KV cache 操作让预训练模型并发观察、思考和响应成为可能。机制已有实验,生产运行时仍在形成,状态治理将成为新的系统边界。
KV-cache manipulation lets pretrained LLMs observe, reason, and respond concurrently. The mechanism works; the production runtime is still emerging.