

TL;DR

StreamingDialogue is introduced, which compresses long dialogue history into conv-attn sinks with minimal losses, and thus reduces computational complexity quadratically with the number of sinks, and has the potential to handle more than 200K of utterances, resulting in a prolonged dialogue learning.

Abstract

Standard Large Language Models (LLMs) struggle with handling dialogues with long contexts due to efficiency and consistency issues. According to our observation, dialogue contexts are highly structured, and the special token of \textit{End-of-Utterance} (EoU) in dialogues has the potential to aggregate information. We refer to the EoU tokens as ``conversational attention sinks'' (conv-attn sinks). Accordingly, we introduce StreamingDialogue, which compresses long dialogue history into conv-attn sinks with minimal losses, and thus reduces computational complexity quadratically with the number of sinks (i.e., the number of utterances). Current LLMs already demonstrate the ability to handle long context window, e.g., a window size of 200K or more. To this end, by compressing utterances into EoUs, our method has the potential to handle more than 200K of utterances, resulting in a prolonged dialogue learning. In order to minimize information losses from reconstruction after compression, we design two learning strategies of short-memory reconstruction (SMR) and long-memory reactivation (LMR). Our method outperforms strong baselines in dialogue tasks and achieves a 4 $\times$ speedup while reducing memory usage by 18 $\times$ compared to dense attention recomputation.

Figures (11)

• Figure 1: Attention map visualization. (a) Llama-2-7B/Chat with "</s>" and "\\n" as EoU ("</s>" counts as one token, "\\n" as two). (b) StreamingLLM versus StreamingDialogue attention on Llama-2-7B with "</s>" as EoU.
• Figure 2: StreamingDialogue framework. SMR & LMR strategies co-train the model by adjusting attention mechanisms. In supervised learning, the SMR & LMR-trained model is fine-tuned with dialogue datasets. During inference, only specific tokens are cached, with critical historical dialogue information in bold italics for clarity.
• Figure 3: Fluency, coherence, and consistency in human evaluations: ours vs StreamingLLM.
• Figure 4: Average perplexity and BLEU for StreamingLLM and StreamingDialogue on the MSC test set across varying utterance counts.
• Figure 5: Per-token latency and memory usage by method on MSC for varying input lengths, with memory reported as total minus fixed.