Guidance and Control Tokens
Abstract
Use system/prefix tokens, stop sequences, negative prompts, and guidance tokens to steer model behavior without changing weights.
Motivation
- Enforce style, length, or safety behaviors
- Condition diffusion models via classifier-free guidance or text conditioning
Architectures
- System/prefix instructions + stop sequences
- Control tokens (e.g., style/safety tokens) where models support them
- Diffusion guidance scales for image/video
Design Choices
- Strength of guidance vs. creativity
- Stop sequences and delimiters for termination
- Safety tokens and content filters
Pros/Cons
- Pros: Lightweight control, no retraining
- Cons: Provider-specific tokens; brittle across versions
Evaluation Metrics
- Adherence to style/length constraints
- Safety violation rates
Vendor/Tooling
- Provider system prompts; negative prompts for diffusion
- JSON mode/Schema as guidance for structure
Design Checklist
- Define allowed styles/tone and stop rules
- Test stability across decoding settings
- Combine with structured output validation
References
- Title: OpenAI Prompt engineering URL: https://platform.openai.com/docs/guides/prompt-engineering Publisher/Vendor: OpenAI Accessed: 2025-08-14 Version_or_release: provider_reported
- Title: Stable Diffusion guidance (Diffusers) URL: https://huggingface.co/docs/diffusers/index Publisher/Vendor: Hugging Face Accessed: 2025-08-14 Version_or_release: provider_reported