Taxonomy Build

Step 2: Parse All Paper Analyses

For each paper analysis file:

1. Extract the 8-dimensional classification
2. Extract paper metadata (ID, title, authors)
3. Extract the evidence table
4. Collect all classifications

Step 3: Build Hierarchical Structure

Level 1: Method Category

Group papers by Method Category dimension:

• Method Family A
• Method Family B
• Method Family C
• System/Deployment-oriented Methods
• Evaluation/Analysis-oriented Methods
• Other

Level 2: Submethod

Within each Method Category, group by Specific Method:

• Technique A
• Technique B
• Technique C
• etc.

Level 3: Specific Technique

Group by combinations of Specific Method + Training Paradigm or other relevant sub-groupings.

Step 4: Cross-Cutting Dimensions

Also organize by:

• Training/Optimization Paradigm: profile-aligned paradigms
• Core Challenge: profile-aligned challenge set
• Evaluation Focus: profile-aligned metric/task set

Step 5: Coverage Analysis

Calculate statistics:

• Papers per Model Type
• Papers per Evaluation Focus
• Papers per Hardware Co-design
• Most common method categories
• Most common challenges

Step 6: Interconnection Analysis

Identify patterns:

• Which method categories address which challenges?
• Which training paradigms work best for which challenges?
• Any notable method combinations?

Step 7: Generate Output

Write taxonomy.md following templates/TAXONOMY_TEMPLATE.md:

# Taxonomy

## Hierarchical Taxonomy Structure

### Level 1: Method Category

#### Representation Enhancement
##### Level 2: Submethod
###### Learnable Scaling/Offset
...

Step 8: Update Coverage Analysis Table

Include:

• Papers by Model Type
• Papers by Evaluation Focus
• Papers by Hardware Co-design
• Method-Challenge Matrix

Taxonomy Statistics to Report

• Total Papers Analyzed
• Number of Level-1 Categories
• Number of Level-2 Submethods
• Most Common Method Category
• Most Common Challenge
• Coverage gaps (e.g., "no papers on X")

Key Rules

• Derive from evidence: Taxonomy must be grounded in actual paper analyses
• Evidence binding: Each taxonomy node should list supporting paper IDs
• Machine-readable: Output must be parseable by /gap-identify skill
• Hierarchical clarity: Ensure clear parent-child relationships
• Cross-references: Link related categories where methods overlap

Example Output Structure

# Taxonomy: [Survey Topic]

## Hierarchical Taxonomy Structure

### Level 1: Method Category

#### Representation Enhancement
**Definition**: Methods that share a common mechanism under this category
**Papers**: [2401.12345, 2402.23456]

##### Level 2: Submethod

###### Learnable Scaling/Offset
**Definition**: Submethod definition grounded in paper evidence
**Papers**: [2401.12345]
**Key Techniques**: [list concrete techniques from evidence]

###### Rotation/Orthogonal Transform
**Definition**: Methods that apply rotations to weight matrices
**Papers**: [2402.23456]
**Key Techniques**: Hadamard transform, orthogonal matrices

### Level 1: Training Paradigm

#### [Paradigm Name]
**Definition**: Paradigm definition grounded in paper evidence
**Papers**: [2401.12345, 2402.23456, 2403.34567]

...

For survey construction workflow, see /survey-pipeline For gap identification, see /gap-identify