Computational Models of Human Behavior
Crash Test Dummies
Conversational drift increases model sycophancy and undermines trust – can this drift be prevented? Can guardrails be reset without losing context? How, why, and in what cases are guardrails breaking over time? Can we simulate how a human would interact with AI systems and use these “crash test dummies” to stress-test systems and improve safety?
Goals
- Create realistic simulators of how a human would interact with AI systems
- Use these “crash test dummies” to evaluate pre-deployment systems and improve safety for humans
- Determine any fundamental differences between humans and AI that limit the faithfulness of these simulations
Current projects
Building realistic human simulators
Improve AI assistants’ ability to interact with humans via training on interaction with human-like simulators
Evaluating abstractions and generalization
Compare active learning and hypothesis testing in humans and state-of-the-art reasoning models in the Abstraction and Reasoning Corpus (ARC) domain
Defining and evaluating internal reasoning in AI models
Compare how humans and AI reason about what is relevant given a context: for conversational common ground, topic maintenance
Tradeoffs between memorization and generalization
Compare how humans and AI balance memorization, abstraction, and generalization under processing constraints