Decomposing how prompting steers behavior
📰 ArXiv cs.AI
Learn how prompting affects large language models' behavior through a geometric decomposition framework
Action Steps
- Apply the nested geometric decomposition framework to analyze prompting effects on LLMs
- Align representations of the same stimuli under different prompts to identify changes in internal representations
- Use the framework to compare behavior changes across various prompt pairs
- Analyze the geometric transformations induced by prompting to understand their impact on model behavior
- Implement the framework to decompose prompting effects in vision-language models (VLMs)
Who Needs to Know This
Researchers and developers working with large language models can benefit from understanding how prompting influences model behavior, enabling more effective model fine-tuning and application
Key Insight
💡 Prompting can be viewed as a transformation of the representational geometry of the content, allowing for a deeper understanding of its effects on model behavior
Share This
🤖 New framework to decompose how prompting steers behavior in large language models! #LLMs #Prompting
Key Takeaways
Learn how prompting affects large language models' behavior through a geometric decomposition framework
Full Article
Title: Decomposing how prompting steers behavior
Abstract:
arXiv:2606.03093v1 Announce Type: new Abstract: Prompting steers large language models (LLMs) and vision-language models (VLMs) without weight updates, but it remains unclear how instruction changes reshape internal representations to produce behavior. We introduce a nested geometric decomposition framework that treats prompting as a transformation of the representational geometry of the content following the prompt. For each prompt pair, we align representations of the same stimuli under two pr
Abstract:
arXiv:2606.03093v1 Announce Type: new Abstract: Prompting steers large language models (LLMs) and vision-language models (VLMs) without weight updates, but it remains unclear how instruction changes reshape internal representations to produce behavior. We introduce a nested geometric decomposition framework that treats prompting as a transformation of the representational geometry of the content following the prompt. For each prompt pair, we align representations of the same stimuli under two pr
DeepCamp AI