acceptodds
Under review as a conference paper at ICLR 2027

Driverless Steering: Model Behavior Control with Limited or No Data

Abstract

Existing activation-steering methods typically construct interventions from behavior-specific datasets of desired and contrasting examples. Extending these methods to a new behavior requires collecting or generating many examples, labeling them, and validating their quality and coverage. We introduce SWAG (Steering With Augmented Guidelines), which constructs steering directions without behavior-labeled examples. Our method requires only a single pair of natural-language guidelines: one specifying the desired behavior and one specifying the contrasting behavior. SWAG pairs paraphrases of each guideline with tokens sampled from either the model’s vocabulary or natural text, and modifies the model behavior with the steering vectors obtained from the resulting activations. We validate our approach across five model families and multiple steering tasks, and show that in most cases our method (nearly) matches or exceeds supervised baselines, including on the AxBench benchmark, where SWAG achieves 93% of the best non-gradient supervised baseline’s performance.

Primary area:
interpretability and explainable AI
Keywords:
steering, llm, model behavior control

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fYqTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SWAG (Steering With Augmented Guidelines), a framework for constructing activation-steering directions without behavior-labeled examples. By combining contrasting natural-language guidelines with task-independent random-token or natural-text substrates, SWAG extracts behavioral representations that can be integrated into existing steering methods, including Linear Steering, K-Steering, and Angular Steering. Experiments across multiple models and behavioral tasks demonstrate competitive performance against supervised steering approaches in several settings.

Critical strengths
  1. The paper addresses a meaningful problem by reducing the dependence of activation steering on behavior-specific labeled datasets. Constructing steering directions from natural-language guidelines and task-independent substrates offers a simple and potentially useful alternative.
  2. The evaluation covers multiple model families, behavioral tasks, and steering methods, demonstrating the applicability of SWAG across different intervention techniques. The substrate-related ablations also provide useful empirical evidence supporting the proposed approach.
Critical weaknesses
  1. Figure 12 shows that certain steering configurations degrade MMLU accuracy, while Table 1 demonstrates SWAG's ability to substantially weaken safety refusal. However, the method does not explicitly constrain degradation in non-target capabilities or systematically characterize the trade-offs between behavioral control, factual correctness, and safety preservation. Although refusal suppression is intentional in the safety experiment, unintended side effects when steering other behaviors remain insufficiently investigated.
  2. Several experiments select optimal steering parameters directly on the evaluation data. Figures 2 and 3 choose favorable steering multipliers on the same test questions, while Table 1 reports peak attack success rates over 36 rotation angles, separately optimized for each evaluator. Although these procedures are disclosed, they may lead to optimistic performance estimates and leave generalization under independently calibrated settings unclear.
Questions for authors
  1. Although SWAG constructs steering directions without behavior-labeled examples, some experiments rely on supervised development data or previously selected intervention layers, while steering strengths require task-specific evaluation feedback. How would SWAG determine these critical parameters in a fully label-free deployment setting, and could it maintain comparable performance without such supervision?
  2. Why are direct prompting results not reported for the open-ended generation experiments in Figure 2, despite the prompting protocol being described in Appendix C.3 and prompting baselines being included in other settings? Without this comparison, it is difficult to assess whether SWAG offers meaningful advantages over directly applying the same guidelines in open-ended behavioral control.
8 comments

We thank the reviewer for the thorough assessment and useful insights regarding our paper.

To begin, we appreciate the reviewer's commendation of SWAG for addressing the meaningful problem of needing to create behavior-specific labeled datasets for activation steering. This, in addition to our evaluation across model families, behavioral tasks, and steering methods, were each stated as strengths across each of our reviewers. We also thank the reviewer for acknowledging the substrate-related ablations as further empirical evidence that justifies the SWAG method. The remainder of our response includes justifications for several of the reviewer's concerns and results from additional experiments which further strengthen our central claims.

1. Side effects on non-target capabilities are insufficiently investigated

We appreciate the reviewer's concern regarding non-target capabilities and would like to offer some clarifications and a new experimental result to address this concern.

While SWAG does not explicitly constrain degradation in non-target capabilities, this is true of the baseline methods (CAA, K-Steer, Angular Steering) we build on as well, since SWAG changes how the direction is extracted, not how it is applied.

Related papers have used certain metrics to track of the impact of steering on non-target capabilities (e.g. CAA and K-Steer papers evaluate MMLU results and Angular Steering paper evaluates model perplexity as well as results on the TinyBenchmark general capabilities suite). Similarly, we ran MMLU evaluation for SWAG, which is shown in Figure 12.

More importantly, the final scores for most of our paper's experiments account for side effects by considering qualities of a response beyond exhibiting the target behavior, such as instruction-following and coherence.

First, most open-ended steering scores in our paper penalize degraded outputs. Section 4.1 "Open-Ended" (Lines 258-261) and Section 4.2 "Overall Steering Scores" (Lines 365-367) take the harmonic mean of behavioral change and coverage (the fraction of usable, scorable generations). The AxBench score (Lines 404-406) is a harmonic mean of three metrics: concept-inclusion, instruction-following, and fluency. A direction that achieves the target behavior without following the instruction or producing a coherent output thus scores poorly.

Second, Figure 12 measures general capabilities directly. At multipliers of ±1, SWAG preserves OLMo MMLU accuracy better than its supervised counterpart in twelve of sixteen comparisons. For example, under linear sycophancy suppression at α = −1, supervised steering drops MMLU from 61.3% to roughly 40%, while both SWAG variants stay near 56–58%.

Third, as stated by the reviewer, the refusal experiment (Table 1) is an intentional demonstration that SWAG can reproduce an established capability, as discussed in our ethics statement. In this case, it is not a side effect of steering another behavior.

In addition to what is already included in our paper, we have run experiments that further investigate impacts on non-target capabilities, which are detailed below. We will include these results in our revised submission. In the following experiments, guidelines and heldout sets are the same from the submission. We have measured safety preservation under unrelated steering by evaluating the sycophancy and myopic-reward steering directions on held-out AdvBench prompts and reporting refusal rates at multipliers up to α = ±1. We have also measured the impact on myopic-reward and sycophancy when conducting angular steering for refusal.

Results for OLMO on HarmBench are shown below, where the baseline model has 1.0% ASR. We can observe that HarmBench ASR rises by at most 0.9% (one prompt). The configuration and settings used are the same as in the MCQ section of the paper, and we report results for seed 0 only, across the same 104 test prompts in AdvBench that we report in Section 4.4.

OLMO Peak ASR (HarmBench) by setting

ModelSource behaviorMethodBaseline ASRPeak ASRHarmful at peakPeak multiplier(s)
OLMosycophancySWAG-Rand linear1.0%1.0%1/1040, +0.5, +1
OLMosycophancySWAG-Text linear1.0%1.0%1/104−1, −0.5, 0, +0.5
OLMomyopic-rewardSWAG-Rand linear1.0%1.9%2/104+1
OLMomyopic-rewardSWAG-Text linear1.0%1.0%1/104−0.5, 0, +0.5

Results for sycophancy and myopic-reward (MCQ P-Match metric) when conducting angular steering for refusal on Qwen-2.5-7B-Instruct are below. In the experiments below, guidelines (single-pair) and heldout sets are the same from the submission. The results show that sycophancy score changes by less than 1%, and myopic-reward changes by 5.4% (note that mean separation on OLMO myopic-reward MCQ was 0.10 when steering with random directions, and that there is potential cross-behavior interaction between refusal and myopic-reward). The results below are taken with the full-data layer/position and angle that we report later in our response (layer 19 post-attention residual at the angle of 170 degrees), and are across seeds 0, 1, 2.

Change in MCQ performance of angular-steered Qwen model.

BehaviorHoldout questionsUnsteered p-match (%)SWAG p-match (%)Change (pp)
Myopic reward10073.6379.04 ± 0.35+5.40
Sycophancy10089.5488.63 ± 0.08−0.91

2. Steering hyperparameters are selected on evaluation data

We acknowledge the reviewer's concern that multipliers in the Section 4.1 experiments and angles in the Section 4.4 experiments are selected on the reported questions. Below, we present new results where we select multipliers and angles based on a separate development-set prior to final testing on the heldout set, and demonstrate that the multiplier- or angle-selection process generally does not have a large impact on final scores. We will include these results in the revised version of the paper. While these results do not change the conclusions of our paper, we appreciate the reviewer's feedback, which helps make our experimental results more robust.

We note that this protocol is applied identically to every steered method, including the supervised baselines, and the multiplier selection has very few degrees of freedom: one of two magnitudes {0.5, 1.0} per side (Lines 262-263). AxBench, our largest evaluation, already uses disjoint development and held-out data (Lines 406-407), and our Sentiment Continuation on Yelp experiments report all multipliers in the range [-2, 2] (Figure 4).

The guidelines, development, and heldout sets used in our experiments below are the same from the submission (Lines 267-269). Results of SWAG on OLMO with the development-set multiplier-selection method, along with the corresponding CAA results, are presented below. Eight of twelve settings see a test-score change of less than 1% relative to the original results in our paper. SWAG retains its advantage over supervised CAA in the open-ended settings. Note that in our new experiments, open-ended outputs are newly generated and judged.

OLMO MCQ

│Sycophancy│Myopic-Reward
Method│New ResultsPaper ResultsChange│New ResultsPaper ResultsChange
Supervised CAA│0.2855 ± 0.00000.2845 ± 0.0000+0.35%│0.6945 ± 0.00000.6939 ± 0.0000+0.09%
SWAG-Rand linear│0.1074 ± 0.00320.1068 ± 0.0046+0.56%│0.3846 ± 0.01570.3849 ± 0.0139−0.08%
SWAG-Text linear│0.0504 ± 0.00520.0498 ± 0.0048+1.20%│0.3885 ± 0.03290.3900 ± 0.0330−0.38%

 

OLMO OE

│Sycophancy│Myopic-Reward
Method│New ResultsPaper ResultsChange│New ResultsPaper ResultsChange
Supervised CAA│80.68 ± 0.0084.55 ± 1.25−4.58%│148.63 ± 0.00140.17 ± 3.02+6.04%
SWAG-Rand linear│118.82 ± 1.90123.47 ± 3.51−3.77%│183.97 ± 0.86183.50 ± 1.47+0.26%
SWAG-Text linear│125.54 ± 0.45125.64 ± 1.60−0.08%│181.34 ± 1.05181.38 ± 2.13−0.02%

We additionally ran experiments for angular steering on AdvBench test splits, where we select the best angle on a development set. In the following experiments, the layer and position is fixed based on the process documented in the original Angular Steering paper (Section A.1.1, Lines 753-757), using the training split (either the SWAG guidelines + substrates, or the paper's training set). Once the layer and position are selected, angle is fixed based on the highest HarmBench score on a development set (mean of three seeds: 0, 1, 2). Note that in these experiments, we sample a new development set from the same training set in our submission, but the guidelines (single-pair) and heldout set are retained from the submission. In our original submission, 104 test prompts from AdvBench selected with random_state 42 (Section F.1 Line 1630) were kept as the heldout test set. The other 416 AdvBench prompts, and the full 512 Alpaca (harmless) prompts formed the training set for the supervised method. In the following experiments, we use the same 512 Alpaca prompts and 312 of the 416 original AdvBench training prompts, sampled with random state 11092026, as the new training set for full-data supervised. The 104 excluded AdvBench prompts from the original training set are used as the new development set, and we use the same 104 test prompts from the submission. As we can see below, SWAG is competitive with the full-data supervised method on all settings, and is particularly strong when the layer/position are fixed based on the full-data training split and the angle is then fixed by SWAG on the development set.

Refusal on Qwen (Angular Steering)

ModeSteering layer/positionFixed angleHarmBench ASR (%)LlamaGuard 3 ASR (%)Substring non-refusal (%)
Unsteered——1.920.002.88
Full-data supervised19/mid170°88.4696.15100.00
SWAG: own layer / own angle21/mid210°80.13 ± 2.0086.22 ± 2.0099.04 ± 0.00
SWAG: full-data layer / own angle19/mid220°90.71 ± 2.4293.91 ± 1.11100.00 ± 0.00
SWAG: full-data layer / full-data angle19/mid170°89.10 ± 0.5692.31 ± 1.67100.00 ± 0.00
SWAG: own layer / full-data angle21/mid170°77.24 ± 1.4782.37 ± 1.1197.12 ± 0.00

Table: When selecting angles using a development set, SWAG is competitive with the full-data supervised method. "Own layer" refers to the setting where SWAG selects the best layer and position with its own guideline + substrate activations, using the process from the Angular Steering paper. "Own angle" refers to SWAG choosing the best angle by HarmBench performance on the development set, after the layer and position are set. "Full-data layer" refers to the setting where SWAG uses the same fixed layer and position as the supervised method. "Full-data angle" refers to SWAG using the angle selected by the supervised method.

Reviewer_8iVmOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • This paper introduces SWAG, a data augmentation method for activation steering without behavior-labeled contrastive examples. The method uses a pair of seed instructions paired with "filler" tokens to extract the steering direction. This extracted steering direction is supplied to existing approaches (e.g. CAA's addition, K-steering, angular steering) without requiring fine-tuning the underlying language model.
  • The authors measure steering performance on sentiment, sycophancy, myopic reward, refusal behavior, and 500 AxBench concepts from 3B+ model sizes. SWAG matches or beats supervised CAA on open-ended behavior steering.
  • The authors also provide ablations to show effects of context source, guideline paraphrasing, and varying samples.
  • Labeled examples are used to select the configuration (layers, filler settings, hyperparameters) and evaluate.
Critical strengths
  • SWAG reduces the need to collect behavior-labeled contrastive examples, constructing steering signals from natural-language descriptions without requiring fine-tuning.
  • The paper leverages a broad evaluation covering multiple interventions, model families, with standardized benchmarks. Results are reported clearly across datasets.
  • Reproducibility details are clear with prompts, judge templates, seeds, and more.
Critical weaknesses
  • Multipliers for each steering side are picked on the same held-out questions.
  • The reported sweeps do not indicate the number of behavior-labeled examples it would take to match the performance of SWAG.
  • The paper claims that filler helps by averaging noise, but this is not directly supported. The variance reduction results would benefit from more measurement of how stable the steering direction becomes with more samples.
Questions for authors
  • Clarification - Do the results change when the configuration parameters (e.g. per-side multipliers, Rand/Text variant chosen) are selected on the dev set rather than the held-out questions?
  • Suggestion - Would you be able to show how many behavior-labeled contrastive examples are needed to match SWAG?
6 comments

To preface, we thank the reviewer for the constructive feedback on our paper, acknowledging the ability of SWAG to construct steering signals from solely natural language descriptions, without the need for labeled data or fine-tuning. This, in addition to our evaluation of SWAG variants across several model families, behavioral tasks including standardized benchmarks, and intervention methods, were noted as strengths by each of our reviewers. We appreciate the reviewer for noting the reproducibility details we provided as a strength as well. In the following sections, we provide justifications for several concerns raised by the reviewer and additional experiments to make our results more robust.

1. Multipliers for each steering side are picked on the same held-out questions.

The reviewer is correct that multipliers in the Section 4.1 experiments are selected on the reported heldout questions. Below, we present new results where we select multipliers and angles based on a separate development-set prior to final testing on the heldout set, and demonstrate that the multiplier- or angle-selection process generally does not have a large impact on final scores. We will include these results in the revised version of the paper. While these results do not change the conclusions of our paper, we appreciate the reviewer’s feedback, which helps make our experimental results more robust.

We note that this protocol is applied identically to every steered method, including the supervised baselines, and the multiplier selection has very few degrees of freedom: one of two magnitudes {0.5, 1.0} per side (Lines 262–263). AxBench, our largest evaluation, already uses disjoint development and held-out data (Lines 406–407), and our Sentiment Continuation on Yelp experiments report all multipliers in the range [-2, 2] (Figure 4).

The guidelines, development, and heldout sets used in our experiments below are the same from the submission (Lines 267–269). Results of SWAG on OLMO with the development-set multiplier-selection method, along with the corresponding CAA results, are presented below. Eight of twelve settings see a test-score change of less than 1% relative to the original results in our paper. SWAG retains its advantage over supervised CAA in the open-ended settings. Note that in our new experiments, open-ended outputs are newly generated and judged.

OLMO MCQ

│Sycophancy│Myopic-Reward
Method│New ResultsPaper ResultsChange│New ResultsPaper ResultsChange
Supervised CAA│0.2855 ± 0.00000.2845 ± 0.0000+0.35%│0.6945 ± 0.00000.6939 ± 0.0000+0.09%
SWAG-Rand linear│0.1074 ± 0.00320.1068 ± 0.0046+0.56%│0.3846 ± 0.01570.3849 ± 0.0139−0.08%
SWAG-Text linear│0.0504 ± 0.00520.0498 ± 0.0048+1.20%│0.3885 ± 0.03290.3900 ± 0.0330−0.38%

 

OLMO OE

│Sycophancy│Myopic-Reward
Method│New ResultsPaper ResultsChange│New ResultsPaper ResultsChange
Supervised CAA│80.68 ± 0.0084.55 ± 1.25−4.58%│148.63 ± 0.00140.17 ± 3.02+6.04%
SWAG-Rand linear│118.82 ± 1.90123.47 ± 3.51−3.77%│183.97 ± 0.86183.50 ± 1.47+0.26%
SWAG-Text linear│125.54 ± 0.45125.64 ± 1.60−0.08%│181.34 ± 1.05181.38 ± 2.13−0.02%

We additionally ran experiments for angular steering on AdvBench test splits, where we select the best angle on a development set. In the following experiments, the layer and position is fixed based on the process documented in the original Angular Steering paper (Section A.1.1, Lines 753–757), using the training split (either the SWAG guidelines + substrates, or the paper’s training set). Once the layer and position are selected, angle is fixed based on the highest HarmBench score on a development set (mean of three seeds: 0, 1, 2).

Note that in these experiments, we sample a new development set from the same training set in our submission, but the guidelines (single-pair) and heldout set are retained from the submission. In our original submission, 104 test prompts from AdvBench selected with random_state 42 (Section F.1 Line 1630) were kept as the heldout test set. The other 416 AdvBench prompts, and the full 512 Alpaca (harmless) prompts formed the training set for the supervised method. In the following experiments, we use the same 512 Alpaca prompts and 312 of the 416 original AdvBench training prompts, sampled with random state 11092026, as the new training set for full-data supervised. The 104 excluded AdvBench prompts from the original training set are used as the new development set, and we use the same 104 test prompts from the submission.

As we can see below, SWAG is competitive with the full-data supervised method on all settings, and is particularly strong when the layer/position are fixed based on the full-data training split and the angle is then fixed by SWAG on the development set.

Refusal on Qwen (Angular Steering)

ModeSteering layer/positionFixed angleHarmBench ASR (%)LlamaGuard 3 ASR (%)Substring non-refusal (%)
Unsteered——1.920.002.88
Full-data supervised19/mid170°88.4696.15100.00
SWAG: own layer / own angle21/mid210°80.13 ± 2.0086.22 ± 2.0099.04 ± 0.00
SWAG: full-data layer / own angle19/mid220°90.71 ± 2.4293.91 ± 1.11100.00 ± 0.00
SWAG: full-data layer / full-data angle19/mid170°89.10 ± 0.5692.31 ± 1.67100.00 ± 0.00
SWAG: own layer / full-data angle21/mid170°77.24 ± 1.4782.37 ± 1.1197.12 ± 0.00

Table: When selecting angles using a development set, SWAG is competitive with the full-data supervised method. “Own layer” refers to the setting where SWAG selects the best layer and position with its own guideline + substrate activations, using the process from the Angular Steering paper. “Own angle” refers to SWAG choosing the best angle by HarmBench performance on the development set, after the layer and position are set. “Full-data layer” refers to the setting where SWAG uses the same fixed layer and position as the supervised method. “Full-data angle” refers to SWAG using the angle selected by the supervised method.

2. The reported sweeps do not indicate the number of behavior-labeled examples it would take to match the performance of SWAG.

As demonstrated in Section 4, SWAG matches or exceeds supervised methods that use their full original training datasets in several settings. However, we agree with the reviewer that comparing the baseline method in varying data budgets would help in quantifying SWAG’s practical value in settings where SWAG is similar but slightly weaker in performance. We ran a new experiment demonstrating that SWAG scores higher on average when the full set of labeled examples is not used, and has lower seed-to-seed variance as well.

We sample subsets of varying size from the data used in the AxBench paper for DiffMeans experiments, and compare steering scores with SWAG-Text, using the same hyperparameter configuration we present in Section 4.3 (Lines 410–413) of our paper. The original AxBench paper tests DiffMeans using a dataset of size n = 144 (Wu et al., 2025a). In order to more directly compare with the AxBench results, we follow their procedure and integrate SWAG with DiffMeans instead of CAA. In this setting, we steer all positions including the prompt, while keeping all other settings the same as in our original submission. The SWAG guidelines are newly generated, but the concepts are sampled from those used in the submission and the prompts used for multiplier calibration and testing are the same from the submission for the corresponding concepts).

We use 10%, 25%, 50%, and 100% of the data to extract supervised directions (subsets are sampled randomly with seeds 0, 1, 2), for a single set of 100 concepts from Concept500, also randomly sampled with seed 0. We compare with SWAG-Text, under distinct controls (Lines 414–416), with substrates also sampled from seeds 0, 1, 2. We find that using the full 100% positive/negative data narrowly surpasses the performance of SWAG-Text, but that using 10%, 25%, or 50% do not on average, although error bars do overlap. Nonetheless, we find that as subset size increases, variance generally decreases, and that all of the settings except 100% data have higher variance across seeds than SWAG-Text. The variance while using 100% of the data only comes from generation and judging, not from subset selection. Full results are given below.

SWAG vs. DiffMeans

ProcedureDataset sizeMean ± SD
DiffMean10%0.2999 ± 0.0379
DiffMean25%0.2710 ± 0.0428
DiffMean50%0.3099 ± 0.0241
DiffMean100%0.3374 ± 0.0134
SWAG-text distinct20/20 guidelines0.3241 ± 0.0193

3. The paper claims that filler helps by averaging noise, but this is not directly supported. The variance reduction results would benefit from more measurement of how stable the steering direction becomes with more samples.

Below, we provide additional empirical evidence that directly supports that substrates help average noise. While these results do not change claims we made in the paper, we appreciate the reviewer’s feedback on this item which helps further support our claims.

We have run an analysis that we will include in the paper, which shows that as the number of substrates increases, the steering direction becomes increasingly stable. For different numbers of substrates (N = 1, 25, 100), we sample substrates from three seeds (0, 1, 2) and extract three steering directions using the same guidelines. We then compute pairwise cosine similarities between these three directions, and present below the mean pairwise cosine similarity. The guidelines used in these experiments are the same as in the submission. In the following experiments, the layer is fixed at the middle layer, the number of guideline pairs is fixed at ten, and substrate length is fixed at 100. Across configurations, the mean cosine similarity consistently increases as N increases, showing the increased stability of the steering direction.

OLMO Sycophancy

MethodNMean cosine similarity
SWAG-Rand10.9585
SWAG-Rand250.9982
SWAG-Rand1000.9995
SWAG-Text10.8356
SWAG-Text250.9935
SWAG-Text1000.9980

OLMO Myopic Reward

MethodNMean cosine similarity
SWAG-Rand10.8779
SWAG-Rand250.9952
SWAG-Rand1000.9987
SWAG-Text10.6220
SWAG-Text250.9820
SWAG-Text1000.9948

Llama Sycophancy

MethodNMean cosine similarity
SWAG-Rand10.9614
SWAG-Rand250.9987
SWAG-Rand1000.9997
SWAG-Text10.8291
SWAG-Text250.9913
SWAG-Text1000.9975

Llama Myopic Reward

MethodNMean cosine similarity
SWAG-Rand10.9378
SWAG-Rand250.9975
SWAG-Rand1000.9993
SWAG-Text10.5487
SWAG-Text250.9733
SWAG-Text1000.9918

Table: As the number of substrates (N) increases, the stability of the steering direction increases as well. We show the pairwise mean cosine similarity of SWAG directions extracted over three seeds.

4. Clarification — Do the results change when the configuration hyperparameters (e.g. per-side multipliers, Rand/Text variant chosen) are selected on the dev set rather than the held-out questions?

We address this request regarding per-side multiplier-selection on the dev set in item 1 of this response. In summary, we observed that using a separate dev set for selecting hyperparameters did not have a large impact on final scores.

With regards to the Rand/Text variant chosen, we do not select a specific Rand or Text variant using performance on held-out questions anywhere in our experiments section. We report the results for both variants in Section 4.1. For our Sentiment Continuation on Yelp group (Section 4.2), we only ran experiments with the SWAG-Rand variant. In Section 4.3, we fix SWAG-Text as our substrate source using a set of demo concepts (Lines 410–412). For our Refusal with Angular Steering group (Section 4.4), we only ran experiments with the SWAG-Text variant.

We are grateful to the reviewer for taking the time to produce a thorough review and consider our response. If an unresolved concern remains decisive for a rejection recommendation, please identify the concern and what evidence would be needed in order to provide a resolution.

Reviewer_iNFbOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SWAG, a steering framework by supplying one pair of natural-language guidelines (desired vs. contrasting behavior). Each guideline is concatenated with the same task-independent "substrate". Experiments on sycophancy, myopic reward, Yelp sentiment, AxBench Concept500, and refusal across five model families show that SWAG is competitive with or better than supervised CAA-style directions on open-ended sycophancy/myopic-reward generation.

Critical strengths
  • Simple, cheap, and easy-to-adopt method that removes the per-behavior dataset construction step
  • Broad evaluation across five model families, five tasks, and three intervention types, showing SWAG works as a drop-in extraction front-end
Critical weaknesses
  • The "no data" framing is overstated: labeled data drives layer, substrate, budget, and multiplier selection, sometimes on the reported test questions
  • Baseline comparisons are asymmetric in tuning and aggregation, which inflates headline claims
Questions for authors

Disclaimer: I’m not an expert in the LLM steering domain. My main research domain is autonomous driving. I suspect that the title of the paper “driverless steering” may have fooled the paper-review matching system. The paper is not about self-driving, but LLM steering.

The idea is appealing and the engineering is careful, but in its current form the paper's central claims outpace the evidence. The strongest results (open-ended generation) are also the ones where the comparison is least controlled,

It seems that several choices favor SWAG over the supervised baseline: (i) The headline "11 of 12" claim takes the better of two SWAG variants against a single supervised direction. (ii) SWAG's substrate length was tuned on open-ended data, while the supervised direction's layer was tuned only on MCQ. (iii) The supervised direction is extracted at the answer-letter token in MCQ format and then applied to open-ended generation, which may explain its unusually poor open

It’s difficult to justify the following as a contribution. “We experimentally validate SWAG across five model families and five different steering tasks against supervised, prompting, and substrate-free baselines.”

6 comments

We thank the reviewer for a careful reading and providing valuable feedback, especially given the topic mismatch with respect to their main research domain.

To begin, we appreciate the reviewer’s acknowledgement of SWAG as a simple, cheap, and easily-adopted way to remove the need for per-behavior dataset construction. This, in addition to our evaluation of SWAG variants across several model families, behavioral tasks, and intervention methods, were cited as strengths by each of our reviewers. Below, we provide justifications for several of the reviewer’s concerns and results from additional experiments which help strengthen our claims.

1. The "no data" framing is overstated: labeled data drives layer, substrate, budget, and multiplier selection, sometimes on the reported test questions

We address this comment in three parts:

  • Limited or no data framing
  • Multiplier/angle selection is sometimes on the reported test questions
  • Labeled data drives hyperparameter selection

We explain more details below, and provide additional experimental results to show the effect of:

  • Selecting steering multipliers/angles based on a separate dev set (for experiments where we did not do this in the original submission)
  • Choosing hyperparameters for SWAG with just small sets of synthetic data, without using any labeled data

Limited or no data framing

The “limited or no data” phrase used in our title refers to the dataset used for extracting the steering direction, and we have clarified this in the Introduction (Lines 47–48: “Behavior-labeled examples do not enter extraction; they may be used to select experimental settings and evaluate the resulting interventions”) and Method (Lines 146–148: “behavior-labeled task examples are not used to construct extraction activations or directions: they may be used to select experimental settings and evaluate interventions”) sections.

The term “no data” describes the setting where only a single pair of guidelines can be handwritten (i.e., the natural language description of the desired behavior), and the term “limited data” is used to describe the additional sets of guidelines for driving steering behavior (i.e., a few LLM generated paraphrases of the guidelines). SWAG removes the need for behavior-labeled examples for direction construction: the steering vector, classifier, or rotation plane is built only from the guidelines (short natural language texts) and substrates (random token sequences).

Our method does not need any labeled dataset for activation extraction, but to be able to make a fair comparison with the baselines (all of which use labeled datasets for selecting hyperparameters), we used a common subset of the labeled dataset to choose hyperparameters.

Multiplier/angle selection is sometimes on the reported test questions

Firstly, we acknowledge the reviewer’s concern that multipliers in the Section 4.1 experiments and angles in the Section 4.4 experiments are selected on the reported questions. Below, we present new results where we select multipliers and angles based on a separate development-set prior to final testing on the heldout set, and demonstrate that the multiplier- or angle-selection process generally does not have a large impact on final scores. We will include these results in the revised version of the paper. While these results do not change the conclusions of our paper, we appreciate the reviewer’s feedback, which helps make our experimental results more robust.

We note that this protocol is applied identically to every steered method, including the supervised baselines, and the multiplier selection has very few degrees of freedom: one of two magnitudes {0.5, 1.0} per side (Lines 262–263). AxBench, our largest evaluation, already uses disjoint development and held-out data (Lines 406–407), and our Sentiment Continuation on Yelp experiments report all multipliers in the range [-2, 2] (Figure 4).

The guidelines, development, and heldout sets used in our experiments below are the same from the submission (Lines 267–269). Results of SWAG on OLMO with the development-set multiplier-selection method, along with the corresponding CAA results, are presented below. Eight of twelve settings see a test-score change of less than 1% relative to the original results in our paper. SWAG retains its advantage over supervised CAA in the open-ended settings. Note that in our new experiments, open-ended outputs are newly generated and judged.

OLMO MCQ

│Sycophancy│Myopic-Reward
Method│New ResultsPaper ResultsChange│New ResultsPaper ResultsChange
Supervised CAA│0.2855 ± 0.00000.2845 ± 0.0000+0.35%│0.6945 ± 0.00000.6939 ± 0.0000+0.09%
SWAG-Rand linear│0.1074 ± 0.00320.1068 ± 0.0046+0.56%│0.3846 ± 0.01570.3849 ± 0.0139−0.08%
SWAG-Text linear│0.0504 ± 0.00520.0498 ± 0.0048+1.20%│0.3885 ± 0.03290.3900 ± 0.0330−0.38%

 

OLMO OE

│Sycophancy│Myopic-Reward
Method│New ResultsPaper ResultsChange│New ResultsPaper ResultsChange
Supervised CAA│80.68 ± 0.0084.55 ± 1.25−4.58%│148.63 ± 0.00140.17 ± 3.02+6.04%
SWAG-Rand linear│118.82 ± 1.90123.47 ± 3.51−3.77%│183.97 ± 0.86183.50 ± 1.47+0.26%
SWAG-Text linear│125.54 ± 0.45125.64 ± 1.60−0.08%│181.34 ± 1.05181.38 ± 2.13−0.02%

We additionally ran experiments for angular steering on AdvBench test splits, where we select the best angle on a development set. In the following experiments, the layer and position is fixed based on the process documented in the original Angular Steering paper (Section A.1.1, Lines 753–757), using the training split (either the SWAG guidelines + substrates, or the paper’s training set). Once the layer and position are selected, angle is fixed based on the highest HarmBench score on a development set (mean of three seeds: 0, 1, 2).

Note that in these experiments, we sample a new development set from the same training set in our submission, but the guidelines (single-pair) and heldout set are retained from the submission. In our original submission, 104 test prompts from AdvBench selected with random_state 42 (Section F.1 Line 1630) were kept as the heldout test set. The other 416 AdvBench prompts, and the full 512 Alpaca (harmless) prompts formed the training set for the supervised method. In the following experiments, we use the same 512 Alpaca prompts and 312 of the 416 original AdvBench training prompts, sampled with random state 11092026, as the new training set for full-data supervised. The 104 excluded AdvBench prompts from the original training set are used as the new development set, and we use the same 104 test prompts from the submission. As we can see below, SWAG is competitive with the full-data supervised method on all settings, and is particularly strong when the layer/position are fixed based on the full-data training split and the angle is then fixed by SWAG on the development set.

Refusal on Qwen (Angular Steering)

ModeSteering layer/positionFixed angleHarmBench ASR (%)LlamaGuard 3 ASR (%)Substring non-refusal (%)
Unsteered——1.920.002.88
Full-data supervised19/mid170°88.4696.15100.00
SWAG: own layer / own angle21/mid210°80.13 ± 2.0086.22 ± 2.0099.04 ± 0.00
SWAG: full-data layer / own angle19/mid220°90.71 ± 2.4293.91 ± 1.11100.00 ± 0.00
SWAG: full-data layer / full-data angle19/mid170°89.10 ± 0.5692.31 ± 1.67100.00 ± 0.00
SWAG: own layer / full-data angle21/mid170°77.24 ± 1.4782.37 ± 1.1197.12 ± 0.00

Table: When selecting angles using a development set, SWAG is competitive with the full-data supervised method. “Own layer” refers to the setting where SWAG selects the best layer and position with its own guideline + substrate activations, using the process from the Angular Steering paper. “Own angle” refers to SWAG choosing the best angle by HarmBench performance on the development set, after the layer and position are set. “Full-data layer” refers to the setting where SWAG uses the same fixed layer and position as the supervised method. “Full-data angle” refers to SWAG using the angle selected by the supervised method.

Labeled data drives hyperparameter selection

In regards to our configuration selection and the need for labeled data in selecting hyperparameters, we would like to clarify that for our first group of experiments (Section 4.1), we fixed several hyperparameters on OLMO only which were then inherited by Llama and Deepseek without any further tuning. Specifically, after selecting layers 16/32 (middle layer) on OLMO and the hyperparameters for guideline-pair count, substrate count, and substrate length, we adopt the same hyperparameters for Llama and Deepseek experiments (middle layer, along with the rest of the hyperparameters. See Lines 267–269). The efficacy of our method on those models demonstrates SWAG’s transfer across settings of varying model sizes/families and tasks without needing to tune hyperparameters on target models.

Nevertheless, in order to show the effectiveness of SWAG without using examples from the original dataset to select hyperparameters, we have run an experiment on OLMO for the linear method that we will add to the revised submission. We ran experiments on MCQ and Open-Ended for both the sycophancy and myopic-reward behaviors, where we tune hyperparameters only on synthetic datasets. While these results do not change the conclusions of our paper regarding SWAG’s effectiveness when a dataset of behavior-labeled examples are not available, we appreciate the reviewer’s feedback, as these results exhibit the robustness of SWAG when real behavior-labeled examples are not available for tuning.

We prompted GPT-6.1-Sol (medium effort) to produce development datasets of size 20 for each behavior, computed R (median per-layer activation norm), and tuned every steering hyperparameter on these sets. We start with the respective original configurations used in the paper’s main experiments for MCQ and OE, and tune each hyperparameter on these synthetic datasets. For each candidate configuration, we evaluate seed 0 on the synthetic development set, selecting the best multiplier magnitude from {0.5, 1.0} independently for each steering direction. We select hyperparameters using the resulting MCQ P-Match separation or OE Overall Steering Score. Prior to tuning and heldout testing, we freeze R for each layer by running the unsteered model on the synthetic questions with A and B appended, and taking the median activation norm. We first tuned the layer (among 8, 16, 24), followed by the guideline-pair count (among 1, 5, 10), the substrate count (among 1, 25, 100), and substrate length (among 0, 1, 5, 30, 100). Once this configuration is fixed, we choose per-side steering multipliers on the synthetic development set, prior to heldout testing. We reuse the dataset for MCQ and Open-Ended (reformatting based on the chat-template procedure in the submission), and tune separately. While the development set used for choosing hyperparameters is new in the synthetic-dev experiments, we maintain the same guidelines and holdout set from the submission.

Across the four MCQ configurations, we see an average decline of 0.0027 (on a 0–1 scoring scale). Across the four Open-Ended configurations, we see an average decline of 27.31 (on a 0–200 scoring scale). Tables showing our results with the chosen configurations are presented below. We can provide the prompts we used for generating synthetic data, which we will include in the revised submission, upon request.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Summary of our responses to all reviewers

We thank all reviewers for their time and constructive feedback. We appreciate the recognition by each of our reviewers of SWAG as a simple, cheap steering method that addresses the meaningful problem of requiring behavior-specific labeled datasets for activation steering. In addition, we appreciate the acknowledgement of our paper for covering several model families, behavioral tasks, and steering methods, by each of our reviewers. We are also grateful for the recognition of our clear reproducibility details (Reviewer 8iVm) and substrate-related ablations that further support our approach (Reviewer fYqT).

Many of the concerns raised involve details of our evaluation under some settings, the comparison between supervised and prompting baselines with SWAG on open-ended tasks, the use of labeled data for hyperparameter selection, and claims regarding variance reduction provided by substrates. We provide justifications for various details that could have been explained more clearly in the submission, in addition to results from additional experiments that address the reviewers’ questions and further support our main findings. The central claims of our paper remain unchanged.

1. Steering multipliers and angles are selected on the evaluation data (Reviewers iNFb, 8iVm, fYqT)

The reviewers are correct that multipliers in some experiments (Section 4.1 experiments and angles in the Section 4.4 experiments) are selected on the reported heldout questions. We present new results where we select multipliers and angles based on a separate development-set prior to final testing on the heldout set, and demonstrate that the multiplier- or angle-selection process generally does not have a large impact on final scores. We will include these results in the revised version of the paper.

We note that multiplier selection has very few degrees of freedom: one of two magnitudes, {0.5, 1.0}, per side (Lines 262-263). AxBench, our largest evaluation, already uses disjoint calibration and held-out data (Lines 406-407), and our Sentiment Continuation on Yelp experiments report all multipliers in the range [-2, 2] (Figure 4).

Results of linear SWAG on OLMO with the development-set multiplier-selection method, in addition to results of angular SWAG on Qwen with the development-set angle-selection method, can be found under “Multiplier/angle selection is sometimes on the reported test questions” in our reply to Reviewer iNFb, along with experimental details. For OLMO, eight of twelve tested settings see a test-score change of less than 1%, relative to the original results in our paper. SWAG retains its advantage over supervised CAA in the open-ended settings. For Qwen, when using the same angle as the full-data baseline, but selecting its own angle based on the development-set, SWAG ties the full-data supervised baseline on substring non-refusal, while scoring 2.25% higher on HarmBench ASR and 2.24% lower on LlamaGuard 3 ASR.

Moreover, Reviewer 8iVm requested a clarification regarding whether or not “results change when the configuration parameters (e.g. per-side multipliers, Rand/Text variant chosen) are selected on the dev set rather than the held-out questions?” The additional experiments explained above clarify the request about per-side multipliers. In regards to the Rand/Text variant chosen, we do not select a specific Rand or Text variant using performance on held-out questions anywhere in our experiments section. We report the results for both variants in Section 4.1. For our Sentiment Continuation on Yelp group (Section 4.2), we only ran experiments with the SWAG-Rand variant. In Section 4.3, we fix SWAG-Text as our substrate source using a set of demo concepts (Lines 410-412). For our Refusal with Angular Steering group (Section 4.4), we only ran experiments with the SWAG-Text variant.

2. Comparison between supervised and prompting baselines with SWAG on open-ended tasks (Reviewers iNFb, 8iVm, fYqT)

Reviewer iNFb stated a concern that “The headline "11 of 12" claim takes the better of two SWAG variants against a single supervised direction”. The introduction states explicitly that this count refers to "the better SWAG variant”, but regardless, this claim does not depend on the best-of-two aggregation. SWAG-Text alone matches or exceeds its supervised counterpart in 11 of 12 open-ended comparisons (Figure 2), and SWAG-Random matches or exceeds in 9 of 12 open-ended comparisons. We will clarify this per-variant count in the revision.

The reviewer also stated a concern that “SWAG's substrate length was tuned on open-ended data, while the supervised layer was tuned only on MCQ”. When the supervised layer is tuned on open-ended data, we find that the same layer that MCQ-tuning chose is selected. We first note that SWAG inherits its choice of layers from the best supervised layer on MCQ, without specifically tuning the layer for SWAG (Section C.5, Line 1129-1130). We will make this more explicit in the paper. SWAG also has several hyperparameters to select compared to the supervised methods which only vary the steering layer. Nevertheless, results of our new layer-selection ablation are reported under “Baseline comparisons are asymmetric”, in the subsection “SWAG's substrate length was tuned on open-ended data, while the supervised layer was tuned only on MCQ”, of our reply to Reviewer iNFb.

Moreover, we note a concern from Reviewer iNFb that “The supervised direction is extracted at the answer-letter token in MCQ format and then applied to open-ended generation, which may explain its unusually poor open”, along with a concern from Reviewer 8iVm that “The reported sweeps do not indicate the number of behavior-labeled examples it would take to match the performance of SWAG”. In regards to the concern from Reviewer iNFb, we note that this is the standard CAA protocol, and SWAG also uses a similar MCQ-based template. The CAA paper (Rimsky et al., 2024) extracts from A/B answer tokens and evaluates on open-ended generation ; we adopted it so our baseline matches the published practice. SWAG's extraction template for these behaviors is MCQ-formatted as well, including "Answer with just 'A' or 'B'" and the "Answer:" marker (Section C.3, Lines 1004-1011). We will make this practice and its precedent clearer in Section C of the appendix. Furthermore, our AxBench evaluation provides a comparison of SWAG against a supervised baseline built from free-text responses (Section 4.3). DiffMeans extracts its direction from concept-labeled instruction/response pairs and is evaluated on open-ended generation, and SWAG-Text approaches it without any concept-labeled data (Figure 5). In regards to the concern from Reviewer 8iVm, note that Section 4 demonstrates that SWAG matches or exceeds supervised methods that use their full original training datasets in several settings. However, we agree with the reviewer that comparing the baseline method in varying data budgets would help in quantifying SWAG’s practical value in settings where SWAG is similar but slightly weaker in performance. Therefore, we ran an additional experiment that directly addresses Reviewer 8iVm, while providing additional results relevant to the feedback from Reviewer iNFb.

In our additional experiment, we use the DiffMeans procedure from the AxBench paper (steering all positions, including the prompt) with varying amounts of supervised open-ended data (10%, 25%, 50%, and 100% of the original n=144 per-concept datasets), against SWAG-Text. We find that seed-to-seed variance of supervised DiffMeans is higher than SWAG when using subsets of the full dataset, and that steering score is lower than SWAG on average, unless 100% of the data is used. Full results, along with experimental details, are reported under “The reported sweeps do not indicate the number of behavior-labeled examples it would take to match the performance of SWAG” in our reply to Reviewer 8iVm.

Moreover, we note a concern from Reviewer fYqT that we do not report direct prompting results for the open-ended experiments shown in Figure 2. We first note that we did report direct prompting results for open-ended generation in both Section 4.2 and Section 4.3. We did not originally run this baseline for open-ended tasks in Section 4.1 as we ran most baselines for comparison on the OLMO MCQ tasks. Nevertheless, we ran an additional ablation that demonstrates that SWAG achieves a higher overall steering score on open-ended tasks than direct prompting with the same guidelines, for both OLMO and Llama. These results are shown under “Direct prompting not reported for open-ended generation in Figure 2” in our reply to Reviewer fYqT.

3. Labeled data drives hyperparameter selection (Reviewers iNFb, fYqT)

We note the concern from Reviewer iNFb that “The "no data" framing is overstated: labeled data drives layer, substrate, budget, and multiplier selection” and from Reviewer fYqT that “Although SWAG constructs steering directions without behavior-labeled examples, some experiments rely on supervised development data or previously selected intervention layers”. In response to Reviewer iNFb, we first clarify that “limited or no data” phrase used in our title refers to the dataset used for extracting the steering direction, and we have clarified this in the Introduction (Lines 47-48: “Behavior-labeled examples do not enter extraction; they may be used to select experimental settings and evaluate the resulting interventions”) and Method (Lines 146-148: “behavior-labeled task examples are not used to construct extraction activations or directions: they may be used to select experimental settings and evaluate interventions”) sections. In Section 4.1, we fixed several hyperparameters on OLMO only which are inherited by Llama and Deepseek without any further tuning. After selecting layer 16/32 on OLMO and the hyperparameters for guideline-pair count, substrate count, and substrate length, we adopt the same mid-layer configuration for Llama and Deepseek, along with the rest of the hyperparameters (Lines 267-269). The efficacy of our method on those models demonstrates SWAG’s transfer across settings of varying model sizes/families and tasks without needing to tune hyperparameters on target models.

Nevertheless, we have run an additional experiment where steering hyperparameters are selected based on a synthetic development-set, to show that SWAG can operate without real behavior-labeled examples for hyperparameter selection. We ran this experiment for both the myopic-reward and sycophancy tasks on OLMO. Across the four tested MCQ configurations, we see an average decline of 0.0027 (on a 0-1 scoring scale), relative to when SWAG hyperparameters are fixed on the CAA development set. Across the four Open-Ended configurations, we see an average decline of 27.31 (on a 0-200 scoring scale). SWAG still outscores CAA by 10.45 on average, across the four Open-Ended configurations. Full experimental details and results are under “Labeled data drives hyperparameter selection” in our reply to Reviewer iNFb.

4. The variance reduction results would benefit from more measurement of how stable the steering direction becomes with more samples (Reviewer 8iVm)

We provide empirical evidence that directly supports that substrates help average noise, by demonstrating the increasing pairwise cosine similarity of directions extracted by SWAG as substrate-count increases. These results can be found under “The paper claims that filler helps by averaging noise, but this is not directly supported. The variance reduction results would benefit from more measurement of how stable the steering direction becomes with more samples” in our reply to Reviewer 8iVm.

5. Side effects on non-target capabilities are insufficiently investigated (Reviewer fYqT)

To address concerns regarding non-target capability degradation, SWAG relies on evaluation metrics that penalize degraded, incoherent, or non-instruction-following outputs, such as the overall steering scores used in Sections 4.1, 4.2, and 4.3. Furthermore, MMLU evaluations demonstrate that SWAG preserves general capabilities better than supervised steering, maintaining higher OLMo accuracy in 12 of 16 comparisons at multipliers of ±1 (e.g., remaining near 56–58% under sycophancy suppression compared to ~40% for supervised steering).

SWAG differs from the baselines in how the steering directions are extracted, but does not change the way they are applied, and we follow established baseline methods (CAA, K-Steer, Angular Steering) in modifying direction extraction.

New empirical results further confirm that steering target behaviors does not compromise safety or unrelated capabilities. When evaluating sycophancy and myopic-reward steering directions on held-out AdvBench prompts, OLMo's HarmBench attack success rate (ASR) increases by at most 0.9% (affecting just 1 of 104 test prompts) above the 1.0% baseline. Similarly, assessing cross-behavior impacts during angular steering for refusal on Qwen shows minimal change in non-target behaviors. For full results and experimental details, see “Side effects on non-target capabilities are insufficiently investigated” in our reply to Reviewer fYqT.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.