Can LLMs work in a spectrum of thought

Large Language Models (LLMs) are increasingly used in everyday life, yet they often make decisions based on internal assumptions and tendencies that users cannot see. These hidden patterns, called inductive biases, shape how models interpret concepts, judge situations, and respond to people. Understanding these biases is essential for building safer and more trustworthy AI systems.
To explore this, we introduce a novel game-like activity inspired by the psychological dynamics of the party game Wavelength. In this activity, an AI model must place different concepts along a spectrum of opposites, such as cold ? hot or more subjective dimensions like safe ? dangerous or moral ? immoral. This simple interaction offers a powerful way to reveal how LLMs internally organize meaning and make conceptual judgments.
Building on this idea, the project develops a controlled experimental pipeline grounded in mechanistic interpretability techniques. By comparing multiple LLMs, analyzing layer-wise contributions, and applying targeted causal interventions to internal components, we investigate how these interventions shift the model’s conceptual decisions. This allows us to map where certain biases arise and how stable or fragile they are under perturbation.

Faculty Supervisor:

Leila Kosseim

Student:

Partner:

University of Copenhagen

Discipline:

Computer science

Sector:

Education

University:

Concordia University

Program: