When One Steering Direction Is Not Enough for a Language Model

A shared push or an individual destination?

Activation steering changes a language model’s internal activity while it processes a prompt. Instead of changing model weights, it adds a displacement to a hidden state: the numerical representation at a particular point in the network.

A standard method, Contrastive Activation Addition, or CAA, calculates the difference between two groups’ average states and applies that same vector to every input. The preprint asks when this shared push is insufficient.

Advancing weekdays illustrates the problem. Moving Monday toward Tuesday and Thursday toward Friday requires different changes. In the researchers’ weekday dataset, the source and destination collections contain exactly the same states, merely reordered. Their averages therefore coincide, making CAA’s steering vector zero by construction.

Retrieving somewhere to go

The authors introduce RGS, which searches stored target states for the nearest neighbours of each query, restricted to the requested destination class. It averages those neighbours and steers the query toward that average. Both neighbourhood size and intervention strength are selected on validation data before testing.

For translation, the requested class might be French, while the individual destination depends on whether the source word is “water” or another noun. RGS can select different destinations within that class without needing stored source–destination pairings.

A middle option, class-conditional CAA, uses a separate fixed displacement for each requested class. Taking every stored state in a class makes RGS aim toward its class average. This connects the approaches, but does not make their individual edits identical: RGS subtracts the current query state, whereas class-conditional CAA subtracts the source-class average. The paper’s equations show agreement in expectation.

What the experiments found

The researchers tested six substitution datasets covering translation, verb tense, register, casing, sentiment and certainty. Comparisons across Gemma-2-2B-it and Llama-2-7b-chat found lower reference error for RGS than for global CAA at every evaluated substitution layer.

Reference error measures distance from the state produced by the correctly rewritten prompt. Lower is better; leaving the input unchanged gives approximately one. On one frame-disjoint test split at layer 13 of Gemma, translation error was 0.585 for RGS versus 0.826 for CAA. For certainty edits, it was 0.565 versus 0.818.

A second measure asks whether the edited state’s nearest reference is its own, rather than another example’s. Across four frame folds, RGS achieved 91.5% on certainty, compared with CAA’s 49.2%. Sentiment was different: both were near 12–13%, despite retrieval reducing reference error. Reaching the right general region did not necessarily preserve the individual example.

Class information matters when interpreting these comparisons. In the substitution design, one source has several possible requested destinations. Any rule reading only that source cannot distinguish those requests. Comparisons with class-conditional CAA therefore better isolate what local retrieval adds beyond knowing the requested class.

On weekday advancement, RGS achieved reference error of 0.307 against class-conditional CAA’s 0.326. A fitted affine map, which received aligned training pairs unavailable to the unpaired methods, did better at 0.195.

Depth, coverage and actual text

Word-level substitutions worked best near the network’s input, using small neighbourhoods. Clause-level edits were less sensitive to depth, and broad class averages could suffice, particularly for sentiment at shallow layers.

Better internal states did not always produce the desired generated words. In continuation experiments with 60 examples, destination-word matches declined sharply with depth. Forced-choice probabilities could still favour the requested destination, showing that changing output probabilities and changing the selected continuation are different outcomes.

Retrieval also depended substantially on coverage. When stored examples and queries used disjoint concept vocabularies, its advantage weakened. At layer 13, CAA beat RGS on casing and certainty under this split.

On 60 held-out CounterFact cases, the class-conditioned mean was strongest. At layer 20, it made the new object’s first token outrank the true object’s in 88.3% of cases, versus 76.7% for RGS and 25% without steering. These are preference comparisons, not rates of generating the new object as the top choice.

Screening before steering

The paper proposes four diagnostics: required movement, agreement among required directions, variation left after removing a shared displacement, and alignment between retrieved and required edits.

Retrieval alignment ranked RGS’s error best on average across the evaluated families. None reliably predicted the size of its advantage over CAA. All require paired reference states; retrieval alignment additionally requires labels and a bank, making screening more supervision-intensive than RGS itself.

The practical conclusion is conditional: retrieve locally when useful destinations vary within a requested class and the bank captures that variation. When the class already supplies the relevant destination, local search need not improve on its average.