Yanjun Zhao

Work Experience

NEC Labs America

Research Intern · May 2026 – Aug 2026 · Princeton, NJ

Private LLM Agents with Dr. Yanchi Liu and Dr. Haifeng Chen.

Bilibili Inc.

Research Intern · May 2025 – Aug 2025 · Shanghai, China

Efficient LLM reasoning with Dr. Tianjiao Li.

Alibaba, DAMO Academy

Research Intern · Feb 2023 – Feb 2024 · Hangzhou, China

Data mining and time-series forecasting with Dr. Tian Zhou, Dr. Liang Sun, and Dr. Rong Jin.

News

Publications

* equal contribution

• Selected Preprints
• ReContext: Recursive Evidence Replay as LLM Harness for Long-Context Reasoning
  arXiv 2026
  Yanjun Zhao*, Ruizhong Qiu*, Tianxin Wei*, Yuanchen Bei, Zhining Liu, Lingjie Chen, Ismini Lourentzou, Hanghang Tong, Jingrui He
  Key Insight: Recursively replays evidence to support long-context reasoning with an LLM harness.
  Paper Code
  AgentLLMRetrieval
• Code as Agent Harness
  arXiv 2026
  Xuying Ning, Katherine Tieu, Dongqi Fu, Tianxin Wei, Zihao Li, Yuanchen Bei, Jiaru Zou, Mengting Ai, Zhining Liu, Ting-Wei Li, Lingjie Chen, Yanjun Zhao, and collaborators
  Key Insight: Treats executable code as a harness for agent deployment, evaluation, and improvement.
  Paper Project
  AgentLLM
• Conference
• Mem-Gallery: Benchmarking Multimodal Long-Term Conversational Memory for MLLM Agents
  ACL 2026 Main
  Yuanchen Bei, Tianxin Wei, Xuying Ning, Yanjun Zhao, Zhining Liu, Xiao Lin, Yada Zhu, Hendrik Hamann, Jingrui He, Hanghang Tong
  Key Insight: Evaluates memory retention, reasoning, and adaptation across multi-session multimodal conversations.
  Paper Code
  AgentMultimodal
• PaperMind: Benchmarking Agentic Reasoning and Critique over Scientific Papers in Multimodal LLMs
  ACL 2026 Findings
  Yanjun Zhao*, Tianxin Wei*, Jiaru Zou, Xuying Ning, Yuanchen Bei, Lingjie Chen, Simmi Rana, Wendy H. Yang, Hanghang Tong, Jingrui He
  Key Insight: Benchmarks multimodal LLM agents on reasoning about and critiquing scientific papers.
  Paper Code
  AgentMultimodal
• RAMEN: Robust Test-Time Adaptation of Vision-Language Models with Active Sample Selection
  CVPR 2026 findings
  Wenxuan Bao*, Yanjun Zhao*, Xiyuan Yang, Jingrui He
  Key Insight: Uses active retrieval and embedding-gradient caching for robust mixed-domain test-time adaptation.
  Paper
  MultimodalTest-timeRobustness
• SABER: Switchable and Balanced Training for Efficient LLM Reasoning
  AAAI 2026
  Kai Zhao*, Yanjun Zhao*, Jiaming Song, Shien He, Lusheng Zhang, Qiang Zhang, Tianjiao Li
  Key Insight: Enables user-controllable, token-budgeted reasoning through balanced reinforcement learning.
  Paper
  LLMEfficiency
• Learning a Zeroth-Order Optimizer for Fine-Tuning LLMs
  ICML 2026
  Kairun Zhang*, Haoyu Li*, Yanjun Zhao*, Yifan Sun, Huan Zhang
  Key Insight: Learns the optimizer itself to improve zeroth-order LLM fine-tuning.
  Paper Code
  LLMOptimization
• FZOO: Fast Zeroth-Order Optimizer for Fine-Tuning Large Language Models towards Adam-Scale Speed
  ICLR 2026
  Sizhe Dang*, Yangyang Guo*, Yanjun Zhao*, Haishan Ye, Xiaodong Zheng, Guang Dai, Ivor Tsang
  Key Insight: Accelerates zeroth-order fine-tuning with batched one-sided estimates and Rademacher perturbations.
  Paper Code
  LLMOptimizationEfficiency
• Second-Order Fine-Tuning without Pain for LLMs: A Hessian Informed Zeroth-Order Optimizer
  ICLR 2025
  Yanjun Zhao*, Sizhe Dang*, Haishan Ye, Guang Dai, Yi Qian, Ivor Tsang
  Key Insight: Uses a diagonal Hessian to stabilize and accelerate zeroth-order LLM fine-tuning.
  Paper Code
  LLMOptimizationEfficiency
• Does Vector Quantization Fail in Spatio-Temporal Forecasting? Exploring a Differentiable Sparse Soft-Vector Quantization Approach
  KDD 2025
  Chao Chen*, Tian Zhou*, Yanjun Zhao, Hui Liu, Rong Jin, Liang Sun
  Key Insight: Uses differentiable sparse soft-vector quantization for succinct spatio-temporal representations.
  Paper Code
  Time Series
• GCformer: An Efficient Solution for Accurate and Scalable Long-Term Multivariate Time Series Forecasting
  CIKM 2023
  Yanjun Zhao*, Ziqing Ma*, Tian Zhou, Mengni Ye, Liang Sun, Yi Qian
  Key Insight: Combines global convolution and a local Transformer branch for scalable long-term forecasting.
  Paper Code
  Time Series

Awards & Service

Gallery