ForesightKV: Optimizing KV Cache Eviction for Reasoning Models by Learning Long-Term Contribution

Zican Dong, Peiyu Liu, Junyi Li, Zhipeng Chen, Han Peng, Shuo Wang, Xin Zhao

Keywords: KV cache eviction, long-context reasoning, reinforcement learning

TL;DR: We propose ForesightKV, a training-based KV cache eviction method for long-context reasoning tasks, which consists of supervised training and reinforcment learning.

Abstract: Recently, large language models (LLMs) have shown remarkable reasoning abilities by producing long reasoning traces. However, as the sequence length grows, the key-value (KV) cache expands linearly, incurring significant memory and computation costs. Existing cache eviction methods mitigate this issue by discarding less important KV pairs, but often fail to capture complex KV dependencies, resulting in performance degradation. To better balance efficiency and performance, we introduce ForesightKV, a training-based KV cache eviction framework that learns to predict which KV pairs to evict during long-text generations. We first design the Golden Eviction algorithm, which identifies the optimal eviction KV pairs at each step using future attention scores. These traces and the scores at each step are then distilled via supervised training with a Pairwise Ranking Loss. Furthermore, we formulate cache eviction as a Markov Decision Process and apply the GRPO algorithm to mitigate the significant language modeling loss increase on low-entropy tokens. Experiments on AIME2024 and AIME2025 benchmarks with Qwen3-1.7B and Qwen3-4B demonstrate that ForesightKV consistently outperforms prior methods under only half the cache budget, while benefiting synergistically from both supervised and reinforcement learning approaches.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 4237