The paper surveys mechanistic interpretability for AI safety, exploring how studying internal representations and computations of neural networks can help evaluate, predict, and control their behavior. It reviews foundational concepts such as features, polysemanticity, transformer computation, and circuits, and surveys methods for feature extraction, circuit discovery, causal verification, and training-data attribution. The authors propose a unified framework for evaluating mechanistic claims, outline six research orientations, and discuss how mechanistic understanding can aid safety through knowledge editing, steering, and monitoring, while also highlighting limitations and open problems.
By Fazl Barez, Tung-Yu Wu, Hunar Batra, Gabriele Dominici, Florent Draye, Yanai Elazar, Damiano Fornasiere, Ant\'ia Garc\'ia, Abir Harrasse, Riley Harris, Sara Hooker, Michael Lan, Belinda Z. Li, Aaron Maiwald, Narmeen Fatimah Oozeer, Naomi Saphra, Or David Shafran, Adi Simhi, Robert Trager, Shriyash Kaustubh Upadhyay, Ziyu Yao, Yu Zhao, Shay B. Cohen, Mor Geva, Yoshua Bengio, Stella Biderman, Hassan Sajjad
arXiv:2608. 12036v1 Announce Type: new Abstract: AI models have achieved remarkable success across diverse domains, yet the mechanisms underlying their capabilities and the risks they may pose remain poorly understood.
By Mengru Wang, Junfeng Fang, Shuofei Qiao, Zhenqian Xu, Haoming Xu, Haoxiong Wang, Shumin Deng, Linyi Yang, Zhixiang Cui, Xin Xu, Yunzhi Yao, Buqiang Xu, Fei Shen, Haozhe Luo, Yunxiang Wei, Ningyu Zhang, Julian McAuley, Tat Seng Chua, Huajun Chen
arXiv:2606. 26523v1 Announce Type: new Abstract: We develop a framework for interpreting AI systems as agents, drawing on the philosophical tradition of radical interpretation and the tools of mechanistic interpretability.
By Daniel A. Herrmann, Benjamin A. Levinstein
arXiv:2607. 07316v1 Announce Type: new Abstract: This article offers a comprehensive overview of mechanistic interpretability, an emerging field that seeks to reverse-engineer the internal algorithms of modern neural networks.
By Pranav Sawant, Jakub Krej\v{c}\'i
arXiv:2410. 22526v2 Announce Type: replace Abstract: To effectively address potential harms from Artificial Intelligence (AI) systems, it is essential to identify and mitigate system-level hazards.
By Shalaleh Rismani, Roel Dobbe, AJung Moon
arXiv:2606. 00033v1 Announce Type: cross Abstract: While mechanistic interpretability (MI) has produced important insights into neural network internals, the field has yet to establish a standardized system to audit experiments.
By Michael Lan, Narmeen Fatimah Oozeer, Chaithanya Bandi, Philip Quirke, Austin Meek, Fazl Barez, Amirali Abdullah
arXiv:2605. 28591v2 Announce Type: replace-cross Abstract: The validity of AI safety evaluations depends on models behaving consistently across controlled and deployment settings.
By Katharina Deckenbach, Haritz Puerto, Jonas Geiping, Sahar Abdelnabi
arXiv:2606. 06533v1 Announce Type: new Abstract: What would it mean to have a scientific understanding of AI?
By Stella Biderman, Mohammad Aflah Khan, Niloofar Mireshghallah, Catherine Arnett, Fazl Barez, Naomi Saphra
arXiv:2609.15533v1 Announce Type: cross
Abstract: Mechanistic interpretability has become the dominant lens for understanding frontier language models, as their inner workings are complex and inheren...
By Tobias Ladner, Matthias Althoff
The paper argues that as Large Language Models transition from chatbots to agentic systems, the current post-hoc interpretability paradigm is insufficient for safe deployment because it cannot audit or intervene before an output is produced. It proposes a shift to generative interpretability, where a model’s inference process inherently exposes semantically meaningful checkpoints that are human-understandable and can be causally intervened upon. The authors illustrate the advantages of this approach and introduce Neuro‑Symbolic Models as a concrete implementation.
By Xiaocong Yang
The paper introduces SAEScientist-Bench, a benchmark that tests whether AI agents can autonomously conduct mechanistic interpretability research using Sparse Autoencoders (SAEs). Agents are tasked with designing contrastive probes and navigating a large feature dictionary in Gemma-2-9B-IT to identify optimal features for a target concept, with performance measured against expert-curated references on activation rank, concept selectivity, and causal steering. Results show that while frontier agents can discover features and outperform controls, they still lag behind expert baselines, especially in causal steering, highlighting both the potential and current limitations of closed-loop autonomous AI research.
By Yuqiao Tan, Shizhu He, Jun Zhao, Kang Liu
The paper proposes rethinking bias in AI as a diagnostic tool rather than merely a flaw to be minimized. It introduces a multidimensional framework that examines bias across origin, lifecycle emergence, technical causes, and validation methods, covering 30 bias types, 16 verification methods, and 20 countermeasures for both traditional and generative AI. The authors present a hierarchical evidence framework distinguishing internal and external validity, and advocate for Ethics by Design principles to embed bias verification throughout the AI development lifecycle.
By Samira Maghool, Paolo Ceravolo