ForesightKV: Optimizing KV Cache Eviction for Reasoning Models by Learning Long-Term Contribution
This addresses efficiency issues for users of reasoning models by optimizing KV cache management, though it is incremental as it builds on existing eviction methods with a novel learning approach.
The paper tackles the problem of high memory and computation costs from linearly expanding KV cache in large language models during long reasoning tasks by introducing ForesightKV, a training-based eviction framework that learns to predict which KV pairs to evict, resulting in consistent outperformance of prior methods under only half the cache budget on benchmarks like AIME2024 and AIME2025.
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 KV 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 of three reasoning models demonstrate that ForesightKV consistently outperforms prior methods under only half the cache budget, while benefiting synergistically from both supervised and reinforcement learning approaches.