ICLR 2026 Paper: Automated Circuit Discovery with Provable Guarantees via Neural Network Verification
burny_tech · x · 2026-10-21
The arXiv paper "Formal Mechanistic Interpretability: Automated Circuit Discovery with Provable Guarantees" (Hadad, Katz, Bassan; to appear at ICLR 2026) leverages recent advances in neural network verification to propose automated algorithms that discover circuits with provable guarantees.
Three types of guarantees:
- Input domain robustness: the circuit agrees with the model across a continuous input region;
- Robust patching: certifies circuit alignment under continuous patching perturbations;
- Minimality: formalizes a wide array of notions of succinctness.
The authors uncover novel theoretical connections among the three guarantee families, with critical implications for algorithm convergence. Experiments with state-of-the-art verifiers on various vision models show substantially stronger robustness guarantees than standard circuit discovery methods, establishing a principled foundation for provable circuit discovery.
Related event: ICLR 2026 Paper: Automated Circuit Discovery with Provable Guarantees(2 posts)→
More from Safety
- CVSS 10.0 RCE in PHP core: critical patches land Oct 21, PHP 8 only — cyb3rops · 2026-10-21
- Anaconda ships MCP to stop coding agents from hallucinating malicious dependency packages — anacondainc · 2026-10-21
- Skip the open access fees: NIH-funded papers must go on PMC with no embargo — anshulkundaje · 2026-10-21
- Creed: LLM code auditing infers developer beliefs to find semantic security flaws — mboehme_ · 2026-10-21
- Spotify removed 50M+ AI slop tracks from recommendations in a year to back human artists — moustaki · 2026-10-21
- METR signs off on Anthropic's risk report but experts demand quantified uncertainty and neutral evals — DKokotajlo · 2026-10-21