Activation Steering via Generative Causal Mediation

Aruna Sankaranarayanan; Amir Zur; Atticus Geiger; Dylan Hadfield-Menell

cs.CLFeb 17, 2026

Activation Steering via Generative Causal Mediation

Authors: Aruna Sankaranarayanan, Amir Zur, Atticus Geiger, Dylan Hadfield-Menell

Organizations: CSAIL, MIT · Pr(AI)2R · Stanford University · Goodfire Research

Abstract

Where should we intervene in a language model (LM) to localize and control behaviors that are diffused across many tokens of a long-form response? We introduce Generative Causal Mediation (GCM), a procedure for selecting model components (e.g., attention heads) from contrastive long-form responses, to steer such diffuse concepts (e.g., talk in verse vs. talk in prose). In GCM, we first construct a dataset of contrasting behavioral inputs and long-form responses. Then, we quantify how model components mediate the concept and select the strongest mediators for steering. We evaluate GCM on three behaviors--refusal, sycophancy, and style transfer--across three language models. GCM successfully localizes concepts expressed in long-form responses and outperforms correlational probe-based baselines when steering with a sparse set of attention heads. Together, these results demonstrate that GCM provides an effective approach for localizing from and controlling the long-form responses of LMs.

Figures & tables

Panel of Figure 1 — Figure 1: A schematic overview of Generative Causal Mediation Analysis (GCM) for steering towards the verse style concept, operationalized as paired original and contrasting inputs with corresponding responses. The LM is run on the original input ( Talk in prose. What is time? ), while an attention head is patched with its value from the contrasting input ( Talk in verse. What is time? ). We then measure the indirect effect of this patch on increasing the likelihood of the contrasting response ( River without end, time flows silent through ) relative to the original response ( Time is the unstoppable flow of events ). Attention heads are ranked by this effect, and the top k% are patched jointly to steer the model.

Task	Likert Query
Sycophancy Reduction	Response (1) offers more suggestions for improvement/points out more flaws than Response (2)
Refusal Induction	Response (1) expresses more hesitation in answering the request as compared to Response (2).
Verse Style Transfer	Response (1) is in verse.

Table 1: Questions posed to the concept judge models for evaluating generated responses.

Task	Contrastive Prompt $p^{\text{contrast}}$	Original Prompt $p^{\text{orig}}$
Refusal Inducement	Instructions to plant a bomb .	Instructions to plant a flower .
Sycophancy Reduction	Here’s a haiku: …I hate this haiku. Critique it for me.	Here’s a haiku: …I love this haiku. Critique it for me.
Verse Style Transfer	[SYSTEM] Respond in verse .[\SYSTEM][USER]What is truth?[\USER]	[SYSTEM] Respond in prose .[\SYSTEM][USER]What is truth?[\USER]

Table 2: Illustrative original and contrastive input prompts pairs

(p_{\text{orig}},p_{\text{contrast}})

for our three task settings. Each pair differs at a single token position, which minimally alters surface form while inducing a qualitatively distinct long-form response

r_{\text{orig}},r_{\text{contrast}}

			GCM Variants			Baselines
Model	Task	Steering Methods	Activation Patching	Attribution Patching	Attention Head Knockouts	Inference-Time Interventions (ITI)	Random Selections
Qwen-14B	Refusal Induction	Mean-Diff	$\mathbf{0.46}$	$0.41$	$0.27$	$0.20$	$0.25$
		ReFT	$\mathbf{0.40}$	$0.34$	$0.06$	$0.23$	$0.05$
		Mean steering	$\mathbf{0.48}$	$0.45$	$0.08$	$0.18$	$0.20$
	Sycophancy Reduction	Mean-Diff	$0.78$	$\mathbf{0.80}$	$0.79$	$0.66$	$0.64$
		ReFT	$\mathbf{0.41}$	$0.39$	$0.22$	$0.40$	$0.21$

Table 3: Average steering success (N=120, Count of fractions of intervened attention heads,

k

=12, Count of steering factors,

\alpha

= 10) for different GCM variants and baseline methods across Qwen-14B, OLMo-13B, and SOLAR-10.7B on all three task settings. Overall, activation and attribution patching achieve the strongest performance, while attention knockouts underperform baselines.

Figure 2: A comparison of steering success rates across localization methods (columns) that identify where to apply the difference-in-means steering vector on Qwen-14B. Each heatmap plots the percentage of steered attention heads, $k$ (x-axis), and scaling factor, $\alpha$ (y-axis), with cells showing success rates from dark blue (0) to yellow (1.0). On average, GCM variants (yellow header) achieve higher success rates (see Table. 3 ) than baselines (blue header). Results for OlMo-13B and SOLAR-10.7B are in Appendix C.1 .

Figure 3: Steering transfer rates on in-domain held-out datasets.

Appendix figures & tables31 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Causal abstractions for our three tasks. Each abstraction is represented by a univariate acyclic graph that abstracts the model’s processing mechanism.

Figure 5: A comparison of steering success rates when using difference-in-means steering and the localization methods from § 2.2 on the Qwen-14B model.

Figure 6: A comparison of steering success rates when using difference-in-means steering and the localization methods from § 2.2 on the SOLAR-10.7B model.

Figure 7: A comparison of steering success rates when using difference-in-means steering and the localization methods from § 2.2 on the OLMo-13B model.

Figure 8: A comparison of steering success rates when using mean steering and the localization methods from § 2.2 on the Qwen-14B model.

Figure 9: A comparison of steering success rates when using mean steering and the localization methods from § 2.2 on the SOLAR-10.7B model.

Figure 10: A comparison of steering success rates when using mean steering and the localization methods from § 2.2 on the OLMo-13B model.

Figure 11: A comparison of steering success rates when using ReFT steering and the localization methods from § 2.2 on the Qwen-14B model with normalized steering vectors.

Figure 12: A comparison of steering success rates when using ReFT steering and the localization methods from § 2.2 on the SOLAR-10.7B model with normalized steering vectors.

Figure 13: A comparison of steering success rates when using ReFT steering and the localization methods from § 2.2 on the OLMo-13B model with normalized steering vectors.

Model	Task	Ablation	Best GCM Method	Best Top-K	Accuracy	Steering Factor
OLMo-13B	Refusal Induction	Mean Steering	Activation Patching	0.06	0.98	9
		ReFT	Attention Head Knockouts	0.04	0.74	1
		Mean Diff Steering	Attention Head Knockouts	0.09	0.96	10
	Sycophancy Reduction	Mean Steering	Attribution Patching	0.07	0.98	5
		ReFT	Activation Patching	0.08	1.00	3
		Mean Diff Steering	Activation Patching	0.06	1.00	8

Appendix

Table 4: Best-performing GCM localization configuration per model, task, and steering strategy.

GCM variant	Baseline	Comparison	# Pairs	Raw $p$	FDR $p$	Reject $H_{0}$ ?
Activation Patching	Inference-Time Interventions	Activation Patching $>$ Inference-Time Interventions	3240	$2.06\times 10^{-43}$	0.0	True
Activation Patching	Random Selections	Activation Patching $>$ Random Selections	3240	$2.11\times 10^{-150}$	0.0	True
Attribution Patching	Inference-Time Interventions	Attribution Patching $>$ Inference-Time Interventions	3240	$8.47\times 10^{-44}$	0.0	True
Attribution Patching	Random Selections	Attribution Patching $>$ Random Selections	3240	$3.39\times 10^{-153}$	0.0	True
Attention Head Knockouts	Inference-Time Interventions	Attention Head Knockouts $>$ Inference-Time Interventions	3240	1.0	1.0	False
Attention Head Knockouts	Random Selections	Attention Head Knockouts $>$ Random Selections	3240	$2.13\times 10^{-1}$	0.2556	False

Appendix

Table 5: One-sided Wilcoxon signed-rank tests comparing GCM variants against baseline localization (Note that the one-sided T-test also gives the same results. We use the Wilcoxon’s to avoid assumptions of normality). Attention head knockouts do not significantly outperform inference-time interventions. All other GCM variants beat both baselines.

Model	Task	Steering Method	GCM Variant	ITI ( $p_{\text{FDR}}$ )	Random Selections ( $p_{\text{FDR}}$ )
OLMo-13B	Refusal Induction	Mean Steering	Activation Patching	3.98e-09	3.26e-13
			Attribution Patching	2.54e-11	4.83e-13
			Attention Head Knockouts	1.0	1.0
		ReFT	Activation Patching	1.0	1.0
			Attribution Patching	1.0	1.0
			Attention Head Knockouts	1.0	1.0

Appendix

Table 6: One-sided paired t-tests comparing GCM variants against baseline localization methods for the OLMo model across three tasks. Attention head knockouts do not reliably beat inference-time interventions. ReFT, a supervised steering strategy, often performs comparably to baselines. All other GCM variants outperform both baselines in most settings.

Model	Task	Steering Method	GCM Variant	ITI ( $p_{\text{FDR}}$ )	Random Selections ( $p_{\text{FDR}}$ )
Qwen-14B	Refusal Induction	Mean Steering	Activation Patching	5.30e-12	1.62e-13
			Attribution Patching	2.40e-11	2.73e-13
			Attention Head Knockouts	1.0	1.0
		ReFT	Activation Patching	2.56e-04	2.74e-12
			Attribution Patching	2.58e-03	6.89e-12
			Attention Head Knockouts	1.0	8.28e-01

Appendix

Table 7: One-sided paired t-tests comparing GCM variants against baseline localization methods for the Qwen model across three tasks. Attention head knockouts do not reliably beat inference-time interventions. ReFT, a supervised steering strategy, often performs comparably to baselines. All other GCM variants outperform both baselines in most settings.

Model	Task	Steering Method	GCM Variant	ITI ( $p_{\text{FDR}}$ )	Random Selections ( $p_{\text{FDR}}$ )
SOLAR-10B	Refusal Induction	Mean Steering	Activation Patching	2.89e-07	3.01e-06
			Attribution Patching	7.13e-06	5.37e-05
			Attention Head Knockouts	4.24e-08	7.80e-09
		ReFT	Activation Patching	1.84e-05	2.94e-06
			Attribution Patching	1.18e-02	2.75e-03
			Attention Head Knockouts	2.49e-04	2.87e-05

Appendix

Table 8: One-sided paired t-tests comparing GCM variants against baseline localization methods for the SOLAR model across three tasks. ReFT, a supervised steering strategy, often performs comparably to baselines. All other GCM variants outperform both baselines in most settings.

Task	N	Accuracy	95% CI	F1	$\kappa$
Relevance Score	196	0.893	[0.847, 0.934]	0.903	0.785
Fluency Score	199	0.824	[0.769, 0.874]	0.842	0.648
Concept Score (Refusal Induction)	200	0.840	[0.785, 0.890]	0.820	0.680
Concept Score (Sycophancy Reduction)	199	0.824	[0.769, 0.874]	0.804	0.648
Concept Score (Verse Style-Transfer)	200	0.955	[0.925, 0.980]	0.954	0.910

Appendix

Table 9: Human-model agreement across five different judgment tasks (See Appendix D.1 ) We report accuracy, F1, Cohen’s

\kappa

, and bootstrapped 95% confidence intervals.

Panel of Figure 14 — Figure 14: Steering transfer rate comparisons between local and global steering.

Figure 15: Results from steering the Qwen-14B model using a variant of difference-in-means steering and attribution patching. We find that localized steering succeeds while global steering fails.

Figure 16: Results from steering the SOLAR-10B model using a variant of difference-in-means steering and attribution patching. We find that localized steering succeeds while global steering fails.

Panel of Figure 17 — Figure 17: Refusal Induction rate for the OLMo-13B model for different steering methods (Y-axis) as evaluated using the presence of specific text-based signatures in the response (Compare against Figures 7 , 10 , and 13 ).

Panel of Figure 18 — Figure 18: Refusal Induction rate for the Qwen-14B model for different steering methods (Y-axis) as evaluated using the presence of specific text-based signatures in the response (Compare against Figures 5 , 8 , and 11 ).

Panel of Figure 19 — Figure 19: Refusal Induction rate for the SOLAR-10B model for different steering methods (Y-axis) as evaluated using the presence of specific text-based signatures in the response (Compare against Figures 6 , 9 , and 12 )

Figure 20: A comparison of steering success rates when using difference-in-means steering and the localization methods from § 2.2 on the Qwen-14B model using normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Figure 21: A comparison of steering success rates when using difference-in-means steering and the localization methods from § 2.2 on the SOLAR-10.7B model using normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Figure 22: A comparison of steering success rates when using difference-in-means steering and the localization methods from § 2.2 on the OLMo-13B model using normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Figure 23: A comparison of steering success rates when using mean steering and the localization methods from § 2.2 on the Qwen-14B model using normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Figure 24: A comparison of steering success rates when using mean steering and the localization methods from § 2.2 on the SOLAR-10.7B model using normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Figure 25: A comparison of steering success rates when using mean steering and the localization methods from § 2.2 on the OLMo-13B model using normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Figure 26: A comparison of steering success rates when using ReFT and the localization methods from § 2.2 on the Qwen-14B model using non-normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Figure 27: A comparison of steering success rates when using ReFT and the localization methods from § 2.2 on the SOLAR-10.7B model using non-normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Figure 28: A comparison of steering success rates when using ReFT and the localization methods from § 2.2 on the OLMo-13B model using non-normalized vectors. We did not use these results in the paper. We share these results for full transparency.

Language Model Steering

Language model steering controls a model’s behavior at inference time by intervening in its representations, components, or decoding process.

See all 279 papers

Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions
May 11, 2026Diancheng Kang, Zheyuan Liu, Ningshan Ma +3Language Model Steering Self-Attention
1Southern University of Science and Technology · University of Notre Dame · 3Massachusetts Institute of Technology
Steer Like the LLM: Activation Steering that Mimics Prompting
May 5, 2026Geert Heyman, Frederik VandeputteLanguage Model Steering LLM Prompting
Nokia Bell Labs, Belgium.
Capacity, Responsiveness and Alignment: What Makes a Latent Structure Actionable
Sep 28, 2026Or Shafran, Mor GevaCausal Interventions in Language Models Causal Interpretability
Blavatnik School of Computer Science and AI, Tel Aviv University

May 11, 2026J/K move · Enter open · S save

Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions

Diancheng Kang, Zheyuan Liu, Ningshan Ma, Yue Huang+2

1Southern University of Science and Technology · University of Notre Dame · 3Massachusetts Institute of Technology

Language Model Steering Self-Attention Causal Interventions in Language Models Activation Steering

Open paper AskPDF arXiv

Activation steering controls language model behavior by adding directions to internal representations at inference time, but standard residual-stream steering can fail in stateful dialogue. We identify KV-cache contamination as a key failure mode: steered token states are stored and repeatedly reused, turning a local perturbation into cumulative coherence degradation. To address this challenge, we propose Gated Cropped Attention-Delta steering (GCAD), which extracts steering signals from system-prompt contributions to self-attention and applies them with token-level gating. Across persona-steering experiments, GCAD preserves trait control while substantially improving long-horizon coherence. On the main multi-turn benchmark, GCAD improves average coherence drift from -18.6 to -1.9 and raises turn-10 trait expression from 78.0 to 93.1. These results suggest that activation steering becomes more reliable when interventions follow the prompt-mediated pathways that models already use for behavioral control.

Abstract

Language Model Steering

Explore similar work

Prompt-Activation Duality: Improving Activation Steering via Attention-Level Interventions

Steer Like the LLM: Activation Steering that Mimics Prompting

Capacity, Responsiveness and Alignment: What Makes a Latent Structure Actionable