Summary of Unveiling and Manipulating Prompt Influence in Large Language Models, by Zijian Feng et al.
Unveiling and Manipulating Prompt Influence in Large Language Models
by Zijian Feng, Hanzhang Zhou, Zixiao Zhu, Junlang Qian, Kezhi Mao
First submitted to arxiv on: 20 May 2024
Categories
- Main: Computation and Language (cs.CL)
- Secondary: Artificial Intelligence (cs.AI)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary In this paper, researchers propose a new approach called Token Distribution Dynamics (TDD) to study how individual tokens in prompts affect the responses of Large Language Models (LLMs). TDD uses the language model head to analyze input saliency and project tokens into an embedding space. This allows it to estimate token significance based on distribution dynamics. The authors introduce three TDD variants: forward, backward, and bidirectional, each providing unique insights. Extensive experiments show that TDD outperforms state-of-the-art baselines in understanding prompt-token relationships. Additionally, the authors apply TDD to two prompt manipulation tasks: zero-shot toxic language suppression and sentiment steering. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us understand how prompts influence LLM responses. The researchers created a new method called Token Distribution Dynamics (TDD) that uses the language model head to figure out which parts of the prompt matter most. They showed three ways TDD can work: forward, backward, and bidirectional. This helped them see how prompts affect LLMs better than other methods did. They also tested TDD on two tasks: making text less toxic and steering sentiment. |
Keywords
» Artificial intelligence » Embedding space » Language model » Prompt » Token » Zero shot
