Axiom is a research initiative investigating the fundamental mechanisms behind how Large Language Models rank and prioritize information. In an era where LLMs increasingly mediate access to knowledge, understanding whether their responses are predominantly shaped by pretraining data or can be meaningfully steered by contextual signals has become a critical research question.
How do pretraining data and contextual signals interact to determine LLM outputs, and can we develop principled methods to influence this interaction?
Modern LLMs generate responses through a complex interplay between:
Current understanding of how these components interact remains limited, hindering the development of principled methodologies for influencing LLM outputs in reliable, predictable ways.
Develop statistical methodologies to determine whether LLM responses to specific queries are primarily influenced by pretraining data or contextual signals.
Establish frameworks to quantify the relative contributions of different information sources in LLM ranking decisions.
Create theoretically-grounded algorithms that can provably steer LLM rankings toward desired outcomes.
Axiom employs a multi-faceted methodology grounded in statistical learning theory:
Our research has significant implications for:
Axiom brings together researchers from machine learning, statistics, information retrieval, and human-computer interaction. We actively collaborate with academic institutions and industry partners to ensure our research addresses real-world challenges while advancing theoretical understanding.
The Axiom project aims to transform how we understand and interact with AI systems, moving from opaque black boxes to transparent, steerable partners in information discovery.