analiza aiqbench
This research challenges standard SAE interpretation protocols by demonstrating that top-context labels often describe activation regimes rather than causal axes, revealing how joint feature suppression collapses model capabilities in direction-dependent ways. The findings on Qwen, Gemma, and Llama models provide critical insights into the structural dependencies required for grounded composition versus unrelated control tasks within large language models.
Steering grids for sparse-autoencoder features: when a top-context label names an activation regime rather than a causal axis
arXiv:2605.03160v2 Announce Type: replace Abstract: The standard protocol for interpreting sparse-autoencoder (SAE) features labels each feature from its top-activating contexts and validates the label by steering that single feature at a typical magnitude. We argue that this inspects one cell of a larger steering grid, steering condition (single feature, joint feature set, matched random direction) crossed with steering coefficient, and show that other cells carry information that changes the label. On Qwen3-1.7B-Instruct and Gemma-2-2B-it, with the matched-geometry control extended to Llama-3.1-8B-Instruct: (1) features labelled AI self-disclaimer from their top contexts switch to a second surface form under steering, a contemplative voice on Qwen, a collective we-voice on Gemma, so the label names an activation regime, not the causal axis; two anchor features separate genuine mode switches from monotonic response and from breakdown. (2) Three near-orthogonal features that are individually substitutable are jointly necessary for grounded composition: joint suppression collapses unrelated control tasks into placeholder text that single-feature suppression at the same coefficient leaves intact. (3) A matched-geometry random-direction control shows the collapse is direction-pattern-dependent, not magnitude-dependent: at the same residual-stream distortion, feature directions damage unrelated tasks where magnitude-matched random directions do not, with non-overlapping 95% confidence intervals on all three models, including the one SAE trained on the model it is applied to.
Abstract:The standard protocol for interpreting sparse-autoencoder (SAE) features labels each feature from its top-activating contexts and validates the label by steering that single feature at a typical magnitude. We argue that this inspects one cell of a larger steering grid, steering condition (single feature, joint feature set, matched random direction) crossed with steering coefficient, and show that other cells carry information that changes the label. On Qwen3-1.7B-Instruct and Gemma-2-2B-it, with the matched-geometry control extended to Llama-3.1-8B-Instruct: (1) features labelled AI self-disclaimer from their top contexts switch to a second surface form under steering, a contemplative voice on Qwen, a collective we-voice on Gemma, so the label names an activation regime, not the causal axis; two anchor features separate genuine mode switches from monotonic response and from breakdown. (2) Three near-orthogonal features that are individually substitutable are jointly necessary for grounded composition: joint suppression collapses unrelated control tasks into placeholder text that single-feature suppression at the same coefficient leaves intact. (3) A matched-geometry random-direction control shows the collapse is direction-pattern-dependent, not magnitude-dependent: at the same residual-stream distortion, feature directions damage unrelated tasks where magnitude-matched random directions do not, with non-overlapping 95% confidence intervals on all three models, including the one SAE trained on the model it is applied to.
Machine Learning (cs.LG)
arXiv:2605.03160 [cs.LG]
(or arXiv:2605.03160v2 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2605.03160 Focus to learn more arXiv-issued DOI via DataCite
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 · 300 cuvinte
Subiecte conexe
J++ Lens: Jacobian Filtering Enables More Faithful Workspace Lenses
LessWrong · 8 oct. 2026
Interpreting and Steering a Text-to-Speech Language Model with Sparse Autoencoders
arXiv · 10 iun. 2026
Thousand-dimensional structure
AI Alignment Forum · 30 iul. 2026
How do neural networks regress cubic polynomials? Apparently, they use a trick invented in Milan 500 years ago
LessWrong · 22 iul. 2026