AgentKV: Phase-Aware KV Eviction for Agentic LLMs¶

L1 core mechanism diagnostic. Phase-aware scoring runs. No SGLang changes or CUDA kernels were run.

Paper information¶

Field	Details
Paper	AgentKV: Phase-Aware KV Eviction for Agentic LLMs
Institution	University of Cambridge (first author affiliation)
First public release	2026-09-14 (arXiv v1)
Original source code	Yes: LiuTaowen-Tony/agentkv
Adapter/method	`agentkv`
Local reproduction code	`src/auto_research/foundation_latest_20260916.py`

Original paper summary¶

Background and main changes¶

Keep a representative query buffer for each agent phase, and use relevance across phase queries to decide KV eviction.

The local reference kernel retains defining gates, weights or state transitions and exposes auditable intermediates for comparison under a shared budget.

flowchart LR
  I[Public inputs and state] --> M[AgentKV core mechanism]
  M --> A[Auditable intermediates]
  A --> O[Output or policy update]

Key figure from the paper¶

Figure 1: Overall method and components. From the paper; copyright belongs to the authors. Click for the source.

Core mechanism¶

The implementation retains defining state transitions and emits relevant weight, gate, depth, stage or cache statistics. A fixed budget and three seeds check mechanism invariants, not paper-scale training.

Offline and online results¶

Capability, speed and cost claims are attributed to the paper only. Local metrics are not directly compared or extrapolated to online gains.

Local reproduction¶

See three-seed results. The unified receipt marks diagnostic_only=true, so the public dashboard does not treat this as a full capability gain.

Scope¶

Phase-aware scoring runs. SGLang was not modified and no CUDA kernels were run.