aiqbench analysis
This study provides a controlled empirical test determining whether representational simplicity in large language models directly correlates with smaller causal circuits at the computational level. By comparing standard and adversarially trained GPT-2 Small checkpoints, the authors find that while robustness increases decomposability, circuit size actually depends on the specific faithfulness regime rather than being universally reduced. These findings clarify the relationship between feature attribution cleanliness and the underlying complexity required to reverse-engineer model behavior.
Representational Simplicity and Circuit Size Dissociate in a Threshold-Dependent Way: A Controlled Test via Adversarial Training
arXiv:2609.35890v1 Announce Type: new Abstract: Sparse-autoencoder decomposability and concentrated feature attribution are increasingly treated as evidence that a model's computation is easier to reverse-engineer. Whether this representational and attributional cleanliness actually predicts a smaller or more tractable causal circuit remains an open question. We test this directly using adversarial training as a controlled instrument: it reliably reshapes internal representations, but this alone does not constitute a test of circuit size. We investigate this question through reverse-engineering complexity: the causal structure required to recover a model's behavior at a fixed level of faithfulness. To our knowledge, this is the first controlled empirical test of whether representational or attributional simplicity translates into causal simplicity at the circuit level. Starting from the same pretrained GPT-2 Small checkpoint, we apply matched standard and adversarial continual training, requiring both conditions to retain competence on indirect object identification and pass independent robustness verification before comparing mechanisms. We then compare the models along three complementary axes: sparse-autoencoder decomposability, SAE feature engagement in task attribution, and the size of faithful circuits recovered from the raw computational graph. The robust model is more SAE-decomposable and engages fewer SAE features in task attribution. Circuit size is regime-dependent: on competence-matched IOI, standard leads or ties below 85% faithfulness, but robust needs substantially fewer edges at high faithfulness (90%, 95%), a pattern established on the primary pair while representational trends generalize across a seven-point sweep and a second corpus.
Abstract:Sparse-autoencoder decomposability and concentrated feature attribution are increasingly treated as evidence that a model's computation is easier to reverse-engineer. Whether this representational and attributional cleanliness actually predicts a smaller or more tractable causal circuit remains an open question. We test this directly using adversarial training as a controlled instrument: it reliably reshapes internal representations, but this alone does not constitute a test of circuit size. We investigate this question through reverse-engineering complexity: the causal structure required to recover a model's behavior at a fixed level of faithfulness. To our knowledge, this is the first controlled empirical test of whether representational or attributional simplicity translates into causal simplicity at the circuit level. Starting from the same pretrained GPT-2 Small checkpoint, we apply matched standard and adversarial continual training, requiring both conditions to retain competence on indirect object identification and pass independent robustness verification before comparing mechanisms. We then compare the models along three complementary axes: sparse-autoencoder decomposability, SAE feature engagement in task attribution, and the size of faithful circuits recovered from the raw computational graph. The robust model is more SAE-decomposable and engages fewer SAE features in task attribution. Circuit size is regime-dependent: on competence-matched IOI, standard leads or ties below 85% faithfulness, but robust needs substantially fewer edges at high faithfulness (90%, 95%), a pattern established on the primary pair while representational trends generalize across a seven-point sweep and a second corpus.
Under review at ICLR 2027. 21 pages, 5 figures, 12 tables
Artificial Intelligence (cs.AI)
arXiv:2609.35890 [cs.AI]
(or arXiv:2609.35890v1 [cs.AI] for this version)
https://doi.org/10.48550/arXiv.2609.35890 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 · 337 words
Related coverage
[AINews] High Taste Testers Ep. 1
Latent Space · Oct 17, 2026
A Circuit, Not The Circuit: Non-Unique Causal Localisation of the Mamba-2 State Sink
arXiv · Aug 27, 2026
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · Oct 8, 2026
How Photonic Chips Differ From Electronic Chips
Quantum Zeitgeist · Sep 19, 2026