research

current projects

digital twins

we are building user mimic models from unlabeled data (e.g., chats with AI assistants, recordings). accurate mimic models can be used to simulate user behavior, including predicting their preferences and performing actions on their behalf.

publications

are large reasoning models interruptible?

tsung-han wu*, mihran miroyan*, david chan, trevor darrell, narges norouzi, joseph gonzalez

preprint

we evaluated large reasoning models (e.g., gpt-oss) under dynamic interruptible settings in math and coding tasks.

search arena: analyzing search-augmented llms

mihran miroyan*, tsung-han wu*, logan king, tianle li, jiayi pan, xinyan hu, wei-lin chiang, anastasios angelopoulos, trevor darrell, narges norouzi, joseph gonzalez

ICLR 2026

we collected and analyzed user-llm interactions and preferences in search-augmented chat settings.

parastudent: generating and evaluating realistic student code by teaching llms to struggle

mihran miroyan*, rose niousha*, joseph gonzalez, gireeja ranade, narges norouzi

preprint

we trained a student simulator to predict student-written code sequences.

analyzing pedagogical quality and efficiency of llm responses with ta feedback to live student questions

mihran miroyan*, chancharik mitra*, rishi jain, gireeja ranade, narges norouzi

SIGCSE 2025

we evaluated the pedagogical effectiveness and efficiency of a rag-powered llm assistant through expert teaching assistant evaluations.

retllm-e: retrieval-prompt strategy for question-answering on student discussion forums

chancharik mitra*, mihran miroyan*, rishi jain*, vedant kumud, gireeja ranade, narges norouzi

AAAI (EAAI) 2024

we built a retrieval-augmented prompting framework to generate context-aware responses for student questions in course forums.