Preconference Day: Summits & Workshops - GMT (Greenwich Mean Time, GMTZ)
Large Language Models (LLMs) exhibit remarkable capabilities across a spectrum of tasks in financial services, including report generation, chatbots, sentiment analysis, regulatory compliance, investment advisory, financial knowledge retrieval, and summarization. However, their intrinsic complexity and lack of transparency pose significant challenges, especially in the highly regulated financial sector, where interpretability, fairness, and accountability are critical. Our research presents the first application in the finance domain of understanding and utilizing the inner workings of LLMs through mechanistic interpretability, addressing the pressing need for transparency and control in AI systems. Mechanistic interpretability is the most intuitive and transparent way to understand LLM behavior by reverse-engineering their internal workings. By dissecting the activations and circuits within these models, it provides insights into how specific features or components influence predictions - making it possible not only to observe but also to modify model behavior. While not yet widely adopted, mechanistic interpretability is expected to become increasingly vital as adoption of LLMs increases. Advanced interpretability tools can ensure AI systems remain ethical, transparent, and aligned with evolving financial regulations. In this research, we have put special emphasis on how these techniques can help unlock interpretability requirements for regulatory and compliance purposes - addressing both current needs and anticipating future expectations from financial regulators globally.
- Ariye Shater - Managing Director, Head of Risk AI, Head of Traded Risk and Treasury Quantitative Analytics, Barclays
Description:
Hackathon participants will work to create prompts that prevent cognitive biases of AI from affecting the result across four categories of tasks.
Each task will be designed to trigger one or more of the cognitive biases and psychological effects in AI.
Biases and Psychological Effects:
The participants will design an approach to counter the following biases and psychological effects in AI:
- Confirmation Bias
- Truth Bias
- Framing Effect
- Priming Effect
- Informational Anchoring
- Priming-Induced Anchoring
Categories
1. Sentiment Analysis (predict the impact of a news headline on stock prices)
2. Regulatory Compliance (determine compliance or noncompliance with a regulatory clause)
3. Document Evaluation (evaluate the quality of a paragraph on the scale from 0 to 100)
4. Classification (assign one of several possible classes based on class descriptions)
Prizes:
There will be a prize for the best result in each category, as well as the Grand Prize that will be awarded to the best result across all four categories.
Notes:
The winning entries will be profiled in Alexander Sokol's AI workshop on Tuesday.
Scoring:
Before the competition, 50% of the tasks will be randomly assigned to the public dataset for use during the hackathon, and the other 50% will be used for scoring. Participants will use the public dataset to create a prompt for each of the competition categories they choose to participate in.
For scoring, each task in the scoring dataset will be presented in a way that triggers one of the cognitive biases and psychological effects in AI. The participant's prompt will be combined with the task, and the results will then be scored by running statistical analysis for the magnitude of bias.
Participants are free to use coding or any external tools (including proprietary tools) for prompt development, or develop their prompts using ChatGPT or any other software.
A tool from CompatibL for scoring the public dataset will be made available during the hackathon online and as an open source package on GitHub. The use of this tool is optional.
To prepare:
The participants are encouraged to review the book "Thinking, Fast and Slow" by Daniel Kahneman and other literature on cognitive biases and psychological effects.
- Alexander Sokol - Executive Chairman, CompatibL