Publications

Nona Ghazizadeh, E. Rahmati, M. Dehghani, and P. Piray. Beyond a Penalty: Regularization as Cognitive Control in Alignment. NeurIPS 2026 Position Paper Track, 2026. (Accepted)

Nona Ghazizadeh, E. Rahmati, A. S. Ziabari, K. Ahrabian, P. Piray, and M. Dehghani. DIALECT: Continuous-Space Communication Within and Across Language Models. Under review at ICLR 2027, 2026.

A. S. Ziabari*, Nona Ghazizadeh*, Z. Sourati, F. Karimi-Malekabadi, P. Piray, and M. Dehghani. Reasoning on a Spectrum: Aligning LLMs to System 1 and System 2 Thinking. COLM 2026.

Z. Sourati*, M. H. Wu*, Nona Ghazizadeh, J. Kaplan, M. Dehghani, and S. A. Nastase. Cognitive Expert Language Models Better Align with the Corresponding Brain Systems. Under review at ICLR 2027, 2026.

E. Rahmati, Nona Ghazizadeh, Z. Sourati, N. Rouhani, and M. Dehghani. Structural Abstraction as an Inductive Bias for Non-Stationary Language Model Training. Under review at ICLR 2027, 2026.

P. Golazizian, E. Rahmati, J. Trager, Z. Sourati, Nona Ghazizadeh, G. Chochlakis, J. J. Alcocer, K. Bennett, A. V. Devnani, P. Hejabi, H. G. Muttram, A. K. Padte, M. Saadatinia, C. Wu, A. S. Ziabari, M. Sierra-Arevalo, N. Weller, S. Narayanan, B. A. T. Graham, and M. Dehghani. The Subjectivity of Respect in Police Traffic Stops: Modeling Community Perspectives in Body-Worn Camera Footage. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics, 2026.

M. M. Abootorabi, Nona Ghazizadeh, A. Dalili, A. Ghahramani, M. Dehghani, and E. Asgari. AIMA at SemEval-2024 Task 10: History-Based Emotion Recognition in Hindi-English Code-Mixed Conversations. In Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024), Mexico City, Mexico, June 2024. Association for Computational Linguistics.

A. Ghahramani Kure, M. Dehghani, M. M. Abootorabi, Nona Ghazizadeh, S. A. Dalili, and E. Asgari. AIMA at SemEval-2024 Task 3: Simple Yet Powerful Emotion Cause Pair Analysis. In Proceedings of the 18th International Workshop on Semantic Evaluation (SemEval-2024), Mexico City, Mexico, June 2024. Association for Computational Linguistics.

M. Akhi and Nona Ghazizadeh. HDNA: A graph-based change detection in HTML pages (Deface Attack Detection), 2023.

Research Experience

Research Assistant at University of Southern California

Reasoning on a Spectrum: Aligning LLMs to System 1 and System 2 Thinking

This project examines LLM reasoning along a spectrum of fast, intuitive (System 1) and slow, deliberative (System 2) thinking, and evaluates how alignment affects reasoning behavior across benchmarks.

DIALECT: Continuous-Space Communication Within and Across Language Models

DIALECT explores communication through continuous representations between agents using the same or different language models. It transfers stochastic mixtures of token embeddings directly between same-model agents or through a learned bridge between different models, keeping both LLMs frozen. Evaluation spans 13 benchmarks in mathematical reasoning, knowledge question answering, and code generation.

Cognitive Expert Language Models Better Align with the Corresponding Brain Systems

This project investigates how language models specialized in cognitive domains align with corresponding brain systems, connecting language-model representations with cognitive neuroscience.

Mechanistic Representations of Structure-Mapping Theory

This project examines how LLMs represent the relationships described by Gentner’s Structure-Mapping Theory. It uses a multi-domain analogical reasoning dataset and contrastive activation analysis to study structural mapping representations.

Beyond a Penalty: Regularization as Cognitive Control in Alignment

This position paper connects reference-based regularization in LLM alignment with computational accounts of cognitive control. It argues for examining what regularization preserves and how its structure should vary across contexts.

Cognitive-Inspired Self-Evaluation Reasoning Framework

This project explores reflection token tagging and eye-tracking data as ways to identify points of cognitive reflection and inform self-evaluation in language-model reasoning.

Structural Abstraction as an Inductive Bias for Non-Stationary Language Model Training

This project studies structural abstraction as an inductive bias for language models learning from changing data. Abstraction-Augmented Training (AAT) is a replay-free online continual learning approach that uses abstraction-aware loss optimization to support stability and generalization.

The Subjectivity of Respect in Police Traffic Stops

This project examines how community perspectives shape interpretations of respect in LAPD body-worn camera footage. It combines a multi-perspective dataset with procedural justice-grounded rubrics to study perspective-aware modeling of respect.

Under the Supervision of Prof. Morteza Dehghani at Brain and Cognitive Science Department.

August 2024 – Present

Research Assistant at Sharif University of Technology

Emotion Recognition in Code-Mixed Conversations

This project studies emotion recognition in code-mixed conversations using pretrained language models and GRU networks to incorporate prior and future context alongside conversational sequence information.

Emotion-Cause Pair Extraction in Multimodal Conversations

This project examines emotion-cause relationships in multimodal conversations using causality matrices, transformer-based encoders, and prompt engineering.

Multilingual Medical and News Question Answering (RAG)

This project explores retrieval-augmented generation for multilingual medical, drug, and news question answering, using PubMed QA, PubMed summarization, and news datasets.

Bug Issue Ranking with Traceability and Call Graphs

This project investigates GitHub issue ranking through software traceability, call graphs, and similarity between NLP embeddings.

HDNA: Graph-Based Change Detection in HTML Pages

HDNA studies change detection in HTML pages through node identification and weighted tree representations of the DOM, including dynamically generated content.

Under the Supervision of Dr. Ehsaneddin Asgari and Prof. Abbas Heydarnoori at Computer Engineering Department.

August 2022 – July 2024

Research Assistant at University of New South Wales Sydney

Automated Medical Report Generation for FFA Images

This project explores automated report generation from Fundus Fluorescein Angiography (FFA) images using visual feature extraction, cross-modal memory, and reinforcement learning.

Under the Supervision of Dr. Imran Razzak

June 2023 – September 2023