Dehai Min   闵德海

👋 I am a second-year CS Ph.D. student (Aug. 2025–) at University of Illinois Chicago (UIC), advised by Prof. Philip S. Yu (Fellow of ACM, IEEE and AAAS). I am currently a Student Researcher at Microsoft CoreAI - GitHub Copilot team. Previously, I was a Research Scientist Intern at ByteDance (San Jose) Doubao/Seed Agent MaaS Team in Summer 2026. I received my M.S. from Southeast University, advised by Prof. Guilin Qi, and my B.S. from Hefei University. I am deeply grateful for the previous guidance from Prof. Chenyu You from Stony Brook University.

My current research interests include:

1. Agent Trajectory Evaluation, Environment Reconstruction & Verifier Synthesis
2. Retrieval-Augmented Generation (RAG) & Information Retrieval
3. Efficient LLMs & LLM Uncertainty Quantification

Previously, during my master's studies, I worked on LLMs with heterogeneous data (text, graphs, and tables).

I am always open to research discussions — feel free to reach out!

🎉 News

• [2026.11] New paper (co-first): Veritas — A Verifier Generation Harness for Agent Tasks. 📄 Paper
• [2026.10] Invited by Apple to give a research talk on TraceDance! 🎤
• [2026.10] Excited to share that I'm joining Microsoft’s CoreAI - GitHub Copilot team as a student researcher, working on coding agents post-training and evaluation!
• [2026.09] New paper released: TraceDance — An Automated System for Building Agent Behavior Benchmarks from Real-World Agent Deployment Traces. Paper and project website are now available! 📄 Paper / 🌐 Website
• [2026.08] Two papers accepted: TSRouter by COLM 2026 (📄 Paper / 📦 Code) & EpiQAL by EMNLP 2026! 🎉
• [2026.06] New paper: Stop When Reasoning Converges — Semantic-Preserving Early Exit for Reasoning Models (PUMA). 📄 Paper / 📦 Code / 🤗 Models & Datasets
• [2026.06] New paper (co-first): CorVer — Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering. 📄 Paper / 📦 Code
• [2026.05] Moved to San Jose, CA (Bay Area) to start as a Research Scientist Intern @ ByteDance, working with the Doubao/Seed Agent MaaS Team at AML-Volcano Engine-Ark.
• [2026.05] Two papers (Latent Thoughts Tuning & MVI-Bench) accepted by ICML 2026! 🎉
• [2026.04] One paper (QuCo-RAG) accepted by ACL Findings 2026! 📄 Paper / 📦 Code / Covered by 🔥 新智元. Thanks to Dr. Cheng for her guidance!
• [2026.02] New paper on latent reasoning: Latent Thoughts Tuning — Bridging Context and Reasoning with Fused Information in Latent Tokens. 📄 Paper / 📦 Code
• [2025.12] Selected as Outstanding Reviewer of 2024–2025 for IEEE TNNLS! 📦 Certificate
• [2025.12] New paper released: MVI-Bench — Evaluating Robustness to Misleading Visual Inputs in LVLMs. 📄 Paper / 📦 Code
• [2025.08] Moved to Chicago 🇺🇸.
• [2025.01] 🎉 One paper (UniHGKR) accepted by NAACL 2025 as Oral, with the highest meta-review score of 5! 📦 Code / 🤗 Models & Datasets
• [2024.12] One paper accepted by AAAI 2025!

🎓 Education

University of Illinois Chicago (Chicago, IL, USA) , Ph.D. , 08/2025 - present

Southeast University (Nanjing, China) , M.S. , 09/2022 - 06/2025

Hefei University (Hefei, China) , B.S. , 09/2017 - 06/2022

🧑‍🏫 Internship

Microsoft , 10/2026 -
Student Researcher , United States · Remote
CoreAI - GitHub Copilot Team.
Coding agents post-training and evaluation.

ByteDance (TikTok) , 05/2026 - 08/2026
Research Scientist Intern , San Jose, California
Doubao/Seed Agent Model-as-a-Service (MaaS) Team, AML / Volcano Engine-Ark.
Data flywheel from deployment trajectories back into Doubao/Seed agent evaluation and environment.

📝 Publications and Preprints

(* denotes equal contribution.)

First-author Publications

Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models Preprint 2026

Dehai Min*, Giovanni Vaccarino*, Huiyi Chen, Yongliang Wu et al.

Efficient Reasoning 📦 Code / 🤗 Models & Datasets /

QuCo-RAG: Quantifying Uncertainty from the Pre-training Corpus for Dynamic Retrieval-Augmented Generation ACL Findings 2026

Dehai Min, Kailin Zhang, Tongtong Wu, Lu Cheng.

RAGUncertainty Quantification 📦 Code /

UniHGKR: Unified Instruction-aware Heterogeneous Knowledge Retrievers NAACL 2025, Oral Highest Meta-Review Score: 5

Dehai Min, Zhiyang Xu, Guilin Qi, Lifu Huang et al.

Information Retrieval 📦 Code / 🤗 Models & Datasets

Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering Preprint 2026

Shicheng Fan*, Haochang Hao*, Dehai Min*, Weihao Liu et al.

Post-Training 📦 Code

Can ChatGPT replace traditional KBQA models? An in-depth analysis of the question answering performance of the GPT LLM family ISWC 2023, Oral 370 Citations

Yiming Tan*, Dehai Min*, Yu Li, Wenbo Li et al.

LLM Evaluation 📦 Code / 🎤 Slides

Selected Collaborative Publications

TSRouter: Dynamic Modality-Model Selection for Time Series Reasoning COLM 2026

Fangxu Yu, Tao Feng, Dehai Min, Lu Cheng et al.

Model Routing 📦 Code

Latent Thoughts Tuning: Bridging Context and Reasoning with Fused Information in Latent Tokens ICML 2026

Weihao Liu, Dehai Min, Lu Cheng.

Latent Reasoning 📦 Code

MVI-Bench: A Comprehensive Benchmark for Evaluating Robustness to Misleading Visual Inputs in LVLMs ICML 2026

Huiyi Chen, Jiawei Peng, Dehai Min, Changchang Sun et al.

Multimodal Evaluation 📦 Code

🏅 Competitions

( I was an active participant in programming competitions during my undergraduate years. )

• ACM-ICPC (International Collegiate Programming Contest) , Asian Regional Contest (including all participating universities in mainland China), 2019-2020, 2 times Silver Medal. individual / team
• CCPC-Finals (China Collegiate Programming Contest National Finals), 2019, 21st place, Silver Medal
• Anhui Province Collegiate Programming Contest, First Prize (2 times, 2019, 2020), Champion🏆 (1 time, 2020)
• Codeforces Rating: 2108 (Master), link

🎤 Invited Talks

Invited research talk on TraceDance: From Deployment Traces to Agent Behavior Benchmarks 🎤

Apple · 2026.10

👨‍💻 Academic Services

• Outstanding Reviewer of 2024–2025, IEEE TNNLS 📦 Certificate

• Reviewer of ACL, EMNLP, NAACL, ICML, NeurIPS, ICLR, AAAI

• Reviewer of IEEE Transactions on Neural Networks and Learning Systems (TNNLS)