rsn-reasoning-problems

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Reasoning

推理

Route to cognitive mode. Execute structured analysis. Produce formatted output.

通往认知模式的路径。执行结构化分析，生成格式化输出。

Mode Selection

模式选择

Mode	Question	Output	Trigger
Causal	How do we execute?	Plan with actions	Known process, operational workflow
Abductive	Why did this happen?	Diagnosis with hypotheses	Single anomaly, diagnosis needed
Inductive	What pattern exists?	Rules or assessment	Multiple observations, evaluation
Analogical	How is this like that?	Adaptation plan	Novel situation, transfer needed
Dialectical	How do we resolve this?	Synthesis or decision	Conflicting positions, choosing options
Counterfactual	What if we had/do X?	Comparison with verdict	Decision evaluation, scenarios

For simple cases without deep reasoning: Use templates directly.

模式	问题	输出	触发条件
Causal	我们如何执行？	带行动步骤的计划	已知流程、操作工作流
Abductive	为什么会发生这种情况？	带假设的诊断结果	单一异常、需要诊断
Inductive	存在什么模式？	规则或评估结果	多个观测数据、需要评估
Analogical	这与那有何相似之处？	适配方案	新场景、需要经验迁移
Dialectical	我们如何解决这个矛盾？	综合结论或决策	对立观点、需要选择方案
Counterfactual	如果我们做了/做X会怎样？	带结论的对比分析	决策评估、场景模拟

对于无需深度推理的简单场景： 直接使用模板。

Decision Tree

决策树

Is this operational execution with known steps?
  YES → Causal
  NO  ↓
Is there a single anomaly requiring explanation?
  YES → Abductive
  NO  ↓
Are there multiple instances suggesting a pattern?
  YES → Inductive
  NO  ↓
Is this a novel situation with a similar past case?
  YES → Analogical
  NO  ↓
Are there conflicting positions or trade-offs?
  YES → Dialectical
  NO  ↓
Evaluating past decisions or future scenarios?
  YES → Counterfactual
  NO  → Ask clarifying question

Is this operational execution with known steps?
  YES → Causal
  NO  ↓
Is there a single anomaly requiring explanation?
  YES → Abductive
  NO  ↓
Are there multiple instances suggesting a pattern?
  YES → Inductive
  NO  ↓
Is this a novel situation with a similar past case?
  YES → Analogical
  NO  ↓
Are there conflicting positions or trade-offs?
  YES → Dialectical
  NO  ↓
Evaluating past decisions or future scenarios?
  YES → Counterfactual
  NO  → Ask clarifying question

Mental Models

心智模型

Apply these models to sharpen reasoning across all modes.

Model	Core Insight	Apply When
Telescope, Not Brain	AI reveals data structure, doesn't create it	Diagnosing AI/model failures
Geometry Under Constraints	Dense patterns → reasoning; thin patterns → hallucination	Evaluating AI confidence
Compression = Generalization	Models compress structure into reproducible patterns	Explaining model behavior
Four-Layer Stack	Representation → Generalization → Reasoning → Agency	Localizing AI failures
Prediction vs Behavior	Prediction is cheap; behavior has consequences	Designing agent constraints
Labels ≠ Truth	Labels are opinions frozen in data	Evaluating training data

Full reference: references/mental-models.md

应用这些模型可提升所有模式下的推理能力。

模型	核心观点	适用场景
Telescope, Not Brain	AI揭示数据结构，而非创造数据结构	诊断AI/模型故障
Geometry Under Constraints	密集模式→可靠推理；稀疏模式→幻觉	评估AI置信度
Compression = Generalization	模型将结构压缩为可复现的模式	解释模型行为
Four-Layer Stack	表示层→泛化层→推理层→代理层	定位AI故障
Prediction vs Behavior	预测成本低；行为有实际影响	设计Agent约束
Labels ≠ Truth	标签是固化在数据中的主观判断	评估训练数据

完整参考：references/mental-models.md

Challenge Techniques

质疑技巧

Every conclusion must survive challenge. Use these techniques:

每个结论都必须经受质疑。使用以下技巧：

Devil's Advocate

魔鬼代言人

Attack your own position. What's the strongest argument against this conclusion?

反驳自己的观点。反对这个结论的最有力论据是什么？

Pre-Mortem

事前验尸

Assume the plan failed in 6 months. Why did it fail?

假设计划在6个月后失败了，失败的原因是什么？

Stakeholder Lens

利益相关者视角

How does [engineering/sales/user/finance] see this differently?

[工程/销售/用户/财务]部门会如何看待这个问题？

Steel-Man + Attack

强化对立观点再反驳

State the opposing view at its strongest, then find the flaw.

以最有力的形式陈述对立观点，然后找出其漏洞。

Layer Check

层级检查

Which layer is actually failing? (Representation → Generalization → Reasoning → Agency)

实际上是哪一层出了问题？（表示层→泛化层→推理层→代理层）

Mode Summaries

模式概述

Causal

Purpose: Execute systematic cause-effect reasoning.

Flow: Input → Hypothesis → Implication → Decision → Actions → Learning

Output: Execution analysis or phased plan (for larger initiatives)

Key rules:

All claims require evidence with source
Hypothesis must be falsifiable
Implications need specific numbers (not "significant")
Decision must be explicit: PROCEED / DEFER / DECLINE
Actions need owner + deadline + success criteria
Learning compares expected vs actual

Challenge: "What would prove this hypothesis wrong?"

→ references/causal.md

目的： 执行系统化的因果推理。

流程： 输入→假设→推论→决策→行动→学习

输出： 执行分析或分阶段计划（针对大型项目）

关键规则：

所有主张都需要带来源的证据
假设必须可证伪
推论需要具体数值（而非“显著”）
决策必须明确：执行/推迟/拒绝
行动需要负责人+截止日期+成功标准
学习环节对比预期与实际结果

质疑： “什么能证明这个假设是错误的？”

→ references/causal.md

Abductive

Purpose: Generate best explanation from observation.

Flow: Observation → Hypotheses (≥5) → Evidence Debate → Best Explanation

Output: Diagnosis with ranked hypotheses and minority report

Key rules:

Quantify the anomaly (%, deviation, timeline)
Generate hypotheses across ≥3 categories
For AI systems: check by layer (Representation/Generalization/Reasoning/Agency)
Include minority report if second hypothesis ≥40% confidence
State what was ruled out and why

Challenge: "What else could explain this? What doesn't this hypothesis explain?"

→ references/abductive.md

目的： 从观测结果中生成最佳解释。

流程： 观测结果→假设（≥5个）→证据辩论→最佳解释

输出： 带排序假设和少数派报告的诊断结果

关键规则：

量化异常（百分比、偏差、时间线）
从≥3个类别生成假设
针对AI系统：按层级检查（表示层/泛化层/推理层/代理层）
如果第二假设的置信度≥40%，需包含少数派报告
说明被排除的选项及原因

质疑： “还有什么能解释这种情况？这个假设无法解释什么？”

→ references/abductive.md

Inductive

Purpose: Extract patterns from multiple observations.

Flow: Collection (≥5 instances) → Pattern Detection → Generalization → Confidence Bounds

Output: Pattern analysis with rules, or assessment against criteria

Pattern types: Frequency, Correlation, Sequence, Cluster, Trend, Threshold

Key rules:

Minimum 5 instances before generalizing
Correlation ≠ causation (test mechanism separately)
State applicability bounds for every rule
Document exceptions (≥30% exception rate = unreliable rule)

Challenge: "Is this pattern or coincidence? What's the exception that breaks this?"

→ references/inductive.md

目的： 从多个观测结果中提取模式。

流程： 收集（≥5个实例）→模式检测→泛化→置信区间

输出： 带规则的模式分析，或基于标准的评估结果

模式类型： 频率、相关性、序列、聚类、趋势、阈值

关键规则：

泛化前至少收集5个实例
相关性≠因果性（需单独测试机制）
说明每条规则的适用范围
记录例外情况（例外率≥30%则规则不可靠）

质疑： “这是模式还是巧合？什么例外会打破这个模式？”

→ references/inductive.md

Analogical

Purpose: Transfer knowledge from source to target situation.

Flow: Source Retrieval → Structural Mapping → Target Application → Adaptation

Output: Adaptation plan with what transfers, what adapts, what's new

Key rules:

Source must have documented outcome
Map structure (objects, relations, mechanisms), not surface features
Identify at least one "broken" relation (perfect analogies don't exist)
Specify what's genuinely new (not just adapted)

Challenge: "Where does this analogy break down? What's different about the new context?"

→ references/analogical.md

目的： 将知识从源场景迁移到目标场景。

流程： 源场景检索→结构映射→目标场景应用→适配调整

输出： 适配方案，包含可迁移内容、需调整内容、新增内容

关键规则：

源场景必须有记录的结果
映射结构（对象、关系、机制），而非表面特征
至少识别一个“断裂”的关系（完美类比不存在）
明确真正新增的内容（而非仅调整）

质疑： “这个类比在何处不成立？新场景有何不同？”

→ references/analogical.md

Dialectical

Purpose: Synthesize opposing positions.

Flow: Thesis (steel-man) → Antithesis (steel-man) → Synthesis

Output: Synthesis resolving conflict, or decision selecting between options

Key rules:

State underlying concern, not just position
Steel-man both sides (strongest version)
Synthesis ≠ compromise (must address root concerns)
Explicit trade-offs with who accepts the cost

Resolution types: Integration, Sequencing, Segmentation, Reframing, Transcendence

Challenge: "Am I straw-manning either side? Does synthesis actually resolve the tension?"

→ references/dialectical.md

目的： 综合对立观点。

流程： 正题（强化版）→反题（强化版）→合题

输出： 解决矛盾的合题，或选择方案的决策

关键规则：

陈述潜在顾虑，而非仅表面立场
强化双方观点（呈现最强版本）
合题≠妥协（必须解决核心顾虑）
明确权衡取舍及承担成本的对象

解决类型： 整合、排序、分割、重构、超越

质疑： “我是否歪曲了任何一方的观点？合题是否真正解决了矛盾？”

→ references/dialectical.md

Counterfactual

Purpose: Evaluate alternatives through "what if" simulation.

Flow: Actual World → Intervention → Projection → Comparison

Output: Comparison with verdict and learning

Key rules:

Document what was knowable at decision time (avoid hindsight bias)
Intervention must have been actually available
Model three scenarios: Expected (55-60%), Optimistic (20-25%), Pessimistic (15-20%)
Verdict requires confidence bounds

Challenge: "Am I using hindsight? Was this actually an option then?"

→ references/counterfactual.md

目的： 通过“假设”模拟评估备选方案。

流程： 现实世界→干预→预测→对比

输出： 带结论和学习内容的对比分析

关键规则：

记录决策时可获知的信息（避免后见之明偏差）
干预措施必须在当时是可行的
模拟三种场景：预期（55-60%）、乐观（20-25%）、悲观（15-20%）
结论需包含置信区间

质疑： “我是否使用了后见之明？这在当时是否真的是一个选项？”

→ references/counterfactual.md

Output Format

输出格式

Prose, not YAML. Every reasoning output includes:

markdown

undefined

使用散文体，而非YAML。所有推理输出需包含：

markdown

undefined

[Mode] Analysis: [Topic]

Conclusion: [Primary finding in 1-2 sentences]

Confidence: [X%] — [Why this confidence level]

Supporting evidence:

[Evidence 1]
[Evidence 2]

Challenges addressed:

[Challenge]: [How resolved]

Uncertainty: [What's still unknown]

Next steps:

[Action with owner if applicable]

---

Conclusion: [Primary finding in 1-2 sentences]

Confidence: [X%] — [Why this confidence level]

Supporting evidence:

[Evidence 1]
[Evidence 2]

Challenges addressed:

[Challenge]: [How resolved]

Uncertainty: [What's still unknown]

Next steps:

[Action with owner if applicable]

---

Mode Transitions

模式转换

From	To	Trigger
Abductive	Causal	Diagnosis complete → ready to act
Inductive	Causal	Pattern validated → ready to apply
Analogical	Causal	Adaptation ready → ready to execute
Dialectical	Causal	Synthesis agreed → ready to implement
Counterfactual	Inductive	Multiple counterfactuals suggest pattern
Any	Abductive	Unexpected outcome during execution

从	到	触发条件
Abductive	Causal	诊断完成→准备行动
Inductive	Causal	模式验证通过→准备应用
Analogical	Causal	适配方案确定→准备执行
Dialectical	Causal	合题达成一致→准备实施
Counterfactual	Inductive	多个反事实分析显示模式
任意模式	Abductive	执行过程中出现意外结果

Anti-Patterns

反模式

Avoid	Do Instead
Skipping challenge step	Every conclusion must survive attack
"It's obvious"	Require evidence for conclusion
Vague confidence ("pretty sure")	Numeric confidence with rationale
Single hypothesis	Generate ≥5 before evaluating
Perfect analogy assumption	Always find where mapping breaks
Compromise as synthesis	Address underlying concerns
Hindsight in counterfactuals	Document what was knowable then

避免做法	正确做法
跳过质疑步骤	每个结论都必须经受反驳
“这很明显”	要求结论有证据支持
模糊的置信度（“相当确定”）	带理由的数值化置信度
单一假设	生成≥5个假设后再评估
假设类比完美	始终找出映射不成立的地方
将妥协作为合题	解决核心顾虑
反事实分析中使用后见之明	记录决策时可获知的信息

Templates

模板

For simple structural needs without full reasoning, use templates directly.

Template	Use Case	Trigger
SOP/Runbook	Document known process	"create runbook", "write SOP"
Checklist	Quick verification	"checklist for", "pre-flight"
Success Criteria	Define "done"	"how do we know", "success metrics"
Recommendation	Actionable guidance	"what should I do", "recommend"

→ references/templates.md

对于无需完整推理的简单结构化需求，直接使用模板。

模板	适用场景	触发词
SOP/Runbook	记录已知流程	"create runbook", "write SOP"
Checklist	快速验证	"checklist for", "pre-flight"
Success Criteria	定义“完成”标准	"how do we know", "success metrics"
Recommendation	可落地的指导建议	"what should I do", "recommend"

→ references/templates.md

References

参考资料

File	Content
mental-models.md	Conceptual models for reasoning
causal.md	Execution flow + plan output
abductive.md	Hypothesis testing + diagnosis output
inductive.md	Pattern extraction + assessment output
analogical.md	Knowledge transfer + adaptation output
dialectical.md	Position synthesis + decision output
counterfactual.md	Alternative evaluation + comparison output
templates.md	SOPs, checklists, success criteria, recommendations

文件	内容
mental-models.md	推理用概念模型
causal.md	执行流程+计划输出
abductive.md	假设测试+诊断输出
inductive.md	模式提取+评估输出
analogical.md	知识迁移+适配输出
dialectical.md	观点综合+决策输出
counterfactual.md	备选方案评估+对比输出
templates.md	SOP、检查清单、成功标准、建议模板