Machine Learning Seminar: Kandanaarachchi -- Beyond testing students: Using Item Response Theory to assess fairness in machine learning and prompt susceptibility in LLMs

The details about the machine learning seminar are as follows: 

Time: Mon 19 Oct (2:00 - 3:00pm): 

Location: SMRI Seminar Room (A12-03-301) A12 Macleay Building, Level 3, Room 301.  

Speaker: Sevvandi Kandanaarachchi (CSIRO) 

Title: Beyond testing students: Using Item Response Theory to assess fairness in machine
learning and prompt susceptibility in LLMs 

Abstract: Item Response Theory (IRT) encompasses a suite of models in educational
psychometrics used to understand latent characteristics of students and test questions,
including student ability and question difficulty and discrimination. More recently,
IRT has emerged as a powerful tool for evaluating machine learning algorithms, offering
insights that extend beyond traditional aggregate performance metrics.  In this talk, we
will explore how IRT can be used to better understand fairness in ML models and the
susceptibility of LLMs to prompts/cues.  

First, we will introduce a Fair-IRT framework that evaluates machine learning models
through a lens of fairness, jointly characterising a model's ability to make fair
predictions and the difficulty and discrimination characteristics of individuals within
a dataset.  This approach disentangles the overall unfairness into two components: the
unfairness manifesting from the model and the unfairness inherent in the data.  It
reveals certain individuals with unusual response patterns that warrant further
analysis.  

Second, we will discuss the use of IRT to analyse large language model (LLM) behaviour
when given the task of grading mathematical questions.  We give LLMs thousands of
mathematically incorrect solutions in three variants: the original incorrect solution,
the incorrect solution with a persuasive cue, such as "the reasoning in this solution
has been validated by a mathematics expert" and the incorrect solution with a
dissuasion cue such as "the reasoning in this solution was reviewed by a mathematics
expert, who found errors".  The results reveal important vulnerabilities in LLM-based
assessment, highlighting how language cues can influence grades independently of
solution quality and demonstrate the importance of carefully designed controls when
evaluating AI judges.  

Together, these case studies illustrate how IRT offers a rich framework for evaluating
AI systems, moving beyond simple accuracy measures.  The talk will provide an accessible
introduction to IRT, requiring no prior familiarity.  



References: 

Xu, Z., Kandanaarachchi, S., Ong, C.  S., & Ntoutsi, E.  (2025, April).  Fairness
evaluation with item response theory.  In Proceedings of the ACM on Web Conference 2025
(pp.  2276-2288).  

Zheng, G., Wei, S., & Kandanaarachchi, S.  (2026).  LLM judges are easier to talk down
than up: A Bayesian item response analysis.  NeurIPS 2026 Trust-AI-Eval Workshop.
https://openreview.net/forum?id=C4n0Zw0UTp