KV Cache Steering 作为无梯度潜空间注入口:Mind 到 Face 的物理控制通道
Evaluation modality
Spec-levelA spec-motivation / governance borrow. Evaluated by spec review + contract tests, not A/B or ablation.
- Primary owner
- —
- Phase-A verdict
- —
- Shadow profile
- —
- Source papers
- KV Cache Steering 2025
- Specs
- docs/specs/temporal-abstraction.mddocs/specs/expression-layer.md
Blind spot (现状盲点)
COG-8 提到把 Mechanistic Interpretability 作为只读的健康监控,并警告不要直接 steering 绕过 owner。但我们需要一条安全、有界且不破坏 R4(不修改 Token 表达)的物理通道,让 Metacontroller(Mind)向 Expression Layer(Face)传递 $z_t$ 指令。
Adoptable suggestions (可落地动作)
- 1.在 [`docs/specs/temporal-abstraction.md`](../specs/temporal-abstraction.md) 和 [`docs/specs/expression-layer.md`](../specs/expression-layer.md) 中,将 KV Cache Steering 定义为 substrate residual 之外的"第二条有界注入口"。PROPOSED
Not a runnable A/B candidate — evaluated by the path above, not ablation.
- 2.在**完全冻结基底(R2)**的情况下,通过干预特定层的 KV Cache 来注入控制信号,而不是通过 prompt 拼接。PROPOSED
Not a runnable A/B candidate — evaluated by the path above, not ablation.
- 3.确保该注入通道是单向的(Mind -> Face),且不传递梯度,严格遵守 Mind/Face 隔离(OA-2)。PROPOSED
Not a runnable A/B candidate — evaluated by the path above, not ablation.
Traceability
No plugins / runs linked yet. Scaffold a suggestion to start.
Expected benefit (预期收益)
- 提供一种比 prompt 拼接更底层、更稳定、更抗注入的控制指令传递方式。 - 完美契合 R2(冻结基底)和 R4(内部控制不在 token 空间)的设计哲学。
Cited paper (引用论文)
**C2-08. KV Cache Steering for Frozen LLMs. arXiv:2507.08799, 2025.** - 关键观点:one-shot 的 KV 干预触发 CoT,提供了一种稳定的"无梯度控制器"通道。 ---