Imported from AwareLiquid/liquid-research-skill (SKILL.md). Install upstream with npx skills add AwareLiquid/liquid-research-skill. Copyright stays with the author.

Liquid R&D Principles

Distilled decision framework from the model-R&D engineering review. Apply these principles when evaluating architecture work, approving research directions, or scoping the liquid-architecture roadmap. Full notes and the component-decomposition template: references/rd-principles.md.

The three consensus conclusions

1. Brain-like architecture = correct inductive bias. The value is not "imitating the brain" — it is injecting inductive biases that match cognitive / physical laws. This raises per-task generalization and some general-intelligence dimensions, and makes pure scale-up more efficient than the autoregressive baseline.
2. Autoregression is the ceiling. As long as the fitting target is "predict the next token," every network-layer or parameter-stacking optimization is only a quantitative gain. No qualitative leap is possible without changing this target.
3. Paradigm change has a very high bar. Diffusion proved the value of correct inductive bias, but migrating it to general / temporal intelligence requires a full training-methodology rebuild — a physics-mechanism-level challenge, not an architecture tweak.

Decision rules

• Judge a proposal by its fitting target, not its topology. Ask: "Does this change what the model is fitting (the distribution target), or only how it fits the same distribution?" Only the former can be paradigm-level; the latter is at best marginal scale-up.
• Reject vague research goals. Never approve 立项 phrased as "develop a brain-like architecture" or "optimize model performance." Require the componentized decomposition below.
• Inductive bias is the scale-up lever. Rank proposals by the strength and correctness of the inductive bias they introduce — this, not parameter count, predicts scale-up efficiency.
• Change the data form with the architecture. Data form is the fitting target — next-token text re-imports the autoregressive ceiling. When reviewing an architecture change, ask whether the supervision signal changed too; if not, the change is at best quantitative. See references/data-form.md.

Componentized R&D method (mandatory for 立项)

Never approve a vague goal like "develop a brain-like architecture." Translate it into concrete components by abstracting the human brain's functional systems, then plan each one.

From vague goal to 立项 tasks (3 steps):

1. Replace the vague goal with a list of human-brain functional systems (memory, attention, gating, neuromodulation, plasticity, predictive coding, deliberation — not "intelligence").
2. Map each system to one computable component with defined input / output boundaries.
3. Fill the 7-field template for every component (functional goal, system goal, input, output, cross-component interaction, inductive bias, verification). A component is 立项-ready only when all 7 fields are concrete and its verification is independently testable.

The full map — separated into a cross-cutting mechanism layer and a functional-system layer, connected by loop topology (not a sparse list) — lives in references/brain-component-map.md.

Overview:

机制层 (横切, 不单独立项): 连续时间动力学 (LTC) · 可塑性规则 (STDP/Hebbian) · 神经调质广播 · 预测编码 (贯穿原则)

功能系统层 (可立项的脑模块):

第四维度 — 数据形态/监督层: 数据形态直接决定拟合目标。换架构必须同步审视数据形态与监督 信号（否则自回归瓶颈从数据入口回流）。组件 → 数据形态 → 监督信号的完整映射见 references/data-form.md。

Liquid-architecture specifics

• Double down on the existing advantage. The liquid architecture already embeds 连续时间动力学 (continuous-time dynamics) as a native inductive bias — the root cause of its industrial-temporal generalization and parameter efficiency vs. Transformers. Keep deepening it.
• Pilot diffusion in state evolution. Optionally prototype a "gradual-correction, diffusion-like" mechanism in the liquid state-evolution step to further strengthen the physical inductive bias and sidestep the next-token bottleneck — small-scale validation only.
• Do not migrate diffusion wholesale. It would require redesigning the entire training methodology; treat it as an inductive-bias borrowing, not an architecture swap.