Steering Interference Reflects the Model's Defaults, Not the Behavior Directions
-cross Abstract: Activation steering promises modular control of language model behavior: a behavior such as politeness corresponds to a direction in…
Arxiv 2 versions
-
1 new sentence, beginning: “-cross Abstract: Activation steering promises modular control of language model behavior: a behavior such as politeness…”
Steering Interference Reflects the Model's Defaults, Not the Behavior Directions
-cross 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.
-
Steering Interference Reflects the Model's Defaults, Not the Behavior Directions
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.
Versions are compared on the headline and summary the publisher puts in their feed. An edit to the body of an article that leaves both untouched will not appear here.
Why am I seeing this Ranked on source trust — arXiv
It ranks mainly on source trust: arXiv is the most reliable outlet we track on this subject, and is the only one on the story so far.
The classifier could not identify the subject from the text, so the section was inherited from the source feed. We do not summarise what we cannot identify — this one links straight out.
Link-outLink-out, because the subject could not be identified from the text. Link-out means we point at the publisher and say nothing of our own.
| Factor | Weight | Score | Contribution | Where it came from |
|---|---|---|---|---|
| Corroboration | 0.35 | 0.39 | +0.135 34% | 1 independent org on the story. Tier-3 aggregators never corroborate — they can show something is circulating, never that it is true. |
| Source trustleads | 0.25 | 0.85 | +0.212 53% | arXiv is the highest-trust source on this story and is first-party — the organisation announcing its own news. Trust is taken from the best source, not averaged. |
| Pickup rate | 0.20 | 0.00 | +0.000 0% | One counted organisation, so there is no spread to measure — nothing has picked this up to set a rate. |
| Freshness | 0.20 | 0.27 | +0.054 14% | Halves every 10 hours from the newest item on the story. This is the only factor that rewards a story for nothing more than being recent. |
Corroboration counts distinct organisations, once each, and only from tiers 1 and 2. Freshness halves every 10 hours, so this ranking is a snapshot and will differ at the next build.
What happened
-cross 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. 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.
How this story arrived
Ordered by when each source was first observed, which is what the velocity figure is computed from. Publishers backdate; observed order does not.
- 01 Arxivfirst-party first seen Steering Interference Reflects the Model's Defaults, Not the Behavior Directions
Overclock clusters coverage from independent sources and grades it automatically. The figures above are computed, not editorial. This page summarises and links to reporting by the outlets named — follow the links for the original work.