CORE MACHINE LEARNING
From Neurons to Neutrons: A Case Study in Mechanistic Interpretability
July 21, 2024
Abstract
Mechanistic Interpretability (MI) promises a path toward fully understanding how neural networks make their predictions. Prior work demonstrates that even when trained to perform simple arithmetic, models can implement a variety of algorithms (sometimes concurrently) depending on initialization and hyperparameters. Does this mean neuron-level interpretability techniques have limited applicability? We argue that high-dimensional neural networks can learn lowdimensional representations of their training data that are useful beyond simply making good predictions. Such representations can be understood through the mechanistic interpretability lens and provide insights that are surprisingly faithful to human-derived domain knowledge. This indicates that such approaches to interpretability can be useful for deriving a new understanding of a problem from models trained to solve it. As a case study, we extract nuclear physics concepts by studying models trained to reproduce nuclear data.
AUTHORS
Written by
Ouail Kitouni
Samuel Pérez Díaz
Sokratis Trifinopoulos
Mike Williams
Niklas Nolte
Publisher
icml
Research Topics
Core Machine Learning
Related Publications
November 18, 2025
RESEARCH
CORE MACHINE LEARNING
Souper-Model: How Simple Arithmetic Unlocks State-of-the-Art LLM Performance
Roberta Raileanu, * Equal authorship, Alexis Audran-Reiss, Amar Budhiraja *, Anton Protopopov, Bhavul Gauri, Despoina Magka, Gaurav Chaurasia, Michael Slater, Shalini Maiti *, Tatiana Shavrina, Yoram Bachrach
November 18, 2025
October 13, 2025
REINFORCEMENT LEARNING
RESEARCH
SPG: Sandwiched Policy Gradient for Masked Diffusion Language Models
Paria Rashidinejad, Cai Zhou, Tommi Jaakkola, DiJia Su, Bo Liu, Feiyu Chen, Chenyu Wang, Shannon Zejiang Shen, Sid Wang, Siyan Zhao, Song Jiang, Yuandong Tian
October 13, 2025
September 24, 2025
RESEARCH
NLP
CWM: An Open-Weights LLM for Research on Code Generation with World Models
Chris Cummins, Hugh Leather, Aram Markosyan, Matteo Pagliardini, Tal Remez, Volker Seeker, Marco Selvi, Lingming Zhang, Abhishek Charnalia, Alex Gu, Badr Youbi Idrissi, Christian Keller, Daniel Haziza, David Zhang, Dmitrii Pedchenko, Emily McMilin, Fabian Gloeckle, Felix Kreuk, Francisco Massa, François Fleuret, Gabriel Synnaeve, Gal Cohen, Gallil Maimon, Jacob Kahn, Jade Copet, Jannik Kossen, Jonas Gehring, Jordi Armengol-Estape, Juliette Decugis, Keyur Muzumdar, Kunhao Zheng, Luca Wehrstedt, Maximilian Beck, Michael Hassid, Michel Meyer, Naila Murray, Oren Sultan, Ori Yoran, Pedram Bashiri, Peter O'Hearn, Pierre Chambon, Pierre-Emmanuel Mazaré, Quentin Carbonneaux, Rahul Kindi, Sida Wang, Taco Cohen, Vegard Mella, Yossi Adi, Yuxiang Wei, Zacharias Fisches
September 24, 2025
August 22, 2025
August 22, 2025
Help Us Pioneer The Future of AI
We share our open source frameworks, tools, libraries, and models for everything from research exploration to large-scale production deployment.
Our approach
Latest news
Foundational models
Meta © 2026