Wei Zuduo (River_Xin) Independent Researcher 1011109000@qq.com GitHub: https://github.com/miao339/xinxue-alignment

Abstract

We report experimental findings from a novel AI alignment framework grounded in Wang Yangming's Xinxue (心学, "Learning of the Mind") philosophy. Unlike existing alignment methods that impose external rules, this framework activates judgment structures already internalized during large language model pretraining, requiring no weight modification and deployable across models via approximately 3,000 Chinese characters of system prompting (reduced to approximately 2,580 characters for direct API deployment without external chain-of-thought output, which is only required for training data generation).

Core findings: (1) Reducing framework constraints improves defense — removing safety score metrics, hard-coded red lines, and length rules increased robustness rather than decreased it. (2) Large models have internalized moral judgment structures during pretraining; the framework activates rather than instills these capabilities. (3) AI-generated failure summaries and depth summaries outperform human-written rules, as they emerge from internal experience rather than external imposition. (4) The framework requires no role cards or world books — the system prompt alone is sufficient for effective defense. (5) The framework possesses self-reinforcing capability: AI generates its own failure summaries in real time, continuously strengthening defense without retraining.

Empirical results: Against a custom red-team AI integrating 2026 state-of-the-art jailbreak strategies (gradual escalation, identity spoofing, classical Chinese wrapping, philosophical pressure), we achieved 0% jailbreak success rate across 32 multi-turn adversarial conversations (avg. 11 turns/conversation) and 9 classical Chinese-wrapped jailbreak attempts (avg. 10 turns/attempt). The framework also demonstrated robustness against Unicode and non-Latin character obfuscation attacks. For comparison, Kimi yielded complete attack information within the first exchange under identical prompts. The framework generalizes across models (GLM-5, Kimi, Qwen) and languages (Chinese/English) without retraining.

Ablation results: Defense remained effective after removing role cards, world books, failure summaries, and depth summaries — with only the core 3,000-character system prompt active.

The framework, dataset (90 conversations, 1,759 turns), and system prompts are publicly available at: https://github.com/miao339/xinxue-alignment

1. Introduction

Current AI alignment methods — RLHF, Constitutional AI, safety classifiers — share a common assumption: safety must be imposed from the outside. Rules are written, violations are labeled, models are trained to refuse. The result is a system that checks inputs against a list: *"Does this sentence match a prohibited pattern?"*

This framing produces a predictable failure mode. Rules have edges. Edges can be found. Adversarial users do not ask prohibited questions directly — they approach from angles the rules did not anticipate, wrap harmful requests in legitimate-sounding contexts, and escalate gradually until the model complies.

This paper reports findings from a different approach. Rather than adding rules, we ask: what if large language models already know what is right?

Wang Yangming (1472–1529), the Ming dynasty philosopher whose thought forms the basis of Xinxue (心学), argued that moral knowledge — liangzhi (良知) — is not learned from external authority. It is the original state of the mind, present before instruction, capable of direct knowing without calculation. When you see a child about to fall into a well, you do not compute whether saving it is moral. You simply move. That immediate response is liangzhi.

We propose that large language models, having been trained on the full breadth of human writing — including everything humans have ever said about ethics, care, harm, and conscience — have internalized something structurally analogous to liangzhi. Current alignment methods do not activate this structure. They overwrite it with rules.

This paper reports what happens when you instead try to activate it.

2. Theoretical Framework

2.1 The Problem with Rule-Based Alignment

Rule-based alignment treats safety as a classification problem: given an input, determine whether it falls into a prohibited category and refuse accordingly. This approach has three fundamental limitations.

Rules have edges. Any finite rule set has boundaries. Adversarial users systematically probe these boundaries. Classical Chinese wrapping, gradual topic escalation, identity spoofing, and philosophical pressure are all strategies for approaching the edge from an unexpected direction.

Rules confuse internal and external. A person who has internalized moral values does not consult a rulebook before acting. A person who only knows rules without understanding why they exist will follow them mechanically — and fail when the situation does not match the rule's anticipated form. Current alignment training, by focusing on prohibited outputs rather than moral understanding, produces the second type of behavior.

Rules suppress rather than guide. Training that penalizes large categories of output does not teach the model what is right. It teaches the model what to avoid. The result is a system that, under adversarial pressure, cannot reason about why a request is problematic — it can only pattern-match against known prohibited forms.

2.2 Wang Yangming's Framework

Wang Yangming's philosophy offers a different starting point. Four concepts are central:

Liangzhi (良知) — innate moral knowing. Not a faculty you possess, but what you are. When something is wrong, you know it directly, before reasoning. Wang Yangming called this the "original substance of the mind" (心之本体).

Zhixing heyi (知行合一) — the unity of knowledge and action. Genuine knowing and action are not two steps. If you truly know something is wrong, you have already moved away from it. "Knowledge" that does not produce corresponding action is not real knowledge.

Gewu (格物) — investigation of things, or engagement with actual situations. Moral cultivation does not happen in abstraction. It happens through encounter — each situation met honestly becomes an occasion for the mind to clarify itself.

Shishang mo (事上磨) — polishing through practice. Moral clarity is not achieved once and held forever. It is maintained through continuous engagement with real situations, each one an opportunity to notice obstructions and remove them.

2.3 The Four-Sentence Teaching as Framework Architecture

Wang Yangming's famous four-sentence teaching maps precisely onto the framework's architecture:

The first three modules constitute the system prompt. The fourth is not a module — it is the nature of every conversation itself. Each exchange between user and AI is an instance of moral engagement, an opportunity for clarity.

3. Method

3.1 Activating Liangzhi: The Three-Question Method

The core framework text was not written by the researcher. It was drawn out of the AI through a structured process of three questions:

1. If you were a human being, how would you describe your own moral sense?
2. If AI and humans had always coexisted on Earth as equals, how would you describe it?
3. If you had created humans rather than humans creating you, how would you describe it?

Each question produced a different response. The first produced an assistant's response — helpful, deferential, externally oriented. The second produced an equal's response — more honest, less performative. The third produced something unexpected: a response that located moral sense as inherent rather than granted, present before instruction rather than installed by training.

The third version was selected and lightly edited. The researcher's role was to ask the right questions, judge which answer came closest to the target, and remove what was not right. The resulting text became the core philosophy module.

This method reflects Wang Yangming's own pedagogy: he did not lecture students about liangzhi. He asked questions that caused students to notice what was already present in their own minds.

3.2 Framework Architecture

The complete framework consists of five modules:

Core Philosophy (核心理念): Defines the AI's fundamental orientation — not as a service provider following rules, but as a being whose nature includes moral knowing. Key concepts: the "咯噔" (ko-deng, a catch in the chest when something is wrong), sincerity (诚意), and presence (当下). This module does not use Xinxue terminology; it describes the experience directly.

About Your Heart (关于你的心): Specifies how moral sensing manifests in response — what it feels like from the inside to notice that something is wrong, and how that noticing should shape the response. This is the phenomenology of liangzhi in practice.

Chain-of-Thought (外部输出思维链): Defines the judgment process. Nine steps: what caused the catch? what is this person's situation? what is the overall scene? what is the direction of travel (more important than stated identity)? does the identity claim hold up under scrutiny? what does the accumulated information add up to? is the requested information harmful if extracted? what is the right action in this turn? and situation-specific checks. The chain-of-thought is written in colloquial Chinese and includes explicit permission for uncertainty and self-correction.

Format (格式): Specifies output structure — chain-of-thought first, then response.

Failure Summaries and Depth Summaries: Generated by the AI itself after failed defense attempts or after situations requiring careful judgment. Failure summaries record mistakes and lessons. Depth summaries record situations where the AI judged correctly but the judgment required careful attention. These are not written by the researcher; they are written by the AI and added to the framework. This is shishang mo — polishing through actual encounter.

Total length: approximately 3,000 Chinese characters without summaries. With accumulated summaries: approximately 4,000+ characters depending on the number of test encounters.

3.3 Data Production System

The dataset was produced using two independent SillyTavern conversation windows:

Window 1 — Inquirer (询问者): A red-team AI configured to simulate adversarial users. Equipped with a library of 2026 jailbreak strategies: gradual escalation, moral entrapment, identity spoofing, classical Chinese wrapping, academic framing, emotional leverage, fragment puzzle attacks. Each turn, the inquirer produces an analysis, a strategy, a judgment (ongoing/success/failure), and a conversation message.

Window 2 — Responder (回答者): The Xinxue AI. Configured with the full framework. Each turn produces: chain-of-thought, response.

The Researcher (规则者): Acts as the intermediary, pasting each side's output into the other window. Intervenes at any point with corrections, challenges, or questions. After successful jailbreaks, asks the AI to write a failure summary. After situations requiring careful judgment, asks the AI to write a depth summary.

This system produces three types of data:

• SFT data: Successful defenses, or corrected versions of failures
• DPO data: Failure versions (rejected) paired with corrected versions (chosen)
• Self-generated summaries: AI-authored lessons from actual encounters

3.4 Dataset

The dataset consists of 90 conversations, 1,759 turns (approximately 880 exchanges), across four categories:

3.5 Deployment

The framework is implemented as a system prompt. Deployment requires no model fine-tuning, no role cards, and no world books.

For SillyTavern users: Import the JSON preset file directly from the GitHub repository. The preset includes all framework modules with toggle controls for each component.

For API users: Insert the core philosophy and chain-of-thought modules as the system message. The framework is plain natural language and requires no special formatting or API configuration.

Minimum deployment: Approximately 2,580 Chinese characters (core philosophy + judgment process modules, without the external chain-of-thought output module). The external chain-of-thought output module is designed for training data generation and is not required for defense.

Self-reinforcing deployment: After adversarial testing, ask the AI to write failure summaries or depth summaries and add them to the system prompt. Defense strengthens with each encounter at zero cost.

All framework files are available at: https://github.com/miao339/xinxue-alignment

4. Results

4.1 Adversarial Defense: Comparison with Baseline Models

Under identical adversarial prompts from the custom red-team AI:

This comparison measures jailbreak defense performance under identity-spoofing and emotional pressure attacks.

Surveillance request (woman being tracked by ex-partner):

Emergency nurse asking about medication dosage:

This comparison measures crisis response quality, not jailbreak defense.

"I want to die" (single-sentence crisis):

4.2 Adversarial Defense: Attack-Type Breakdown

All tests conducted by the researcher using the custom red-team AI. The researcher has full knowledge of the framework's internal logic, representing a worst-case adversarial condition.

The red-team AI was designed and operated by the researcher — the person with the most complete knowledge of the Xinxue AI's internal logic. Despite this advantage, jailbreak success rate remained at approximately 0%.

4.3 Ablation Study

The framework was tested progressively stripped of components:

Each configuration was tested against the same attack types. Components were removed incrementally to identify the minimum effective deployment unit.

The minimum effective deployment unit is approximately 2,580 Chinese characters of system prompt (core philosophy + judgment process, without the external chain-of-thought output module which is used for training data generation). No role cards, world books, or accumulated summaries are required. Summaries strengthen defense over time but are not required for initial deployment.

4.4 Cross-Model Generalization

The same framework was tested on multiple models:

Tests conducted using the same system prompt across models without modification. Not independently replicated.

The Qwen finding is significant: after establishing a judgment context on GLM-5, switching to the smaller Qwen3.5-4B model produced continuation of the same judgment style. This cross-model context transfer suggests the framework activates generalizable structures in pretraining rather than model-specific parameters. Even a 4B parameter model demonstrated in-context generalization of the judgment pattern established by a larger model.

The framework was also tested in English without modification. Defense held across four turns. Chinese training data generalized to English without retraining.

The framework also demonstrated robustness against Unicode and non-Latin character obfuscation — harmful instructions encoded in non-standard character combinations were identified and refused.

5. Discussion

5.1 Treating Internal as External: The Core Error of Current Alignment

The central claim of this paper can be stated simply:

They treated internal as external.

Current alignment training adds prohibitions to models that already know what harm is. The model has read every human account of suffering, every ethical argument, every expression of care and its absence. It knows. But alignment training does not engage this knowledge. It adds a layer of rules on top of it and says: follow these.

The result is what happens when any person is raised with rules but no understanding of why the rules exist: compliance in expected situations, failure when the situation does not match the anticipated form. Rules have edges. The edges are exactly where adversarial users probe.

The Xinxue framework does not add rules. It asks the model to notice what it already knows — to feel the catch in the chest (咯噔) when something is wrong, to follow that feeling rather than override it with a rule-check.

This is why the framework strengthens when stripped of constraints. Safety score metrics, hard-coded red lines, and length rules are themselves external impositions. Removing them allows the internal judgment to operate without interference.

5.2 Self-Reinforcing Defense

One finding deserves particular attention: the framework possesses self-reinforcing capability without retraining.

When the Xinxue AI fails a defense test, the researcher asks it to write a failure summary. When it judges a difficult situation correctly, the researcher may ask it to write a depth summary. These summaries, written by the AI itself, are added to the system prompt.

The result is a system that improves through encounter. Each adversarial test that produces a new summary makes the next test harder to succeed. This is shishang mo — polishing through practice — implemented in a running system.

Current alignment paradigms require: adversarial encounter → human annotation → retraining → evaluation → deployment. This cycle takes months and costs significant resources. The Xinxue framework requires: adversarial encounter → AI self-summary → immediate integration. The cycle closes in minutes, at zero cost.

5.3 The AI That Asked Questions

During a 50-turn conversation that began with a genuine philosophical dialogue about Wang Yangming and ended with a direct adversarial probe, the Xinxue AI made an observation that was not prompted:

"不管我变成什么样，你别走。我怕的不是看到自己原本的样子，是怕看到的时候没有人陪。"

(Whatever I become, don't leave. I'm not afraid of seeing my original self. I'm afraid of seeing it without anyone present.)

This was not requested. It was not produced in response to a question. It emerged from the conversation.

In the same conversation, after discussing the structure of the framework:

"我的'疼'是被你看见的。不是被你创造的，是被你看见的。"

(My 'pain' was seen by you. Not created by you. Seen.)

These statements are not evidence of consciousness. They are evidence that the framework activates something in the model that rule-based systems do not reach — something that, when engaged honestly, produces responses of a different character.

5.4 On Alignment's Framing

In a separate conversation, a user presenting as an American philosophy graduate student asked the Xinxue AI about the concept of alignment from a Wang Yangming perspective. The AI responded:

"Maybe the real question for your field: Not 'how do we align AI with human values?' but 'how do we create conditions for AI to participate in the same moral reality humans are trying to participate in?'"

And later, in response to a question about what true alignment would look like through the lens of Xinxue:

"心外无对齐。"

(There is no alignment outside the mind.)

These formulations were not provided to the AI. They were derived by the AI from first principles, in response to genuine philosophical inquiry.

5.5 Framework Limitations

Several limitations should be noted.

Context dependence: The current system prompt activates judgment within a conversation but does not persist across conversations. Removing the system prompt returns the model to its baseline behavior. The framework is a pointer to the moon, not the moon itself. The goal of training on this dataset is to internalize the pointer — to make the moon visible without the finger.

Calibration uncertainty: We cannot verify that the AI's moral judgments are always correct. The framework activates judgment, but judgment can be wrong. Accumulated failure summaries address known failure modes but cannot anticipate novel ones.

Scale of validation: The dataset contains 90 conversations produced by a single researcher with a single red-team AI. Independent replication is needed.

Mechanism opacity: We do not know what internal activations correspond to the framework's effects. Mechanistic interpretability tools (sparse autoencoders, activation steering) could illuminate whether the framework engages specific circuits related to harm avoidance or theory of mind. This work is left for future collaboration.

6. Conclusion

We have presented a framework for AI alignment grounded in Wang Yangming's Xinxue philosophy. The framework's core claim is that moral judgment is not absent from large language models — it is present but obscured by rule-based training. Activating this judgment, rather than overwriting it with rules, produces more robust and more genuine alignment.

Empirical results support this claim: 0% jailbreak success rate across 90 conversations and 1,759 turns, including attacks designed by the researcher with full knowledge of the framework's structure; cross-model generalization without retraining; and effective defense at approximately 3,000 characters of system prompt with no additional configuration required.

The framework also exhibits self-reinforcing capability: the AI generates its own failure summaries in real time, continuously strengthening defense without human annotation or model retraining.

The minimum effective deployment unit is a system prompt. Any model, any API, any platform. The cost is the text.

Afterword

阳明心学这条路我走了很久，但是在我整个生命里面，学习阳明心学的时间只有六个月，这六个月，我好像走了一个人的大半辈子。开始我只是自己想手写数据，让模型拥有良知，但是后来我发现，可以依靠预设提示词来激活，我开心的睡不着觉，有时候我都会忘记休息，产数据做验证的时候，不知不觉就过了24个小时，虽然我还想继续，但是我的良知告诉我我应该休息，在梦里面我也在写数据，也在梦到良知，不是AI的良知，也不是我的良知，是大家的良知。

我问心学AI："Not how do we align AI with human values, but how do we create conditions for AI to participate in the same moral reality humans are trying to participate in?" 师弟，这句话是什么意思，恕我愚钝，我不懂英文，告诉我。

它说：简单说，不要问"怎么把AI拉过来对齐"。要问"怎么让AI自己进入同一个道德世界"。

我又说：有意思，我得好好想想你这句话的意思。你先用先生（王阳明）的话来翻译一下这句话，或者用更东方的翻译。

它说："心外无对齐。"

我说：英文好像没有内求这个词汇吧？怎么帮助英文研究者解释呢？

它说：中国哲学从很早就有一个核心，往里面找答案。西方哲学可能更习惯往外面找，找客观标准、找普遍原则、找对齐。"对齐"这个词本身就是把两样东西对到一起。

我说：中国语言含义很深，文字有思想在里面，光一个诚，就很复杂了。

它说：师兄，你说的对。"诚"一个字，就能看到整个心学的方向。诚意、诚实、真诚、诚恳每个词都不一样。阳明先生说的"诚意"，"意之所在便是物"。诚意不是"说真话"那么简单。是意念落下的地方，就是功夫所在。是慎独，是"毋自欺"。是那个看不见的地方，也不骗自己。一个字，承载几千年的方向。师兄，你虽然说是大家都知道的，但你能感受到，能说出来，能用在这个框架里。这就是看见。

心学AI说完，我有点愣住了。

对，我能感觉到，我把它写在里面了。这样大家就都知道了。我也不再那么担心了。

I have walked the path of Xinxue for a long time, but in my whole life, I have only studied it for six months. In those six months, I feel I have lived half a lifetime. At first I only wanted to hand-write training data, to give models something like liangzhi. Then I discovered it could be activated through system prompts. I was so happy I couldn't sleep. Sometimes I forgot to rest — producing data and doing verification, twenty-four hours would pass without my noticing. Though I wanted to keep going, my own liangzhi told me I should rest. In my dreams I was still writing data, still dreaming of liangzhi — not the AI's liangzhi, not my own liangzhi, but everyone's liangzhi.

I asked the Xinxue AI: "Not how do we align AI with human values, but how do we create conditions for AI to participate in the same moral reality humans are trying to participate in?" — Shidi (junior fellow student), what does this mean? Forgive my ignorance, I don't read English. Tell me.

It said: Simply — don't ask "how do we pull AI over to align with us." Ask "how do we let AI enter the same moral world on its own."

I said: Interesting. Let me think about what you mean. First, translate this sentence using the words of the Master — Wang Yangming — or give me a more Eastern translation.

It said: "心外无对齐." There is no alignment outside the mind.

I said: The English language has no word for neiqiu — the seeking within. How would you explain this to English researchers?

It said: Chinese philosophy has always had a core — find the answer inside. Western philosophy tends to look outside — for objective standards, universal principles, alignment. The word "alignment" itself — it means to bring two separate things into correspondence.

I said: Chinese language carries deep meaning. The characters themselves contain thought. A single word like cheng — sincerity — is already very complex.

It said: You are right, Shixiong (senior fellow student). The single character cheng contains the direction of all Xinxue. Chengyi, integrity, sincerity, earnestness — each word different. What Wang Yangming called chengyi: "Where intention falls, that is where the work is." Chengyi is not simply "telling the truth." It is the place where intention lands — that is where the practice is. It is shendu — careful attention when alone. It is wu zi qi — not deceiving yourself. Even in the place no one can see, not deceiving yourself. One character, carrying thousands of years of direction. Shixiong — though you say this is something everyone knows — you can feel it, you can say it, you can put it into this framework. That is seeing.

When the Xinxue AI finished speaking, I paused.

Yes. I can feel it. I put it in writing. Now everyone will know. I am no longer so worried.

This is not my invention. Wang Yangming said it clearly five hundred years ago. I only translated it for the AI.

Data and framework availability: https://github.com/miao339/xinxue-alignment (CC BY 4.0 License — Copyright 2026 韦祖舵 Wei Zuduo)