Hamed Damirchi

Adelaide, Australia

My work is on large language models, both making them more useful in applications and interpreting them, so that we know when they are working, how they arrive at their outputs and where those outputs come from. The concern throughout is performance, reliability and safety.

My earlier work was in robotics, on localisation and sensor fusion. I have also built applied AI systems with engineers and domain experts, in energy forecasting, healthcare and AI policy.

I am completing my PhD at the Australian Institute for Machine Learning, advised by Javen Shi, Yuhang Liu and Zhen Zhang, and supported by the Centre for Augmented Reasoning (CAR) and the Responsible AI Research (RAIR) Centre.

What I work on

Reliable LLM reasoning: telling correct reasoning from errors using how a model’s internal representations change across layers.
Efficient adaptation and model editing: fine-tuning with fewer trainable parameters, and editing models through task vectors.
Applied AI systems: multi-agent LLM deliberation, demand forecasting, healthcare computer vision and robot localisation.

Recent

2026: Truth as a Trajectory was presented as an oral at ACL 2026 and nominated for the Best Paper Award. Paper · project page
2026: Secondary Instructor for Large Language Models and Applications (COMP6012) at Adelaide University, from June.
2026: New preprints on reasoning errors in the residual stream, representation trajectories for OOD detection and vertical fusion for ViT classification.
2025: Synergy and Diversity in CLIP at ICLR 2025.

selected publications

ACL

Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning

Hamed Damirchi, I. Meza De la Jara, E. Abbasnejad, A. Shamsi, Z. Zhang, and J. Q. Shi

In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Long Papers), 2026

Oral · Best Paper nominee Bib HTML Website

Oral presentation and Best Paper Award nominee at ACL 2026.

@inproceedings{damirchi2026truth,
  title = {Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning},
  author = {Damirchi, Hamed and {Meza De la Jara}, I. and Abbasnejad, E. and Shamsi, A. and Zhang, Z. and Shi, J. Q.},
  booktitle = {Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Long Papers)},
  pages = {44774--44790},
  year = {2026},
}

ICLR

Synergy and Diversity in CLIP: Enhancing Performance Through Adaptive Backbone Ensembling

C. Rodriguez-Opazo, E. Abbasnejad, D. Teney, Hamed Damirchi, E. Marrese-Taylor, and A. van den Hengel

In 13th International Conference on Learning Representations (ICLR), 2025

Bib HTML PDF

@inproceedings{rodriguezopazo2025synergy,
  title = {Synergy and Diversity in CLIP: Enhancing Performance Through Adaptive Backbone Ensembling},
  author = {Rodriguez-Opazo, C. and Abbasnejad, E. and Teney, D. and Damirchi, Hamed and Marrese-Taylor, E. and {van den Hengel}, A.},
  booktitle = {13th International Conference on Learning Representations (ICLR)},
  year = {2025},
}

ICRA

A Consistency-Based Loss for Deep Odometry Through Uncertainty Propagation

Hamed Damirchi, R. Khorrambakht, H. D. Taghirad, and B. Moshiri

In IEEE International Conference on Robotics and Automation (ICRA), 2023

arXiv Bib HTML

@inproceedings{damirchi2023consistency,
  title = {A Consistency-Based Loss for Deep Odometry Through Uncertainty Propagation},
  author = {Damirchi, Hamed and Khorrambakht, R. and Taghirad, H. D. and Moshiri, B.},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  pages = {11756--11762},
  year = {2023},
}