About me 👋

This is Yutong Song, a forth-year PhD student at the University of California, Irvine and I will graduate in 2026. I am now living in Irvine with my cute cat Hula. I am working in Health SciTech Group, under the mentorship of Prof.Amir M. Rahmani and Prof.Nikil Dutt.

Working Experience 🔑

Jun 2025 - Dec 2025: Research Scientist Intern in TikTok

Research Intersts 👾

Large language models, Health AI, Personalized text generation, LLM Agent, LLM post-training, User modeling, Rein- forcement learning, Human preference alignment, Agentic RAG, Speech language model, Spoken dialog system, Automatic Speech Recognition

News 🥳

Apr 2026: My paper DemMA: Dementia Multi-Turn Dialogue Agent with Expert-Guided Reasoning and Action Simulation is accpeted by ACL 2026 (findings), see you in San Diego!
Jan 2026: My paper MedSpeak: A knowledge Graph-aided ASR Error Correction Framework for Spoken Medical QA is accpeted by ICASSP 2026, see you in Barcelona, Spain!
May 2025: I will start my internship in Tiktok Inc. as a Machine Learning Engineer Intern from Jun to Dec 2025.
Mar 2025: I went to Philadephia and attended AAAI 2025! I gave a presentation of my recent work about RAG agent on dementia care. You can view my presentation slides here: Dementia-plan.
Jan 2024: I am happy to join the Noyce team about dementia care and empathy-centered conversatioanl system. This is a collaboration between UC Irvine, Purdue University, SmartForward Inc., and NaviGAIT. You can view our poster here.

Awards and Honors 🏆

Tyler’s Tribe Foundation Endowment for ALS Researc, Tyler’s Tribe Foundation 2025
ICS Fellowship, UCI 2022 and 2024
Graduate Research Fellowship, SDSU 2022
The Best Paper Award, IEEE CAMAD 2022

Selected Publications 🚀

DemMA: Dementia Multi-Turn Dialogue Agent with Expert-Guided Reasoning and Action Simulation

Yutong Song, Jiang Wu, Kazi Sharif, Pengfei Zhang, Wenjun Huang, Honghui Xu, Nikil Dutt, Amir Rahmani

DemMA is an LLM-based generative agent that simulates dementia patient interactions using chain-of-thought distillation and explicit action labels, supported by a dialogue dataset spanning multiple dementia subtypes.

Paper Poster Code

CARD: Cluster-level Adaptation with Reward-guided Decoding for Personalized Text Generation

Yutong Song, Jiang Wu, Weijia Zhang, Chengze Shen, Shaofan Yuan, Weitao Lu, Jian Wang, Amir Rahmani, Nikil Dutt, Yu Wang

CARD is a hierarchical personalization text generation framework that combines group-level adaptation with user-specific decoding to generate text better aligned with individual preferences.

Paper Slides Code

Styles+ Persona-plug= Customized LLMs

Yutong Song, Jiang Wu, Shaofan Yuan, Chengze Shen, Jian Wang, Amir Rahmani, Nikil Dutt, Yu Wang

To balance implicit personalization and explicit style, we formulate personalization as a distributional residual and propose PsPLUG, a lightweight plug-in that preserves implicit user personalization while following explicit style instructions in customized LLMs.

Paper Poster Code

MedSpeak: Knowledge Enhanced ASR Error Correction for Spoken Medical QA

Yutong Song, Chenxi Lyu, Ehsan Khatibi, Pengyu Zhang, Nikil Dutt, Amir M. Rahmani

Phonetic + semantic KG retrieval repairs hallucinated medical terminology before QA reasoning, lifting exact match and BLEU on internal SQA benchmarks.

Paper Poster Code