CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences

Large language models (LLMs) excel at processing long sequences, boosting demand for key-value (KV) caching. While recent efforts to evict KV cache have alleviated the inference burden, they often fail to allocate resources rationally across layers with different attention patterns. In this paper, we introduce Cascading and Adaptive KV cache Eviction (CAKE), a novel approach that frames KV cache eviction as a "cake-slicing problem." CAKE assesses layer-specific preferences by considering attention dynamics in both spatial and temporal dimensions, allocates rational cache size for layers accordingly, and manages memory constraints in a cascading manner. This approach enables a global view of cache allocation, adaptively distributing resources across diverse attention mechanisms while maintaining memory budgets. CAKE also employs a new eviction indicator that considers the shifting importance of tokens over time, addressing limitations in existing methods that overlook temporal dynamics. Comprehensive experiments on LongBench and NeedleBench show that CAKE maintains model performance with only 3.2% of the KV cache and consistently outperforms current baselines across various models and memory constraints, particularly in low-memory settings. Additionally, CAKE achieves over 10x speedup in decoding latency compared to full cache when processing contexts of 128K tokens with FlashAttention-2. Our code is available at https://github.com/antgroup/cakekv.

Personalized Topic Selection Model for Topic-Grounded Dialogue

Recently, the topic-grounded dialogue (TGD) system has become increasingly popular as its powerful capability to actively guide users to accomplish specific tasks through topic-guided conversations. Most existing works utilize side information (\eg topics or personas) in isolation to enhance the topic selection ability. However, due to disregarding the noise within these auxiliary information sources and their mutual influence, current models tend to predict user-uninteresting and contextually irrelevant topics. To build user-engaging and coherent dialogue agent, we propose a \textbf{P}ersonalized topic s\textbf{E}lection model for \textbf{T}opic-grounded \textbf{D}ialogue, named \textbf{PETD}, which takes account of the interaction of side information to selectively aggregate such information for more accurately predicting subsequent topics. Specifically, we evaluate the correlation between global topics and personas and selectively incorporate the global topics aligned with user personas. Furthermore, we propose a contrastive learning based persona selector to filter out irrelevant personas under the constraint of lacking pertinent persona annotations. Throughout the selection and generation, diverse relevant side information is considered. Extensive experiments demonstrate that our proposed method can generate engaging and diverse responses, outperforming state-of-the-art baselines across various evaluation metrics.

Position Debiasing Fine-Tuning for Causal Perception in Long-Term Dialogue

The core of the dialogue system is to generate relevant, informative, and human-like responses based on extensive dialogue history. Recently, dialogue generation domain has seen mainstream adoption of large language models (LLMs), due to its powerful capability in generating utterances. However, there is a natural deficiency for such models, that is, inherent position bias, which may lead them to pay more attention to the nearby utterances instead of causally relevant ones, resulting in generating irrelevant and generic responses in long-term dialogue. To alleviate such problem, in this paper, we propose a novel method, named Causal Perception long-term Dialogue framework (CPD), which employs perturbation-based causal variable discovery method to extract casually relevant utterances from the dialogue history and enhances model causal perception during fine-tuning. Specifically, a local-position awareness method is proposed in CPD for inter-sentence position correlation elimination, which helps models extract causally relevant utterances based on perturbations. Then, a casual-perception fine-tuning strategy is also proposed, to enhance the capability of discovering the causal invariant factors, by differently perturbing causally relevant and non-casually relevant ones for response generation. Experimental results on two datasets prove that our proposed method can effectively alleviate the position bias for multiple LLMs and achieve significant progress compared with existing baselines.