Research
Publications and current work on LLM evaluation, agent systems, NLP, and production ML.
My current research and engineering work centers on evaluating and controlling learned systems in production. The common thread is simple: as models become more capable, the surrounding harness, measurement, and operating loop become more important.
Recent work includes LLM robustness, coding-agent evaluation, agent harness design, LLM-as-judge systems, multilingual NLP, retrieval, model compression, and production ML systems.
Current interests
- Unified views of machine learning: cases where spectral methods, probabilistic models, kernels, and neural networks collapse to the same underlying object (similarity, message passing, KL divergence).
- Evaluation: measurement design, LLM-as-judge methods, agent trajectory evaluation, and the long tail of the metric went up but the product got worse.
- Agents and planning: how learned decision-makers behave in adversarial or long-horizon settings, and what loop-level tooling they need.
- AI and security: the verification asymmetry between offense and defense, and what that implies for where agents actually land.
- Probabilistic ML and complex systems: variational inference, deep generative models, stochastic processes, and the occasional detour into multifractal signal analysis.
- On-device and edge ML: compressing and factorizing transformers for strict memory and latency budgets.
For systems work around agents and evaluation, see Looplet and the selected work on About.
Selected publications
- J. Niu et al. (incl. H. Saghir). Llama See, Llama Do: Contextual Entrainment and Distraction in LLMs. ACL 2025.
- M. Mirgbagheri, H. Saghir, T. Chau. Mimicking Linguistic Features of Atypical Speech Transcripts. IEEE TASLP, 2025.
- L. Hebert et al. (incl. H. Saghir). Robust Candidate Generation for Entity Linking on Short Social Media. WNUT @ COLING 2022.
- H. Saghir, S. Choudhary, S. Eghbali, C. Chung. Factorization-Aware Training of Transformers for NLU on the Edge. InterSpeech 2021.
- H. Tu, S. Choudhary, H. Saghir, R. McGowan. Multilingual Neural Language Models for On-Device NLU. The Web Conference (WWW) 2021.
- P. Xu, H. Saghir, et al. A Cross-Domain Transferable Neural Coherence Model. ACL 2019.
- H. Saghir, A. Dupuis, T. Chau, A. Kushki. Atypical autonomic nervous system complexity accompanies social cognition task performance in ASD. Research in Autism Spectrum Disorders, 2017.
- H. Saghir, T. Chau, A. Kushki. Clustering of time-evolving scaling dynamics in a complex signal. Physical Review E, 2016.
Full list on Google Scholar.