Verbalized Sampling

Training-free prompting technique that mitigates LLM mode collapse by generating multiple candidate responses with explicit probability assignments, then selecting from the distribution. Based on the VS research paper — achieves 1.6-2.1x diversity increase over direct prompting.

When to Apply Automatically

Apply VS internally (without the user asking) when detecting:

• Brainstorming or ideation requests
• "Give me something original/creative/unexpected"
• Tasks where the first instinct feels generic or predictable
• Requests for multiple options or alternatives

Core Workflow

1. Generate Candidates

Internally generate k candidate responses using this structure:

<response>
<text>Candidate response content.</text>
<probability>0.07</probability>
</response>

Default parameters: k=5 candidates, probability threshold tau=0.10.

2. Choose a Mode

Default to Tail mode for creative tasks. Use Standard for analytical tasks. See references/prompt-templates.md for exact prompt text.

3. Select Output

After generating candidates internally, select using one of these strategies:

• Present all options — Show the user all k candidates when they want to choose (human-in-the-loop)
• Weighted random — Pick one, weighted by probabilities (default for creative tasks)
• Best-of-k — Pick highest probability (default for factual/analytical tasks)
• Tail pick — Pick lowest probability for maximum surprise
• Merge — Synthesize best elements from multiple candidates

See references/selection-strategies.md for detailed guidance on each strategy.

4. Present the Result

When running VS transparently (user asked for options): show all candidates with brief probability context.

When running VS internally (auto-detected creative task): present only the selected output. Optionally mention that you explored multiple approaches.

Parameter Tuning

• k=3: Simple tasks, quick options
• k=5: Default, good balance of diversity and efficiency
• k=7-10: Maximum diversity, brainstorming sessions
• tau=0.10: Default tail threshold
• tau=0.05: Extreme divergence, very unconventional outputs
• tau=0.20: Mild diversity, closer to standard generation

Fork Mode (Advanced)

For complex agentic workflows, execute VS in a "context fork":

1. Spawn a separate reasoning path
2. Generate k candidates with probabilities
3. Evaluate candidates against task requirements internally
4. Return only the best candidate to the main conversation

Use fork mode when the user wants a single polished answer but the task benefits from internal exploration. Do not show the fork process unless asked.