hierarchical-reasoning（层次推理）

Hierarchical Reasoning

Execute sophisticated reasoning through a three-level cognitive architecture that mirrors human multi-timescale processing: strategic (abstract planning), tactical (approach design), and operational (detailed execution).

Core Principle

Separate concerns by abstraction level while maintaining bidirectional information flow:

• Strategic level formulates problems and sets goals (slow, abstract)
• Tactical level designs approaches and selects methods (medium, structured)
• Operational level performs detailed computations (fast, concrete)

Each level informs the others through iterative refinement until convergence.

When to Use

Apply hierarchical reasoning for:

1. Complex multi-step problems requiring systematic decomposition
2. Strategic questions needing both big-picture framing and detailed analysis
3. Design challenges bridging abstract requirements to concrete implementation
4. Reasoning under uncertainty where confidence tracking matters
5. Deep analysis requiring multiple passes of refinement
6. First principles thinking building from fundamentals to emergent systems

Do NOT use for:

• Simple factual lookups
• Single-step calculations
• Purely creative tasks without logical structure
• Real-time processing requirements

Usage Pattern

1. Problem Structuring

Frame the problem with clarity:

Problem: [Clear statement of what needs to be reasoned about]
Context: [Relevant background, constraints, domain knowledge]
Success Criteria: [How to evaluate reasoning quality]

2. Execute Hierarchical Reasoning

Use the reasoning script:

python scripts/hierarchical_reasoner.py "<problem>" --context '<json_context>'

Or invoke programmatically:

from hierarchical_reasoner import HierarchicalReasoner

reasoner = HierarchicalReasoner(
    max_strategic_cycles=3,
    max_tactical_cycles=5, 
    max_operational_cycles=7,
    convergence_threshold=0.95,
    uncertainty_threshold=0.1
)

result = reasoner.reason(problem, context)

3. Interpret Results

Examine multi-level outputs:

• Strategic state: Problem formulation, goals, high-level insights
• Tactical state: Approaches, methods, reasoning strategies
• Operational state: Detailed analysis, computations, evidence
• Convergence metrics: State stability and confidence at each level
• Final synthesis: Integrated conclusion across all levels

4. Iterate Based on Convergence

If not converged (convergence_score < threshold):

• Review which level has low convergence
• Add relevant context or constraints
• Increase cycle count for that level
• Re-run with refined inputs

Reasoning Flow

┌─────────────────────────────────────────────┐
│ STRATEGIC LEVEL (Abstract Planning)         │
│ • Problem formulation                       │
│ • Goal identification                       │
│ • Success criteria                          │
└────────────┬────────────────────────────────┘
             │ guides ↓    ↑ informs
┌────────────▼────────────────────────────────┐
│ TACTICAL LEVEL (Approach Design)            │
│ • Method selection                          │
│ • Decomposition strategy                    │
│ • Reasoning structure                       │
└────────────┬────────────────────────────────┘
             │ guides ↓    ↑ informs
┌────────────▼────────────────────────────────┐
│ OPERATIONAL LEVEL (Detailed Execution)      │
│ • Concrete computations                     │
│ • Evidence gathering                        │
│ • Detailed analysis                         │
└─────────────────────────────────────────────┘

Convergence Detection

Reasoning converges when:

1. State content stabilizes across iterations (high similarity)
2. Confidence increases above threshold (typically 0.90-0.95)
3. Uncertainty decreases below threshold (typically 0.10-0.15)

Convergence Score Formula:

score = 0.7 × similarity(current, previous) + 0.3 × confidence

Multi-level convergence:

converged = all levels > threshold OR weighted_average > threshold
weighted_avg = 0.5×strategic + 0.3×tactical + 0.2×operational

Configuration Parameters

Cycle Counts (iterations per level):

• max_strategic_cycles: 2-5 (typically 3)
• max_tactical_cycles: 3-7 (typically 5)
• max_operational_cycles: 5-10 (typically 7)

Higher values allow more refinement but increase computation.

Thresholds:

• convergence_threshold: 0.90-0.98 (default 0.95)
• uncertainty_threshold: 0.05-0.20 (default 0.10)

Higher convergence thresholds require more stable reasoning.

Example: Complex Problem Analysis

Problem: "Design a sustainable urban transportation system"

Strategic Output:

- Goal: Minimize environmental impact while maximizing accessibility
- Constraints: Budget, existing infrastructure, citizen adoption
- Success metrics: Emissions reduction, ridership, cost-effectiveness

Tactical Output:

- Approach: Multi-modal integration (public transit + micro-mobility)
- Method: Network optimization with accessibility mapping
- Validation: Simulation before implementation

Operational Output:

- Route optimization calculations
- Demand forecasting analysis
- Cost-benefit quantification
- Environmental impact assessment

Synthesis: Integrated design bridging strategic goals through tactical approaches to operational specifications, with confidence metrics for each component.

Advanced Features

Uncertainty Quantification

Each state tracks:

• Confidence: Self-assessed certainty (0.0 to 1.0)
• Uncertainty: Recognized unknowns and ambiguities
• Convergence: State stability over iterations

Use these for:

• Identifying areas needing more analysis
• Selective trust in conclusions
• Metacognitive awareness of reasoning quality

Reasoning Trace

Enable full diagnostic trace:

python scripts/hierarchical_reasoner.py "<problem>" --trace

Outputs:

• State evolution at each cycle
• Convergence progression
• Confidence/uncertainty dynamics
• Dependency tracking

Useful for:

• Debugging reasoning failures
• Understanding convergence patterns
• Analyzing computational efficiency

Architecture Details

For deep technical understanding of:

• Theoretical foundations
• Cognitive architecture alignment
• Implementation patterns
• Scaling properties
• Future enhancements

See: references/architecture.md

Integration with Other Tools

Hierarchical reasoning complements:

• Web search: Operational level can gather facts
• Code execution: Operational level performs calculations
• Knowledge graphs: Strategic level structures problem space
• Document analysis: Tactical level designs analysis approach

Combine reasoning levels with appropriate tools at each abstraction.

Troubleshooting

Low convergence scores:

• Increase cycle count for non-converging level
• Refine problem statement clarity
• Add relevant context to reduce uncertainty
• Check if problem requires more/fewer abstraction levels

High computational cost:

• Decrease cycle counts
• Increase convergence threshold (accept earlier stopping)
• Use adaptive early stopping
• Parallelize level updates (future enhancement)

Inconsistent multi-level outputs:

• Review information flow (are levels properly informing each other?)
• Check if strategic goals are well-defined
• Verify tactical approaches align with strategy
• Ensure operational details serve tactical plan