Hi, I'm Hoda

PhD student at Linköping University in the AI division (AIICS), working on LLM reasoning and alignment.

My research focuses on understanding how large language models think, from analyzing their internal geometry to improving how they learn through reinforcement learning and verifiable rewards.

Website: hodfa840.github.io

What I work on

• LLM interpretability: detecting reasoning failures from hidden-state geometry (VANE)
• LLM safety: type-aware unlearning of memorized secrets (PSTU)
• RL for language models: GRPO, DPO, SDPO, and flow matching approaches to LLM fine-tuning
• Verifiable rewards: grammar-guided RLVF for low-resource languages (geysir)

Selected work

Project	What it does
VANE	Predicts LLM reasoning correctness from hidden-state trajectory geometry. No training required.
PSTU	Removes memorized secrets from LLMs via per-type saliency in weight space. Training-free, runs in seconds.
geysir	RLVF for Icelandic NLP — Greynir grammar parser as a verifiable reward oracle. Compares GRPO, DPO, SDPO, and Delta-DPO.
ml-observability-argus	Production MLOps platform: drift detection, automated retraining & root-cause analysis on NYC taxi data.

Get in touch

hoda.fakhar@liu.se