AKV lets an agent drop and reposition its own KV cache instead of rewriting its history
AspireAI community · arXiv
As agent trajectories grow, key-value cache I/O and storage become the dominant cost, and they can exceed the model context limit. Prior work shrinks cache by trimming or summarizing history, but editing the trajectory prefix invalidates every later cached state and triggers expensive re-prefilling of the retained context. The paper proposes cache engineering instead: agents proactively use drop to exclude selected cache from future attention and repos to adjust the position encodings of retained keys, without modifying the trajectory or re-prefilling. Because identical token sequences can yield different cache states under different operation histories, the authors co-design the AKV serving engine, which exposes a stateless interface compatible with existing serving APIs while managing cache reuse, eviction and restoration internally. Against conventional threshold-triggered summarization, a tool-response rolling-drop policy raises average benchmark scores by 1.7 points and cuts average cost by 27.7%, with up to 1.2x output throughput and 1.5x faster P95 request latency, and recursively self-improving the policy reduces cost a further 7.1% over expert design. The code is open source and builds on SGLang; the throughput and latency figures come from their own serving stack on NVIDIA GPUs.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.