Chiyu Ma

This is Chiyu (Henry) Ma, a third-year Computer Science Ph.D. student at Dartmouth College 🌲, where I am fortunate to be advised by 🌟 Prof. Soroush Vosoughi. Currently, I am also a Research Intern at Qwen Pilot, focusing on research in large-scale post-training of Large Language Models (LLMs).

Previously, I earned my M.S. in Statistical Science from Duke University 💙. As a member of the Interpretable Machine Learning Lab, I worked under the guidance of 🌟 Prof. Cynthia Rudin and collaborated with 🌟 Prof. Chaofan Chen (UMaine). I hold a B.S. in Statistics with Honors (Computational Finance concentration) from Carnegie Mellon University 🤖, where I conducted research on statistical analysis in biological applications advised by 🌟 Prof. Zach Branson.

 /   /   /  Google Scholar

My prior research centered on Interpretability in Computer Vision Models (CNNs and ViTs) and ensuring the reliability of LLM-as-a-Judge frameworks. Building on this foundation, I have pivoted to Large-scale Post-training, where I design algorithms specifically aimed at eliciting and enhancing the complex reasoning capabilities of LLMs.

[Jan. 28, 2026] Two papers accepted to ICLR 2026. Thanks to my excellent collaborators. See you in Rio!

[Sept. 2025] Two papers accepted to EMNLP 2025 and One paper accepted to NeurIPS 2025. See you in Suzhou and SD!

[Jun. 2025] Started as a research intern at Tongyi Lab Qwen Pilot Team, Hangzhou. Excited to meet friends in Hangzhou!

Some interesting papers about my work and research on LLM Post-training and Reinforcement Learning.

This paper proposes that log-probability difference is a more promising metric than entropy for evaluating how LLMs evolve during training, a claim supported by both empirical and theoretical analysis.

This paper sheds light on the distributional changes induced by RLVR and provides a granular, token-level lens for understanding and improving RL fine-tuning in LLMs.

My earlier works on the robustness of LLMs and LLM-as-Judge.

This paper provides counter-intuitive findings that multi-agent based LLM-as-Judges do not always provide reliable answers as people previously thought. We explored the phenomenon of bias amplifications in both Multi-agent Debate and LLm-as-Meta-Judge settings.

This paper provides a thorough analysis on how position bias spread over pair-wise and list-wise comparisons in the state-of-the-arts LLMs such as Gemini, GPT, and Claude.

This paper proposes knowledge continuity, a novel definition inspired by Lipschitz continuity which aims to certify the robustness of neural networks across input domains (such as continuous and discrete domains in vision and language, respectively).

My earlier works on Interpretability. Although I no longer work in this area, these projects represent a cherished part of my research journey.

This paper propopses an interpretable alorightms with prototypical pairs for tasks with continuous labels. This algorithm is further evaluated in Reward Prediction (RL settings) and Age predictions (Classical Regression tasks).

This paper propopses an interpretable alorightms on ViTs with Prototical Parts. By greedying matching algorithms, we decomposes the prototypical parts into small patches that can freely learn features to represent a more local feature.

This paper propopses an interpretable alorightms on CNNs that provides geometrically equivalent visualizations of prototypical concepts. This is the first work in prototyped based methods that extract concepts from training set.

Conference Reviewer:
ICLR (2025, 2026), ICML (2024–2026), NeurIPS (2024, 2025), AAAI (AISI Track 2023)

Workshop Reviewer:
NeurIPS IAI Workshop (2024), ICLR LLM Reasoning and Planning Workshop (2025)

Journal Reviewer:
Transactions on Machine Learning Research (TMLR)

• Ph.D. in Computer Science, Dartmouth College, 2028 (Expected)
• M.S. in Statistical Science, Duke University, 2023
• B.S. in Statistics (with Honors), Carnegie Mellon University, 2021

Design and source code from this cool guy.