LLM Agents
Reasoning and planning for agents that interact with tools and environments.
Hi, I’m Tianyun Zhong.
I am currently an M.S. student in Electronic Information at the Institute of Software, Chinese Academy of Sciences, and a member of the Chinese Information Processing Laboratory (CIP). I am advised by Professors Xianpei Han and Le Sun, and work closely with Professors Yaojie Lu, Hongyu Lin, and Ben He.
My research interests lie in large language models, long-horizon agents, and post-training. I am particularly interested in building reliable and capable agents for scientific discovery and complex decision-making.
01 / Research interests
Making agents reason, plan,
and work over long horizons.
Long-Horizon Agents
Optimization, continual decision-making, and resource management over extended trajectories.
Post-Training & Evaluation
Preference optimization, LLM-as-a-Judge evaluation, benchmarks, and synthetic data.
02 / Selected publications
LONGGAMEBENCH
Work conducted during internship at Alibaba
Are LLMs Long-Termist? Evaluating Long-Horizon Optimization in Roguelike Environments
A unified Agent–Environment benchmark for Long-Horizon Optimization, evaluated with 15 frontier LLMs across 6 agent harnesses and 1,000+ complete trajectories.
HYPOARENA
Work conducted during internship at Alibaba
Before the Action: Benchmarking LLMs on Prospective Hypothesis Discovery
Introduces Prospective Hypothesis Discovery and Retrospective Context Regression, with Arena-style pairwise evaluation for open-ended hypothesis generation.
Beyond Isolated Dots: Benchmarking Structured Table Construction as Deep Knowledge Extraction
XMC-Agent: Dynamic Navigation over Scalable Hierarchical Index for Incremental Extreme Multi-label Classification
Explainable Detection of Online Sexism by Ensembling Fine-tuning Language Models
03 / Research experience
OCT 2025 — PRESENT
Alibaba Group · Tongyi
Research Intern
Built scientific and real-world coding-agent benchmarks, including executable environments, programmatic evaluation, and Agent-as-a-Judge assessment. Research contributor to LongGameBench and HypoArena.
RESEARCH INTERNSHIP
Baidu · PaddleNLP
Large-Model-Based Vector Retrieval
Built LLM embedding training, retrieval, and evaluation pipelines; proposed layer-contribution-based pruning for efficient inference. Core implementation was merged into the official PaddleNLP repository.
SEP 2024 — PRESENT
Institute of Software, Chinese Academy of Sciences
M.S. in Electronic Information · Advisors: Xianpei Han, Le Sun
GPA 3.88/4.00 · Outstanding Student, UCAS.
04 / Academic service
Reviewer, ACL Rolling Review
2026