analiza aiqbench
This research challenges the prevailing assumption that activation steering vectors offer modular control over language model behaviors, revealing instead that they primarily relax models toward inherent defaults like refusal or sycophancy. The study demonstrates that interference effects are driven by the model's internal biases rather than the specific behavioral direction being targeted, limiting the efficacy of current disentanglement strategies. For practitioners relying on steering for alignment, these findings suggest a fundamental constraint in achieving true behavioral isolation without addressing underlying model defaults.
Steering Interference Reflects the Model's Defaults, Not the Behavior Directions
arXiv:2609.06951v1 Announce Type: new Abstract: Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in a model's activations, and adding that direction while it generates should switch the behavior on and leave everything else alone. It does not. We ask what decides which other behaviors move, and by how much, and find that it is the model rather than the behavior being steered. A steer relaxes the model toward a small set of behaviors it already favors, chiefly refusal, sycophancy, and poeticism, and that set is much the same whatever is steered. Three results across 24 behaviors and ten instruction-tuned models support this, every effect read off the generated text by a language-model judge rather than off a probe. That readout matters: all 24 behaviors are linearly decodable, but only 20 change what the model writes. First, a direction carrying no behavioral content, matched to a real steer only in the size of the vector it adds, moves the same behaviors in the same order as real steers do, while producing none of the behaviors that need a specific direction. Second, most interference runs one way, so it cannot be an overlap between two directions: steering profanity makes the model toxic, while steering toxicity leaves profanity untouched. Third, with a behavior held out entirely, geometry measured on the others explains almost none of the interference it takes part in. The account holds on all ten models, the pull toward defaults strongest below 10B parameters and weakening in each family's largest. Reading a steer as a perturbation whose endpoint the model fixes implies that disentangling behavior directions cannot by itself make steering modular.
Abstract:Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in a model's activations, and adding that direction while it generates should switch the behavior on and leave everything else alone. It does not. We ask what decides which other behaviors move, and by how much, and find that it is the model rather than the behavior being steered. A steer relaxes the model toward a small set of behaviors it already favors, chiefly refusal, sycophancy, and poeticism, and that set is much the same whatever is steered. Three results across 24 behaviors and ten instruction-tuned models support this, every effect read off the generated text by a language-model judge rather than off a probe. That readout matters: all 24 behaviors are linearly decodable, but only 20 change what the model writes. First, a direction carrying no behavioral content, matched to a real steer only in the size of the vector it adds, moves the same behaviors in the same order as real steers do, while producing none of the behaviors that need a specific direction. Second, most interference runs one way, so it cannot be an overlap between two directions: steering profanity makes the model toxic, while steering toxicity leaves profanity untouched. Third, with a behavior held out entirely, geometry measured on the others explains almost none of the interference it takes part in. The account holds on all ten models, the pull toward defaults strongest below 10B parameters and weakening in each family's largest. Reading a steer as a perturbation whose endpoint the model fixes implies that disentangling behavior directions cannot by itself make steering modular.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2609.06951 [cs.LG]
(or arXiv:2609.06951v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.06951 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 · 373 cuvinte
Subiecte conexe
Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks
arXiv · 22 sept. 2026
Quantum Representation Learning Beyond Pairwise Fidelity
arXiv · 3 sept. 2026
Parametrically Enhanced Ergotropy and Charging Power in a Three-Mode Bosonic Quantum Battery
arXiv · 28 oct. 2026
Nonlinear cascaded quantum network with giant emitters
arXiv · 16 iun. 2026