Runyuan He

Gateway, UC Berkeley

Berkeley, CA 94720

I am a senior undergraduate researcher in Computer Science at the UC Berkeley Sky Computing Lab, completing my undergraduate studies at The Chinese University of Hong Kong, Shenzhen. I conduct research under the supervision of Prof. Alvin Cheung, with strong interests in database systems and large language models. I am actively seeking Ph.D. opportunities.

My research spans LLM agents, machine learning systems, and databases. I develop methods for generating coding tasks and improving agents, alongside systems that make LLM workloads more efficient and database testing more reliable.

Recent work includes FrontierSmith, accepted as a NeurIPS 2026 Spotlight (0.95%), Combee for scaling prompt learning, and Continuum for multi-turn agent scheduling.

Open-source models from my recent work are available on Hugging Face, and code/preprints are on GitHub, Google Scholar, and DBLP.

Research Interests

LLM agents: open-ended coding, training data synthesis, and self-improvement.
Machine learning systems: efficient inference and scheduling for multi-turn agents.
Database systems: automated testing and query optimization with LLMs.

Selected Achievements

🥈 Silver Medalist - The 38th National Olympiad in Informatics (NOI 2021)
🌍 24th Place - The 48th ICPC World Final 2024
🏆 8th Place - The 48th ICPC World Final Huawei Challenge 2024
🥇 Gold Medalist - ICPC Regional 2023/2024

selected publications

FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale

Runyuan He^*, Qiuyang Mang^*, Shang Zhou, Kaiyuan Liu, Hanchen Li, Huanzhi Mao, and 11 more authors

NeurIPS 2026 Spotlight (0.95%)

A system that synthesizes open-ended coding problems, test cases, and verifiers for training and evaluating language models.

Abs arXiv Bib

Accepted as a Spotlight at NeurIPS 2026 (0.95%).

Open-ended coding remains a weak spot for LLMs, largely because open-ended training problems are scarce. FrontierSmith is an automated system that synthesizes such problems by evolving competitive programming tasks through modifying goals, restricting outputs, and generalizing inputs. The system employs a quantitative idea divergence metric to select problems that elicit genuinely diverse approaches, and generates corresponding test cases and verifiers. Models trained on FrontierSmith data show significant improvements on FrontierCS and ALE-bench benchmarks.
@inproceedings{he2026frontiersmith, title = {FrontierSmith: Synthesizing Open-Ended Coding Problems at Scale}, author = {He, Runyuan and Mang, Qiuyang and Zhou, Shang and Liu, Kaiyuan and Li, Hanchen and Mao, Huanzhi and Zhang, Qizheng and Li, Zerui and Peng, Bo and Cheng, Lufeng and Fu, Tianfu and Wang, Yichuan and Chai, Wenhao and Shang, Jingbo and Dimakis, Alex and Gonzalez, Joseph E and Cheung, Alvin}, booktitle = {Advances in Neural Information Processing Systems (NeurIPS)}, year = {2026}, }
Combee: Scaling Prompt Learning for Self-Improving Language Model Agents

Hanchen Li^*, Runyuan He^*, Qizheng Zhang, Changxiu Ji, Qiuyang Mang, Xiaokun Chen, and 8 more authors

COLM 2026

A framework for parallel prompt learning that helps language model agents improve from their executions.

Abs arXiv Bib

We present Combee, a framework for scaling prompt learning for self-improving language model agents. Existing prompt learning methods struggle in parallel settings, experiencing quality degradation when learning from multiple agent executions simultaneously. Combee leverages parallel scans and an augmented shuffle mechanism along with dynamic batch size control to maintain quality while processing many agents concurrently, achieving up to 17x speedup over previous methods with comparable or better accuracy.
@inproceedings{li2026combee, title = {Combee: Scaling Prompt Learning for Self-Improving Language Model Agents}, author = {Li, Hanchen and He, Runyuan and Zhang, Qizheng and Ji, Changxiu and Mang, Qiuyang and Chen, Xiaokun and Agrawal, Lakshya A and Liao, Wei-Liang and Yang, Eric and Cheung, Alvin and Zou, James and Olukotun, Kunle and Stoica, Ion and Gonzalez, Joseph E}, booktitle = {Conference on Language Modeling (COLM)}, year = {2026}, }
Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live

Hanchen Li^*, Runyuan He^*, Qiuyang Mang^*, Qizheng Zhang, Huanzhi Mao, Xiaokun Chen, and 4 more authors

ICLR 2026 LLA Workshop

A serving system that schedules multi-turn LLM agents using time-to-live policies for KV caches.

Abs arXiv Bib

We present Continuum, a serving system to optimize job completion time for multi-turn agent workloads by introducing time-to-live mechanism for KV cache retaining. For LLM request that generates a tool call, Continuum selectively pins the KV cache in GPU memory with a time-to-live value determined by considering both the reload cost and ordering preserve benefit of retaining KV cache.
@inproceedings{li2025continuum, title = {Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live}, author = {Li, Hanchen and He, Runyuan and Mang, Qiuyang and Zhang, Qizheng and Mao, Huanzhi and Chen, Xiaokun and Zhou, Hangrui and Cheung, Alvin and Gonzalez, Joseph and Stoica, Ion}, booktitle = {ICLR LLA Workshop}, year = {2026}, }
Automated Discovery of Test Oracles for Database Management Systems Using LLMs

Qiuyang Mang, Runyuan He, Suyang Zhong, Xiaoxuan Liu, Huanchen Zhang, and Alvin Cheung

SIGMOD 2026

A database testing framework that uses LLMs and formal verification to discover sound test oracles.

Abs DOI arXiv Bib

This paper explores the use of large language models (LLMs) to automate the discovery and instantiation of test oracles, addressing a long-standing bottleneck towards fully automated DBMS testing. We introduce Argus, a novel framework built upon the core concept of the Constrained Abstract Query. Argus uses LLMs to generate pairs of SQL skeletons that are asserted to be semantically equivalent, then formally proven using a SQL equivalence solver to ensure soundness.
@article{mang2025argus, title = {Automated Discovery of Test Oracles for Database Management Systems Using LLMs}, author = {Mang, Qiuyang and He, Runyuan and Zhong, Suyang and Liu, Xiaoxuan and Zhang, Huanchen and Cheung, Alvin}, journal = {Proceedings of the ACM on Management of Data (SIGMOD)}, volume = {4}, number = {3}, pages = {Article 140}, year = {2026}, doi = {10.1145/3802017}, }