Longtian Bao

Ph.D. Student in Computer Science at the University of Chicago

I am a Ph.D. student in Computer Science at the University of Chicago and a member of the Chicago Human+AI (CHAI) Lab. My work bridges Natural Language Processing and Machine Learning, with a focus on reinforcement fine-tuning and self-evolving curricula for LLM reasoning, long-context understanding, information retrieval, and efficient state-space models like Mamba-2.

Previously, I completed my M.S. in Computer Science and B.S. in Computer Engineering at UC San Diego, where I was a researcher at the Laboratory for Emerging Intelligence advised by Prof. Leon Bergen and Prof. Ramamohan Paturi. I have experience building full-stack educational platforms and autonomous robotic systems.

Longtian Bao

Selected Publications

* denotes equal contribution

Question Begets Question: Self-Evolving Curriculum for Reinforcement Fine-Tuning on Competition Mathematics
Longtian Bao*, Jianyou Wang*, Yang Zhang*, Youze Zheng*, Ramamohan Paturi
arXiv preprint, 2026
Paper
Multi-Agent LLM Serving Is a Full-Stack Problem: A Small Step Towards Algorithm–System Co-Design
Zhongkai Yu, Jinya Jiang, Longtian Bao, Junxia Cui, Jingyi Zhang, Zhengan Cheng, Zaifeng Pan, Jingbo Shang, Yufei Ding
Under review, 2026
DeepImagine: Learning Biomedical Reasoning via Successive Counterfactual Imagining
Youze Zheng*, Jianyou Wang*, Yuhan Chen, Matthew Feng, Longtian Bao, Hanyuan Zhang, Maxim Khan, Aditya K. Sehgal, Christopher D. Rosin, Umber Dube, Ramamohan Paturi
arXiv preprint, 2026
Paper
CT Open: An Open-Access, Uncontaminated, Live Platform for the Open Challenge of Clinical Trial Outcome Prediction
Jianyou Wang*, Youze Zheng*, Longtian Bao, Hanyuan Zhang, Qirui Zheng, Yuhan Chen, Yang Zhang, Matthew Feng, Maxim Khan, Aditya K. Sehgal, Christopher D. Rosin, Ramamohan Paturi, Umber Dube, Leon Bergen
arXiv preprint, 2026
Paper
Single-Pass Document Scanning for Question Answering
Weili Cao*, Jianyou Wang*, Youze Zheng*, Longtian Bao*, Qirui Zheng, Taylor Berg-Kirkpatrick, Ramamohan Paturi, Leon Bergen
Conference on Language Modeling (COLM), 2025 Oral Spotlight, Top 2%
Paper Code
Measuring Risk of Bias in Biomedical Reports: The RoBBR Benchmark
Jianyou Wang*, Weili Cao*, Longtian Bao, Youze Zheng, Gil Pasternak, Kaicheng Wang, Xiaoyue Wang, Ramamohan Paturi, Leon Bergen
Empirical Methods in Natural Language Processing (EMNLP), 2025
Paper Code

Research Funding

ACCESS Discover Compute Allocation
Principal Investigator · Longtian Bao
August 4, 2026 – August 3, 2027 · Project CIS261451

Self-Evolving Curricula for Reinforcement Fine-Tuning of Language Models on Competition Mathematics. Received in-kind computing support through the NSF-supported ACCESS program, with 750,000 ACCESS credits initially available.

Experience

Ph.D. Student
Department of Computer Science, University of Chicago
Reinforcement fine-tuning for LLM reasoning; multi-agent LLM serving
Mar 2026 – Present
  • Designed Question-begets-Question (QbQ), a scalable data-generation procedure in which a teacher model transforms competition-math problems into diverse variants probing the same skills, enabling reinforcement fine-tuning without oracle reasoning traces.
  • Proposed a self-evolving curriculum that re-seeds QbQ each round from problems the current checkpoint can mostly solve, breaking the static-training ceiling and lifting Qwen2.5-Math-7B from 5.6% to 16.5% pass@1 on AIME with no saturation after 20 rounds.
  • Introduced a unified typed-graph representation of multi-agent LLM systems and serving optimizations achieving up to a 5× speedup while using 54–89% fewer tokens on real GPUs, showing that decode dominance makes phase-aware algorithm–system co-design essential.
Research Assistant
Laboratory for Emerging Intelligence, UCSD
Advisors: Prof. Leon Bergen and Prof. Ramamohan Paturi
Apr 2024 – Mar 2026
  • Co-developed CT Open, an open-access, uncontaminated, live platform for the open challenge of clinical trial outcome prediction, enabling rigorous evaluation of LLMs on prospective post-training-cutoff trials.
  • Contributed to DeepImagine, which teaches LLMs biomedical reasoning by training them to infer how observed clinical-trial results change under controlled perturbations, combining supervised fine-tuning on strict counterfactual pairs with RL from verifiable rewards.
  • Re-architected a Mamba-2 State-Space Model with a classification head, creating a subquadratic model that preserves global context for efficient sentence retrieval in documents up to 256k tokens, with a link-based synthetic data pipeline for long-context QA training.
  • Constructed a fine-grained sentence-retrieval benchmark from 41 existing QA datasets via an LLM-assisted annotation-validation pipeline.
  • Developed the RoBBR benchmark for risk-of-bias assessment in biomedical research from expert judgments on 500+ studies, with subtasks that disentangle the retrieval and reasoning abilities of language models.
Instructional Assistant
UC San Diego
2021 – 2024
  • Data Structures (CSE 12): Spring 21, Winter 23, Winter 24
  • Accel. Intro to Programming (CSE 11): Winter 22, Fall 22, Spring 23
  • Computer Organization (CSE 30): Fall 23
  • Software Tools & Techniques Lab (CSE 15L): Spring 21

Projects

UCSD Autograder

2023–2024
Engineered a full-stack educational platform used by over 1,000 students. Developed containerized autograders and real-time ticket queues for office hours.
Full Stack Docker Database

Autonomous AI Racer

Winter 2024
Built a 1/10-scale autonomous vehicle with Jetson Nano, LIDAR, and OAK-D. Trained behavioral cloning models using TensorFlow to navigate racetracks.
Robotics Computer Vision TensorFlow

DaCe Reproducibility

Fall 2022
Evaluated the DaCe high-performance framework on NPBench using AMD hardware. Ported and debugged NumPy benchmarks to analyze performance portability.
HPC Python Benchmarking

Skills

Languages & Core
Python (Advanced) C/C++ Java TypeScript Bash
Machine Learning & AI
PyTorch TensorFlow Transformers Mamba-2 SSMs Long-context QA OpenCV
Systems & DevOps
Docker AWS (EC2/S3) Linux Git PostgreSQL Distributed Training
Tools & Hardware
LLM APIs Weights & Biases NVIDIA Jetson LIDAR React Flask/FastAPI