diversity-tuning

Diversity Tuning Skill

Triggers

Alternate expressions and non-obvious activations (primary phrases are matched automatically from the skill description):

• "mode collapse" → RLHF pattern mitigation via diversity tuning
• "RLHF collapse" → same
• "VS parameters" → Verbalized Sampling parameter adjustment shorthand

Purpose

Interactively tune Verbalized Sampling parameters to find the right diversity level for a given task.

Parameters

Parameter	Default	Range	Description
k	5	3-20	Number of responses to generate
threshold	0.1	0.01-1.0	Probability filter threshold
variant	standard	standard/cot/multi	VS prompt variant

Presets

Preset	k	threshold	variant	Use Case
conservative	3	1.0	standard	Minimal diversity, high confidence
moderate	5	0.1	standard	Balanced (default)
creative	8	0.05	cot	High diversity for ideation
maximum	12	0.01	multi	Maximum exploration

Usage

# Apply a preset
"tune diversity to creative"

# Custom parameters
"set diversity k=8 threshold=0.05"

# A/B comparison
"compare diversity: moderate vs creative for tagline generation"

A/B Comparison Mode

When comparing presets:

1. Generate outputs using preset A
2. Generate outputs using preset B
3. Show side-by-side comparison
4. Measure diversity metrics:
   • Unique n-gram ratio
   • Average pairwise edit distance
   • Semantic cluster count
5. Recommend the better preset for the task

References

• @$AIWG_ROOT/agentic/code/addons/aiwg-utils/rules/vague-discretion.md — Measurable diversity thresholds and tuning criteria
• @$AIWG_ROOT/agentic/code/addons/aiwg-utils/rules/instruction-comprehension.md — Parsing tuning parameters accurately
• @$AIWG_ROOT/docs/extensions/overview.md — Extension system architecture
• @$AIWG_ROOT/agentic/code/addons/voice-framework/README.md — Voice framework for style diversity context