About me
Hi! I am a computer science Ph.D. student at UCLA in the PLUS Lab, advised by Prof. Nanyun (Violet) Peng. Previously, I received my master’s degree at the University of Tehran advised by Prof. Yadollah Yaghoobzadeh and Prof. Mohammad Taher Pilehvar. In summer 2025, I was a research intern at Adobe Research.
My research focuses on Interpretability for Robust and Controllable LLM Systems. Recently, I have been working on:
- Steering Mixture-of-Experts LLMs: detecting behavior-linked experts and (de)activating them at inference time to control safety and faithfulness without any fine-tuning (SteerMoE, ICLR 2026), and understanding how MoE routing works across languages (ICLR 2026).
- Retrieval and RAG: exposing the short, early, and literal biases that make dense retrievers outrank factual evidence (ACL 2025), and vision-centric evaluation of retrieval-augmented multimodal models (MRAG-Bench, ICLR 2025).
- Interpretability: token attribution in Transformers (ACL 2023, NAACL 2022), and probing the knowledge encoded in pre-trained language models (ACL 2022, BlackboxNLP@EMNLP 2021).
Feel free to email me or check my Curriculum Vitae.
Featured Research
Mohsen Fayyaz, Ali Modarressi, Hanieh Deilamsalehy, Franck Dernoncourt, Ryan Rossi, Trung Bui, Hinrich Schütze, Nanyun Peng
SteerMoE steers Mixture-of-Experts LLMs by detecting experts whose activation differs between paired inputs with opposite behaviors, then turning them on or off at inference time, with no retraining or weight updates. Across 11 benchmarks and 6 LLMs, steering improves safety by up to +20% and faithfulness by up to +27%. In adversarial mode, deactivating safety-linked experts sharply reduces safety, and combined with existing jailbreaks it bypasses all safety guardrails, exposing a new dimension of alignment faking hidden within experts.
[📝 paper] [💻 code]News
- Jan 2026: Two papers accepted to ICLR 2026! 🎉
- Sep 2025: Released SteerMoE, our work from my Adobe Research internship on steering MoE LLMs by (de)activating experts at inference time. [paper] [code]
- Jun 2025: Started a research internship at Adobe Research in Seattle.
- May 2025: Our paper "Collapse of Dense Retrievers: Short, Early, and Literal Biases Outranking Factual Evidence" has been accepted to ACL 2025 main conference! [paper] [dataset]
- Jan 2025: "MRAG-Bench: Vision-Centric Evaluation for Retrieval-Augmented Multimodal Models" has been accepted to ICLR 2025! [paper]
Show older news
- Jul 2024: New preprint arXiv "Evaluating Human Alignment and Model Faithfulness of LLM Rationale". [paper]
- Sep 2023: Started my Ph.D. in Computer Science at UCLA, advised by Prof. Nanyun (Violet) Peng.
- Jul 2023: Our paper "DecompX: Explaining Transformers Decisions by Propagating Token Decomposition" has been accepted to ACL 2023! [paper]
- Nov 2022: Our paper "BERT on a Data Diet: Finding Important Examples by Gradient-Based Pruning" has been accepted to ENLSP @ NeurIPS 2022! [paper]
- Apr 2022: Our paper "GlobEnc: Quantifying Global Token Attribution by Incorporating the Whole Encoder Layer in Transformers" has been accepted to NAACL 2022 main conference! [paper] [video]
- Mar 2022: Our paper "Metaphors in Pre-Trained Language Models: Probing and Generalization Across Datasets and Languages" has been accepted to ACL 2022 main conference! [paper] [video]
- Sep 2021: Our paper "Not All Models Localize Linguistic Knowledge in the Same Place: A Layer-wise Probing on BERToids’ Representations" has been accepted to BlackboxNLP @ EMNLP 2021! [paper]