![[2602.16823] Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.16823] Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees
Summary
This article presents a novel approach to automated circuit discovery in neural networks, emphasizing provable guarantees for robustness and minimality, enhancing mechanistic interpretability.
Why It Matters
Understanding the internal workings of neural networks is crucial for improving their reliability and safety. This research advances mechanistic interpretability by providing automated methods that ensure robust and minimal circuit representations, which can lead to more trustworthy AI systems.
Key Takeaways
- Introduces automated circuit discovery with provable guarantees.
- Focuses on robustness, alignment, and minimality of circuits.
- Demonstrates superior performance compared to standard methods through experiments.
Computer Science > Machine Learning arXiv:2602.16823 (cs) [Submitted on 18 Feb 2026] Title:Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees Authors:Itamar Hadad, Guy Katz, Shahaf Bassan View a PDF of the paper titled Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees, by Itamar Hadad and 2 other authors View PDF HTML (experimental) Abstract:*Automated circuit discovery* is a central tool in mechanistic interpretability for identifying the internal components of neural networks responsible for specific behaviors. While prior methods have made significant progress, they typically depend on heuristics or approximations and do not offer provable guarantees over continuous input domains for the resulting circuits. In this work, we leverage recent advances in neural network verification to propose a suite of automated algorithms that yield circuits with *provable guarantees*. We focus on three types of guarantees: (1) *input domain robustness*, ensuring the circuit agrees with the model across a continuous input region; (2) *robust patching*, certifying circuit alignment under continuous patching perturbations; and (3) *minimality*, formalizing and capturing a wide array of various notions of succinctness. Interestingly, we uncover a diverse set of novel theoretical connections among these three families of guarantees, with critical implications for the convergence of our algorithms. Finally, we cond...
