NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
AI Research
arXiv CS.AI·May 25·8 min read
As X, Do Y: How Persona and Task Combine in Instruction-Tuned LLMs
TL;DR
When language models receive role prompts like "As a doctor, diagnose this," it was unclear how persona and task components combine internally. Researchers discovered that at the prompt-to-answer transition in the residual stream (the model's internal state), persona and task effects decompose into separate, partially orthogonal additive directions in Gemma-2-2B-IT and Qwen models.
✦ Why It Matters
Engineers can steer persona and task contributions independently for fine-grained LLM control, but cannot compress prompts into single vectors for efficiency.
Key Takeaways
- Explore role prompts structured as 'As X, do Y' for better LLM outputs.
- Understand the additive structure of persona and task in LLMs.
- Avoid relying on cached predictions for persona-specific generation.
- Utilize findings to improve interpretability and control in LLM applications.
How It Works
The study reveals that instruction-tuned LLMs decompose persona and task contributions at a specific point in their processing. This occurs at the prompt-to-answer transition, where the last prompt token and the first two generated tokens interact.
The persona effect and task effect are additive but not compressible into a single vector, as the persona's influence is distributed throughout the entire generation process.