00:00: Introduction
00:27: Why two papers?
01:28: What are "simulated AI agents"?
03:34: Applications of simulated AI agents in Marketing?
06:39: Priming language models using personas
07:20: The two questions answered in the papers
08:47: Paper 1
08:55: Two views on LLM applicability in Marketing research
11:02: The need to interpretability methods for LLM outputs
12:32: Explaining the Shapley value method through an example
16:28: Mathematical formulation of Shapley value
19:33: Shapley value properties
20:21: Shapley value method in a nutshell 🌰
20:46: Shapley value—Application 1: A Discrete Choice Experiment
22:42: From p-hacking to prompt-hacking
24:22: Shapley value—Application 2: Existence of Cognitive Biases in LLMs
24:53: Framing effect (or lack thereof) in LLMs
26:33: Revisiting the two views on LLM applicability in Marketing research
27:06: LLMs as distributions of distributions (from a Bayesian perspective)
28:54: Paper 2
29:03: Base vs. Aligned LLMs
31:42: LLMs produce one token at a time
33:00: Experiment 1: Customer Persona and Review Generation
36:16: Sentiment analysis of the reviews generated by simulated customers
37:23: Experiment 2: Semantic-level variation in LLM outputs
39:22: Hypothesis: Low Semantic Diversity ⇒ Low Syntactic Diversity
40:46: RLHF blocks token trajectories (lowers syntactic diversity)
41:13: Experiment 3: Syntactic diversity in LLM outputs
43:28: Entropy analysis of model outputs
44:15: Attractor States
47:00: FAQ
48:06: The End

In this video, I talk about the papers that I have been included in my job market package. These papers are about the applicability of large language models in marketing research studies, and each paper addresses one aspect of the behavior of simulated AI agents.

Paper 1: Wait, It's All Token Noise? Always Has Been: Interpreting LLM Behavior Using Shapley Value
Link: [ Ссылка ]

Paper 2: Creativity Has Left the Chat: The Price of Debiasing Language Models
Link: [ Ссылка ]

свернуть