avatar

Mingyu Jin

mingyu.jin404@gmail.com


About Me

I received my M.S. in Computer Science from Rutgers University–New Brunswick, where I worked on interpretability under the supervision of Professor Dimitris N. Metaxas. My research interests lie in Explainable AI (XAI), with a focus on large language models (LLMs). Before that, I obtained my bachelor’s degree from the University of Liverpool. I currently have the privilege of working on optimization and theory under the guidance of Professor Zhaoran Wang. I also work on interpretability with Professor Zhuoran Yang.

Research Interests

Mechanistic Interpretability, Sparse Autoencoders (SAEs), Science of LLMs, Learning Theory

Research Overview

My research aims to uncover the core mechanisms underlying reasoning in large language models, with the ultimate goal of enabling precise and comprehensive interventions in their reasoning processes.

LLM
Interpretability
Probing Representations
Concept Depth 【COLING ’25】More complex concepts emerge in deeper layers.
SAEs
Explain
SAGE 【EACL ’26 · Industry】Agents test and refine explanations of SAE features.
Steer
SAE-SSV 【EMNLP ’25】Steers model behavior through a small SAE subspace.
Circuit Discovery
All Circuits Lead to Rome 【ICML ’26】Distinct sparse circuits can support the same task.
Interpretable
Reasoning
Memory & Reasoning
Disentangling Memory and Reasoning 【ACL ’25】Separates knowledge recall from logical inference.
Understanding Reasoning Capacity
Minimal Parameter Budget for Implicit Reasoning 【ICML ’26】Links reasoning capacity to data complexity via a scaling law.
Chain-of-Thought
The Impact of Reasoning Step Length on LLMs 【ACL ’24 · Findings】Longer CoT demonstrations can improve reasoning without new information.
Discovering Internal Mechanisms
Massive Values in Attention: Keys to Contextual Knowledge Understanding 【ICML ’25】Large query and key activations enable contextual knowledge understanding.
Farther the Shift, Sparser the Representation 【NeurIPS ’26】Greater distribution shifts yield sparser hidden representations.

News

Internship

Services

Selected Publication

2026

2025

2024

Interest in Life


Powered by Jekyll and Minimal Light theme.