jan 2026
CCAI Collaboration with HCAI Accepted to ICLR 2026 in Rio de Janeiro, Brazil
A joint research collaboration between the Centre for Credible AI and the Human Centred AI Lab (HCAI) at the University of Technology Sydney has resulted in a paper accepted to ICLR 2026.
read publication
jan 2026
CCAI Collaboration with HCAI Accepted to ICLR 2026 in Rio de Janeiro, Brazil
A joint research collaboration between the Centre for Credible AI and the Human Centred AI Lab (HCAI) at the University of Technology Sydney has resulted in a paper accepted to ICLR 2026.
read publication
We are pleased to share that a joint research collaboration between the Centre for Credible AI and the Human Centred AI Lab (HCAI) at the University of Technology Sydney, led by Jianlong Zhou, has resulted in a paper accepted to ICLR 2026 – The Fourteenth International Conference on Learning Representations, which will take place in Rio de Janeiro, Brazil.
The paper addresses a fundamental problem in the evaluation of attribution methods. Many widely used evaluation metrics rely on perturbation techniques, such as naive feature masking, which can produce out-of-distribution samples and distort model behavior, leading to unreliable conclusions about explanation quality.
Proposed Evaluation Framework
To address this issue, the authors propose a new evaluation framework that:
- uses diffusion models to reconstruct masked regions so that evaluation samples remain in-distribution,
- ensures that reconstructed regions do not introduce new evidence affecting the model's prediction, and
- enforces hard constraints on features that must remain unchanged during evaluation.
The work was co-authored by Bartłomiej Sobieski and Przemysław Biecek in collaboration with researchers from the HCAI team.
The paper is available here: https://openreview.net/forum?id=FF14TqjU3e