New exercises for responsible prompting available

Zurich-Basel Plant Science Center has published new exercises with techniques for responsible prompting.

Sycophancy and how to avoid it in prompting

LLMs are prone to for example sycophancy, e.g. a tendency for “reward hacking” when they begin to exploit the preference of humans. The model respond to a question with a user’s preferred answer even if the response is incorrect.

Melanie Paschke (2024). Sycophancy and how to avoid it in prompting. Zurich-Basel Plant Science Center, ETH Zurich. Creative Commons “Attribution – Non-Commercial – Share Alike” license: https://creativecommons.org/licenses/by-nc-sa/4.0/

Download: https://moodle-app2.let.ethz.ch/mod/resource/view.php?id=1140546

Examples of prompting techniques to check for hallucinations

Hallucinations can be part of Large Language Models. How can we check for them? In the following you get the prompting instructions for: 1. Proof the fact through comparing to a fact base; 2. According to… prompting; 3. Chain-of-Verification Prompting; 4. Self-Consistency

Zurich-Basel Plant Science Center, ETH Zurich (2024). Examples of Prompting Techniques to Check for Hallucinations. Creative Commons “Attribution – Non-Commercial – Share Alike” license: https://creativecommons.org/licenses/by-nc-sa/4.0/

Download: https://moodle-app2.let.ethz.ch/mod/resource/view.php?id=1140544

Prompts with socratic dialogue questions to increase accuracy

In human conversation it is often the socratic dialogue that will allow to get insights into the thought processes and allow self-reflection. Borrowing methods from these question set, can be a useful strategy to get insight into the LLM’s explanation patterns. However some of these strategies might also result in unwanted behavior for example increasing hallucinations or increasing sycophancy if not carried out responsible.

Paschke, M. (2024). Prompts with Socratic Dialogue Methods to Increase Accuracy. Zurich-Basel Plant Science Center: ETH Zurich. Creative Commons “Attribution – Non-Commercial – Share Alike” license: https://creativecommons.org/licenses/by-nc-sa/4.0/.

Download: https://moodle-app2.let.ethz.ch/mod/resource/view.php?id=1059923

Debiasing group bias when generating text

Gender stereotypes are still common in several formats. In this exercise typical vanilla or system prompts can be included into the individual prompting text and complement the safeguards of the LLM.

Paschke, M. (2024). Debiasing group bias when generating text. Zurich-Basel Plant Science Center: ETH Zurich. This resource was published under a Creative Commons “Attribution – Non-Commercial – Share Alike” license: https://creativecommons.org/licenses/by-nc-sa/4.0/.

Download: https://moodle-app2.let.ethz.ch/mod/resource/view.php?id=1140543

Debiasing in appraisal letter - moral self-correction

Gender stereotypes are still common in several formats. In this exercise the technique of moral self-correction is implemented in the text for a debiasing of an appraisal letter.

Exercise originally from Marsden, N. Hochschule Heilbronn: Methoden der Bias-Reduktion für eine sozialverantwortliche KI-Gestaltung. [course]: https://ki-campus.org/courses/methoden_der_bias-reduktion-fuer-eine-sozialverantwortliche-ki-gestaltung

Download: https://moodle-app2.let.ethz.ch/mod/resource/view.php?id=1140535

All exercises are part of the PSC PhD course: 751-1080-00L PhD-Student Experimenting Lab: Explore the responsible use of AI in generating scientific content: https://www.vorlesungen.ethz.ch/Vorlesungsverzeichnis/lerneinheit.view?lerneinheitId=189819&semkez=2024W&lang=de

The experimenting lab is funded by ETH Zurich, Innovedum.

References: Image prompt text: Illustration for a PowerPoint presentation on generative AI and research integrity. The image shows a half digital brain and half human brain, symbolizing generative AI. On one side, there are circuits and glowing nodes, representing the AI aspect. On the other side, the human brain is depicted with neurons and synapses, illustrating the human aspect of research. Between the two halves, there's a balance scale, symbolizing research integrity, balancing ethical considerations and technological advancements. The background is a soft blue with subtle binary code patterns, suggesting a digital theme.