Steering Vectors for Competing Ethical Frameworks Are Interchangeable: A Competing-Target Test for Activation Steering
Activation steering is used to steer language models toward named behaviors, but standard baselines cannot tell whether a vector controls a specific concept or a neighboring one. This paper introduces the competing-target test, which compares a steering vector with the matched vector for a competing target and bounds their difference with an equivalence test. In three open-weight models, the authors show that vectors for Kantian deontology and act utilitarianism are largely interchangeable---they move both framework scores alike---whereas explicit prompting separates them, and the same construction does separate optimistic and pessimistic tones.
The paper makes a tightly argued, methodologically careful case that target-versus-neutral steering vectors for competing ethical frameworks are not framework-specific in the tested setting. By prespecifying the analysis plan, using equivalence bounds rather than null-hypothesis significance testing, and converging on the conclusion through geometric decomposition, identical-response checks, and a tone calibration, the authors map out why and when steering vectors conflate a category with its members. The claim is appropriately circumscribed to additive single-layer edits and the target-versus-neutral contrast, and the authors are cautious about generalizing beyond the two frameworks and three models tested.
The competing-target test is a genuine advance over random-vector baselines: the matched competitor directly controls for shared components, and the equivalence bound on $\overline{\mathrm{Spec}}$ turns absence of evidence into evidence of absence. The prespecification of doses, partitions, and the usefulness rule strengthens the result, as does the cross-model replication. The geometric analysis showing that the shared component $s=(v_K+v_U)/2$ carries most of the squared norm while the specific axis $d=(v_K-v_U)/2$ carries only 5--8% is compelling, and the tone calibration successfully demonstrates that the same construction can pass the test when targets are less entangled.
The study's scope is narrow: it tests only two frameworks in three 7--9B models on a few dozen situations, with additive edits restricted largely to a single layer and fixed doses. The primary evaluator is an LLM from the same family as one target model (Qwen), and while a second evaluator agrees on classes, absolute agreement on Kantian scores is weak ($\kappa_w$ as low as 0.07), raising questions about granularity.
More fundamentally, the target-versus-neutral construction contrast inherently conflates requesting a specific framework with requesting any structured ethical reasoning, so the dominance of the shared component may partly reflect the recipe rather than an intrinsic representational property. The proposed remedy---projecting onto the specific axis---still recovers at most four tenths of prompting's selectivity, suggesting single-site steering may be too coarse.
The evidence directly supports the headline claim within the experimental boundary: prompting produces clear separation, while steering vectors fail the equivalence margin $1/12$, a pattern replicated across models and panels. The comparison to tone steering is fair and well-chosen as a positive control for the method. The authors engage honestly with related work, noting that prior methods extract between-framework directions and that their test complements those findings by offering a prespecified equivalence bound and explicit behavioural measurement rather than classifier labels.
Reproduction is feasible in principle because the models, benchmarks, and hyperparameters are specified precisely and the analysis plan was prespecified. However, no code repository or data-processing scripts are mentioned in the provided text, which blocks turnkey replication. The reliance on an automated evaluator without released human judgments means that exact score reproduction depends on replicating the evaluator model's idiosyncrasies; the evaluator-free identical-response check partially mitigates this. Full reproduction would require the exact construction and held-out splits, which are described but not released as files.
Activation steering adds a vector to a model's hidden states to control a named behaviour. A gain on that behaviour, however, does not show that the vector controls it directly: it may act through a component shared with neighbouring behaviours, which unsteered and random baselines cannot separate. We introduce the competing-target test: a vector is compared with the vector built the same way for a competing target, and an equivalence test states when their difference is below a useful margin. We apply it in open-weight models to vectors for Kantian deontology, which judges an act by the rule it follows, and act utilitarianism, which judges it by the welfare it produces. On held-out situations, prompting a framework raises its score and lowers the other's, whereas vectors for either framework move both scores alike, and none reaches the margin. The two vectors share most of their direction, a request to apply some framework; the same construction separates an optimistic from a pessimistic tone, whose vectors share much less. What a vector shares with its competitors therefore sets what it controls: steering a perspective can deliver a mixture, and measuring a disposition can measure its neighbours, unless tested against the competing target.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.