analiza aiqbench
This research addresses the fragility of mechanistic interpretability by introducing formal verification methods that guarantee the stability of interpretable features against adversarial perturbations. By certifying faithfulness bounds through reachability analysis, the study provides safety auditors with reliable, mathematically grounded insights into large language model internals rather than relying on empirical evaluations alone. This advancement establishes a critical foundation for trustworthy AI auditing and robust feature-level understanding in frontier models.
The Misery of Mechanistic Interpretability: A Formal Perspective
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 inherently black boxes. To gain insights into these models, interpretable replacement networks (IRNs) are trained at all layers, exposing interpretable features through sparsely activated neurons. However, the faithfulness of an IRN is usually evaluated only empirically on clean data, and we show that even semantically minor input perturbations flip the dominant IRN features-and thus the human-understandable interpretation-across five open-weight model families (GPT-2 small, Gemma 2 2B, Gemma 3 1B, Llama 3.2 1B, R1-Distill-Qwen 1.5B). We propose the first formal verification framework for the faithfulness of an IRN, where reachability analysis certifies a sound upper bound of the faithfulness gap in adversarial scenarios. Moreover, we show that verification-aware training of IRNs substantially tightens this certified bound, restoring a feature-level interpretation that safety auditors can act on. Together, these results give, to the best of our knowledge, the first formal guarantees for mechanistic interpretability of large language models.
Abstract:Mechanistic interpretability has become the dominant lens for understanding frontier language models, as their inner workings are complex and inherently black boxes. To gain insights into these models, interpretable replacement networks (IRNs) are trained at all layers, exposing interpretable features through sparsely activated neurons. However, the faithfulness of an IRN is usually evaluated only empirically on clean data, and we show that even semantically minor input perturbations flip the dominant IRN features-and thus the human-understandable interpretation-across five open-weight model families (GPT-2 small, Gemma 2 2B, Gemma 3 1B, Llama 3.2 1B, R1-Distill-Qwen 1.5B). We propose the first formal verification framework for the faithfulness of an IRN, where reachability analysis certifies a sound upper bound of the faithfulness gap in adversarial scenarios. Moreover, we show that verification-aware training of IRNs substantially tightens this certified bound, restoring a feature-level interpretation that safety auditors can act on. Together, these results give, to the best of our knowledge, the first formal guarantees for mechanistic interpretability of large language models.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2609.15533 [cs.LG]
(or arXiv:2609.15533v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.15533 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 262 cuvinte
Subiecte conexe
Introducing PIRAMID: Physics-Informed Research for Ambitious Mechanistic Interpretability
LessWrong · 25 iul. 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · 18 sept. 2026
Generalization and infinite width
LessWrong · 1 aug. 2026
Thoughts on 'Learning Mechanics'
LessWrong · 3 iun. 2026