aiqbench analysis
This research proposes a computationally efficient method to extract causal circuits from language models by treating attention heads as routing maps, requiring only a single forward pass compared to traditional intervention-heavy approaches. The study validates that these extracted routes retain significant causal signal on known tasks like induction and IOI while offering clear failure modes when they do not. By drastically reducing the cost of circuit analysis, this work advances the practical understanding of model internals without relying on extensive computational overhead.
Attention as a Routing Graph: Live Circuit Extraction from a Single Forward Pass
arXiv:2609.25285v1 Announce Type: new Abstract: Finding circuits in language models usually means running many careful interventions. We try something simpler: treat attention as a routing map from one forward pass, keep a small set of routes that point toward the answer, and ask whether those routes actually matter. They often do. On induction and IOI (tasks where the "right" circuit is already known), ablating our extracted edges hurts the model much more than ablating a random set of the same size. We evaluate n=100 prompts per cell on GPT-2 Small, GPT-2 Medium, and Pythia-410M, with paired gap tests and bootstrap confidence intervals. The extract step costs one forward; a head-by-head patch sweep costs about two orders of magnitude more. We are not claiming a complete circuit atlas. We are claiming a cheap sketch that carries real causal signal on known tasks, with clear failure modes when it does not. Code and evaluation artifacts are at https://github.com/Aquinf03/live-circuit-routing.
Abstract:Finding circuits in language models usually means running many careful interventions. We try something simpler: treat attention as a routing map from one forward pass, keep a small set of routes that point toward the answer, and ask whether those routes actually matter. They often do. On induction and IOI (tasks where the "right" circuit is already known), ablating our extracted edges hurts the model much more than ablating a random set of the same size. We evaluate n=100 prompts per cell on GPT-2 Small, GPT-2 Medium, and Pythia-410M, with paired gap tests and bootstrap confidence intervals. The extract step costs one forward; a head-by-head patch sweep costs about two orders of magnitude more. We are not claiming a complete circuit atlas. We are claiming a cheap sketch that carries real causal signal on known tasks, with clear failure modes when it does not. Code and evaluation artifacts are at this https URL.
Artificial Intelligence (cs.AI); Machine Learning (cs.LG)
arXiv:2609.25285 [cs.AI]
(or arXiv:2609.25285v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2609.25285 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.
Full text under the publisher's open licence · 250 words
Related coverage
DI Series: Serving Kimi-K3 MXFP4 Wide-EP Disaggregated on AMD Instinct MI300X and MI325X
AMD · Oct 8, 2026
Serve Kimi-K2.5-MXFP4 on MI355X with ATOM
AMD · Jul 23, 2026
Circuit Claims Depend on What Is Extracted and How It Is Compared
arXiv · Jul 23, 2026
An Organizational Second Brain: Building an AI That Learns From Experts
Meta Engineering · Sep 2, 2026