Qifu (Danny) Wen

qfwen@bu.edu

Boston University

Boston, MA, USA

I am a visiting researcher in the ML Group at Boston University, advised by Prof. Reza Rawassizadeh. I completed my M.S. in Computer Science at BU and my B.S. in Mathematics at NYU Shanghai.

My research spans ML training efficiency — GradES achieves 1.2–1.57× faster LoRA fine-tuning by selectively freezing transformer components at convergence — federated learning (gradient stable rank compression), and reinforcement learning. My recent RL work (ECRL) started as an epistemic exploration method and ended up revealing that ensemble variance functions as a controllable regularizer. I have a strong mathematical background in measure-theoretic probability and stochastic processes.

I am a reviewer for ICML. I am currently applying to PhD programs for Fall 2026.

Outside of my research, I enjoy playing basketball, hiking and playing rimworld.

Feel free to reach out if you’d like to chat about research or collaboration!

news

May 01, 2026	ECRL and Pose as a Lens submitted to NeurIPS 2026.
Feb 01, 2026	Paper “Estimation of Distribution Parameters” published in Statistical Papers (Springer), Vol. 67, Article 31.
Jun 01, 2025	Paper “From Clicks to Conversations” published in International Journal of Human-Computer Interaction (Taylor & Francis, October 2025).
May 01, 2025	Paper “Estimation of Distribution Parameters” accepted at Statistical Papers (Springer, JCR Q2).
Mar 15, 2025	OceanEnv submitted to IROS 2026.

selected publications

From Clicks to Conversations: Evaluating the Effectiveness of Conversational Agents in Statistical Analysis

Qifu Wen, Prishita Kochhar, Sherif Zeyada, and 2 more authors

International Journal of Human-Computer Interaction, 2025

Abs DOI Bib HTML

We compared conversational agents to traditional GUIs for statistical analysis through a user study with 51 participants. Conversational agents significantly outperformed GUI-based software on accuracy, completion time, and user satisfaction.
@article{wen2025clicks, title = {From Clicks to Conversations: Evaluating the Effectiveness of Conversational Agents in Statistical Analysis}, author = {Wen, Qifu and Kochhar, Prishita and Zeyada, Sherif and Javaheri, Tahereh and Rawassizadeh, Reza}, journal = {International Journal of Human-Computer Interaction}, publisher = {Taylor \& Francis}, year = {2025}, doi = {10.1080/10447318.2025.2561760}, }
Estimation of Distribution Parameters by Mean Absolute Deviations of a Truncated Distribution Using Quantile Functions

Eugene Pinsky and Qifu Wen

Statistical Papers, Feb 2026

Co-first authors

Abs Bib

Novel parameter estimation method using mean absolute deviations around quartiles for truncated distributions without moments or variance, providing closed-form solutions comparable to MLE for Lévy, Cauchy, and Pareto distributions.
@article{pinsky2026estimation, title = {Estimation of Distribution Parameters by Mean Absolute Deviations of a Truncated Distribution Using Quantile Functions}, author = {Pinsky, Eugene and Wen, Qifu}, journal = {Statistical Papers}, publisher = {Springer}, year = {2026}, month = feb, volume = {67}, number = {31}, note = {Co-first authors}, }
GradES: Significantly Faster Training in Transformers with Gradient-Based Early Stopping

Qifu Wen, Xi Zeng, Zihan Zhou, and 4 more authors

2025

Preprint; targeting ICLR 2026

Abs arXiv Bib

GradES is a gradient-based early stopping algorithm that individually freezes transformer components when their gradients fall below a convergence threshold. Achieves 1.2–1.57× faster LoRA fine-tuning on Llama-3.1-8B and Qwen3-14B.
@article{wen2025grades, title = {GradES: Significantly Faster Training in Transformers with Gradient-Based Early Stopping}, author = {Wen, Qifu and Zeng, Xi and Zhou, Zihan and Liu, Shuaijun and Hosseinzadeh, Mehdi and Su, Ningxin and Rawassizadeh, Reza}, year = {2025}, note = {Preprint; targeting ICLR 2026}, }

One Ocean, All Tasks: A Holistic Simulation Environment for Marine Robotics

Shuaijun Liu, Qifu Wen, Xiang Chen, and 2 more authors

In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026

Under review

Abs Bib

@inproceedings{liu2026ocean,
  title = {One Ocean, All Tasks: A Holistic Simulation Environment for Marine Robotics},
  author = {Liu, Shuaijun and Wen, Qifu and Chen, Xiang and Hu, Xinyu and Su, Ningxin},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year = {2026},
  note = {Under review},
}

Pruning and Quantization Impact on Graph Neural Networks

Khatoon Khedri, Reza Rawassizadeh, Qifu Wen, and 1 more author

Journal of Machine Learning Research, 2025

Under review

Abs arXiv Bib

Empirical evaluation of pruning and quantization compression techniques on graph neural networks across multiple tasks and datasets.
@article{khedri2025pruning, title = {Pruning and Quantization Impact on Graph Neural Networks}, author = {Khedri, Khatoon and Rawassizadeh, Reza and Wen, Qifu and Hosseinzadeh, Mehdi}, journal = {Journal of Machine Learning Research}, year = {2025}, note = {Under review}, }