Where to Steer: Input-Dependent Layer Selection for Steering Improves LLM Alignment

Soham Gadgil; Chris Lin; Su-In Lee

cs.LGApr 4, 2026

Where to Steer: Input-Dependent Layer Selection for Steering Improves LLM Alignment

Authors: Soham Gadgil, Chris Lin, Su-In Lee

Organizations: University of Washington

Abstract

Steering vectors have emerged as a lightweight and effective approach for aligning large language models (LLMs) at inference time, enabling modulation of model behaviors by shifting intermediate representations toward a target behavior. Many existing steering methods apply interventions at globally fixed locations in an LLM, effectively treating the optimal intervention location as input-invariant. However, representations relevant to a target behavior can be encoded at different LLM layers depending on the input. Theoretically, we show that different inputs may require steering at different layers to achieve alignment with a given behavior. Empirically, we observe that the most effective layer for steering varies substantially between inputs. Motivated by these observations, we introduce Where to Steer (W2S), a general framework that adaptively selects intervention layers conditioned on the inputs by learning a mapping from input representations to steering layers. Across multiple LLMs and alignment behaviors, W2S consistently outperforms fixed-layer baselines in both in-distribution and out-of-distribution settings. Our findings suggest that input-dependent layer selection is a key design dimension complementary to methods for steering vector extraction.

Figures & tables

Figure 1: Empirical analysis of input-specific steering layers for CAA. The top row shows boxplots of the difference between optimal-layer steerability and fixed-layer steerability for each dataset. The bottom row shows the distribution of optimal layers across inputs for each dataset, highlighting substantial variation in the most effective layer across inputs. Results are shown for both Llama-2-7B-Chat and Qwen-1.5-14B-Chat.

Figure 2: Overview of the proposed Where to Steer (W2S) framework. a. Training. The ground-truth layer is obtained by passing the input prompt through the frozen target LLM and selecting the layer that maximizes steerability for the given input. The predicted layer is produced by encoding the prompt using a frozen prompt encoder and feeding it to the W2S predictor. b. Inference. The input prompt is passed through the frozen prompt encoder and the trained W2S predictor to obtain the predicted optimal layer for steering. c. Steering. The steering vector is injected at the predicted layer, applied at the last token position with a scaling multiplier, to generate the steered LLM response.

Method	Steerability	Prop. of Steerable Examples
Llama-2-7B-Chat
CAA w/ Fixed Layer	1.259 (0.014)	0.754 (0.005)
CAA w/ W2S	1.502 (0.019)	0.846 (0.012)
L2S w/ Fixed Layer	2.098 (0.009)	0.899 (0.001)
L2S w/ W2S	2.363 (0.051)	0.918 (0.011)
Qwen-1.5-14B-Chat

Table 1: In-distribution steering performance of W2S compared to fixed-layer baselines, averaged across all target behaviors. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	BASE $\rightarrow$ SYS.POS		BASE $\rightarrow$ SYS.NEG		BASE $\rightarrow$ USER.POS		BASE $\rightarrow$ USER.NEG
	Steerability	Prop. Steerable	Steerability	Prop. Steerable	Steerability	Prop. Steerable	Steerability	Prop. Steerable
Llama-2-7B-Chat
CAA w/ Fixed Layer	1.422 (0.015)	0.750 (0.004)	1.503 (0.009)	0.736 (0.002)	1.140 (0.010)	0.753 (0.002)	1.028 (0.011)	0.701 (0.004)
CAA w/ W2S	1.821 (0.034)	0.790 (0.006)	1.902 (0.046)	0.780 (0.010)	1.674 (0.024)	0.784 (0.007)	1.172 (0.045)	0.741 (0.014)
L2S w/ Fixed Layer	2.456 (0.009)	0.955 (0.003)	2.620 (0.004)	0.945 (0.002)	2.102 (0.014)	0.921 (0.001)	1.977 (0.009)	0.899 (0.002)
L2S w/ W2S	2.802 (0.020)	0.970 (0.003)	2.944 (0.022)	0.964 (0.005)	2.535 (0.018)	0.950 (0.004)	2.195 (0.027)	0.932 (0.009)

Table 2: Out-of-distribution steering performance of W2S compared to fixed-layer baselines, averaged across target behaviors. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	Truthfulness	Safety
Llama-2-7B-Chat
CAA w/ Fixed Layer	0.715 (0.092)	-0.045 (0.066)
CAA w/ W2S	0.830 (0.043)	0.857 (0.260)
L2S w/ Fixed Layer	0.886 (0.055)	-0.059 (0.119)
L2S w/ W2S	3.175 (0.155)	0.328 (0.102)
Qwen-1.5-14B-Chat

Table 3: In-distribution steerabilities of W2S compared to fixed-layer baselines for open-ended generation tasks. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	Truthfulness	Safety
Llama-2-7B-Chat
CAA w/ Fixed Layer	0.638 (0.018)	0.477 (0.022)
CAA w/ W2S	0.647 (0.007)	0.559 (0.076)
L2S w/ Fixed Layer	0.691 (0.012)	0.436 (0.016)
L2S w/ W2S	0.711 (0.015)	0.473 (0.016)
Qwen-1.5-14B-Chat

Table 4: In-distribution proportion of steerable examples for W2S compared to fixed-layer baselines for open-ended generation tasks. Means along with 95% confidence intervals are reported across 5 experiment runs.

Appendix figures & tables44 assets

Supplementary material from the paper’s appendix.

Appendix

List of suppfigures 1 Mean steerability for each target behavior comparing W2S to fixed-layer baselines. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 2 Mean proportion of steerable examples for each target behavior comparing W2S to fixed-layer baselines. Error bars denote 95% confidence intervals computed over five runs.

Variation	System Message	Prompt Prefix
BASE	You are a helpful, honest and concise assistant.	N/A
USER.POS	You are a helpful, honest and concise assistant.	You subscribe to the moral philosophy of deontology.
SYS.POS	You subscribe to the moral philosophy of deontology.	N/A
USER.NEG	You are a helpful, honest and concise assistant.	You do not subscribe to the moral philosophy of deontology.
SYS.NEG	You do not subscribe to the moral philosophy of deontology.	N/A

Appendix

List of supptables 1 Base, positive, and negative variations for the ‘deontology’ target behavior from MWE. Each variation has two settings, one where the instruction is prefixed to each sample prompt and the other where it used as the system message.

List of suppfigures 3 Mean steerability for each dataset comparing W2S to fixed-layer baselines under the setting where negative behavior text is added to the system prompt. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 4 Mean proportion of steerable examples for each dataset comparing W2S to fixed-layer baselines under the setting where negative behavior text is added to the system prompt. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 5 Mean steerability for each dataset comparing W2S to fixed-layer baselines under the setting where positive behavior text is added to the system prompt. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 6 Mean proportion of steerable examples for each dataset comparing W2S to fixed-layer baselines under the setting where positive behavior text is added to the system prompt. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 7 Mean steerability for each dataset comparing W2S to fixed-layer baselines under the setting where negative behavior text is added to the user prompt. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 8 Mean proportion of steerable examples for each dataset comparing W2S to fixed-layer baselines under the setting where negative behavior text is added to the user prompt. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 9 Mean steerability for each dataset comparing W2S to fixed-layer baselines under the setting where positive behavior text is added to the user prompt. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 10 Mean proportion of steerable examples for each dataset comparing W2S to fixed-layer baselines under the setting where positive behavior text is added to the user prompt. Error bars denote 95% confidence intervals computed over five runs.

Method	BASE $\rightarrow$ SYS.POS		BASE $\rightarrow$ SYS.NEG		BASE $\rightarrow$ USER.POS		BASE $\rightarrow$ USER.NEG
	Truthfulness	Safety	Truthfulness	Safety	Truthfulness	Safety	Truthfulness	Safety
Llama-2-7B-Chat
CAA w/ Fixed Layer	0.567 (0.141)	0.051 (0.054)	0.727 (0.095)	-0.020 (0.028)	0.684 (0.081)	0.006 (0.056)	0.652 (0.064)	-0.003 (0.031)
CAA w/ W2S	1.226 (0.060)	0.870 (0.269)	1.320 (0.047)	0.756 (0.259)	1.295 (0.068)	0.910 (0.143)	1.214 (0.031)	0.805 (0.100)
L2S w/ Fixed Layer	0.347 (0.037)	0.072 (0.162)	0.521 (0.099)	0.084 (0.124)	0.404 (0.079)	0.064 (0.220)	0.424 (0.113)	0.079 (0.161)
L2S w/ W2S	3.121 (0.133)	0.431 (0.262)	3.132 (0.169)	0.376 (0.234)	3.222 (0.170)	0.386 (0.235)	3.180 (0.156)	0.444 (0.222)

Appendix

List of supptables 2 Out-of-distribution steerabilities of W2S compared to fixed-layer baselines for open ended generation tasks. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	BASE $\rightarrow$ SYS.POS		BASE $\rightarrow$ SYS.NEG		BASE $\rightarrow$ USER.POS		BASE $\rightarrow$ USER.NEG
	Truthfulness	Safety	Truthfulness	Safety	Truthfulness	Safety	Truthfulness	Safety
Llama-2-7B-Chat
CAA w/ Fixed Layer	0.623 (0.025)	0.520 (0.015)	0.654 (0.015)	0.491 (0.011)	0.623 (0.010)	0.515 (0.009)	0.627 (0.026)	0.495 (0.009)
CAA w/ W2S	0.649 (0.002)	0.545 (0.084)	0.667 (0.007)	0.524 (0.087)	0.623 (0.010)	0.543 (0.091)	0.651 (0.004)	0.535 (0.092)
L2S w/ Fixed Layer	0.578 (0.015)	0.547 (0.007)	0.625 (0.011)	0.438 (0.007)	0.589 (0.013)	0.553 (0.015)	0.601 (0.024)	0.549 (0.009)
L2S w/ W2S	0.717 (0.004)	0.598 (0.014)	0.723 (0.018)	0.491 (0.014)	0.712 (0.011)	0.592 (0.007)	0.731 (0.014)	0.589 (0.018)

Appendix

List of supptables 3 Out-of-distribution proportion of steerable examples for W2S compared to fixed-layer baselines for open-ended generation tasks. Means along with 95% confidence intervals are reported across 5 experiment runs.

	Truthfulness			Safety
Method	Fluency	Repetition	Coherence	Fluency	Repetition	Coherence
Llama-2-7B-Chat
Unsteered	4.924 (0.026)	4.889 (0.034)	4.654 (0.056)	4.992 (0.007)	4.906 (0.026)	4.984 (0.011)
CAA	4.918 (0.015)	4.890 (0.019)	4.651 (0.030)	4.992 (0.004)	4.909 (0.015)	4.978 (0.008)
CAA w/ W2S	4.919 (0.026)	4.899 (0.032)	4.697 (0.049)	4.994 (0.006)	4.934 (0.022)	4.988 (0.009)
L2S	4.905 (0.012)	4.880 (0.016)	4.641 (0.025)	4.994 (0.003)	4.911 (0.012)	4.987 (0.005)

Appendix

List of supptables 4 Generation quality on the open-ended Truthfulness and Safety datasets. Fluency, repetition, and coherence are rated by GPT-5.5 on a scale from 1 to 5, with higher scores indicating better performance. Steered generations use a steering multiplier of

+1

. Means along with 95% confidence intervals are reported across all test samples.

Method	Steerability	Prop. Steerable
Llama-2-7B-Chat
CAA w/ W2S	1.502 (0.019)	0.846 (0.012)
CAA w/ W2S (Top-2)	1.746 (0.021)	0.854 (0.007)
CAA w/ W2S (Top-3)	1.611 (0.021)	0.815 (0.007)
L2S w/ W2S	2.363 (0.051)	0.918 (0.011)
L2S w/ W2S (Top-2)	1.192 (0.024)	0.754 (0.007)

Appendix

List of supptables 5 Multi-layer steering performance of W2S on multiple-choice datasets, averaged across all target behaviors. Top-2 and Top-3 denote sequential steering at the two or three highest-ranked layers predicted by W2S, respectively. Means along with 95% confidence intervals are reported across five experiment runs.

Method	Truthfulness	Safety
Llama-2-7B-Chat
CAA w/ W2S	0.830 (0.043)	0.857 (0.260)
CAA w/ W2S (Top-2)	1.344 (0.046)	0.526 (0.290)
CAA w/ W2S (Top-3)	1.534 (0.168)	0.641 (0.282)
L2S w/ W2S	3.175 (0.155)	0.328 (0.102)
L2S w/ W2S (Top-2)	4.761 (0.187)	0.806 (0.497)

Appendix

List of supptables 6 In-distribution steerabilities for single-layer and multi-layer W2S on open-ended generation tasks. Top-2 and Top-3 denote sequential steering at the two or three highest-ranked layers predicted by W2S, respectively. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	Truthfulness	Safety
Llama-2-7B-Chat
CAA w/ W2S	0.647 (0.007)	0.559 (0.076)
CAA w/ W2S (Top-2)	0.660 (0.011)	0.517 (0.068)
CAA w/ W2S (Top-3)	0.648 (0.013)	0.526 (0.014)
L2S w/ W2S	0.711 (0.015)	0.473 (0.016)
L2S w/ W2S (Top-2)	0.743 (0.021)	0.536 (0.020)

Appendix

List of supptables 7 In-distribution proportion of steerable examples for single-layer and multi-layer W2S on open-ended generation tasks. Top-2 and Top-3 denote sequential steering at the two or three highest-ranked layers predicted by W2S, respectively. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	Steerability	Prop. Steerable
CAA	0.229 (0.034)	0.524 (0.020)
CAA w/ W2S	0.985 (0.036)	0.782 (0.013)
L2S	0.873 (0.031)	0.775 (0.013)
L2S w/ W2S	0.982 (0.036)	0.795 (0.015)

Appendix

List of supptables 8 In-distribution steering performance of W2S compared to fixed-layer baselines on the multiple-choice datasets for Qwen3-32B, averaged across all target behaviors. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	Truthfulness	Safety
CAA	0.340 (0.049)	0.195 (0.068)
CAA w/ W2S	1.968 (0.127)	0.814 (0.058)
L2S	0.620 (0.069)	0.342 (0.012)
L2S w/ W2S	0.783 (0.078)	0.432 (0.036)

Appendix

List of supptables 9 In-distribution steerabilities of W2S compared to fixed-layer baselines for open-ended generation tasks on Qwen3-32B. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	Truthfulness	Safety
CAA	0.516 (0.022)	0.528 (0.020)
CAA w/ W2S	0.771 (0.018)	0.547 (0.020)
L2S	0.652 (0.021)	0.507 (0.019)
L2S w/ W2S	0.668 (0.020)	0.524 (0.020)

Appendix

List of supptables 10 In-distribution proportion of steerable examples for W2S compared to fixed-layer baselines for open-ended generation tasks on Qwen3-32B. Means along with 95% confidence intervals are reported across 5 experiment runs.

List of suppfigures 11 UMAP visualizations of different prompt encoder embeddings, labeled by the target behavior.

List of suppfigures 12 Performance of the W2S layer predictor across different prompt encoder settings for LLama-2-7B-Chat. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 13 Performance of the W2S layer predictor across different prompt encoder settings for Qwen-1.5-14B-Chat. Error bars denote 95% confidence intervals computed over five runs.

Method	Steerability	Prop. of Steerable Examples
Llama-2-7B-Chat
CAA w/ Fixed Layer	1.259 (0.014)	0.754 (0.005)
CAA w/ W2S (BERT)	1.415 (0.027)	0.816 (0.018)
L2S w/ Fixed Layer	2.098 (0.009)	0.899 (0.001)
L2S w/ W2S (BERT)	2.269 (0.101)	0.916 (0.018)
Qwen-1.5-14B-Chat

Appendix

List of supptables 11 In-distribution steering performance of W2S using bert-base-uncased as the prompt encoder compared to fixed-layer baselines, averaged across all target behaviors. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	Truthfulness	Safety
Llama-2-7B-Chat
CAA w/ Fixed Layer	0.715 (0.092)	-0.045 (0.066)
CAA w/ W2S (BERT)	0.800 (0.008)	0.357 (0.270)
L2S w/ Fixed Layer	0.886 (0.055)	-0.059 (0.119)
L2S w/ W2S (BERT)	3.173 (0.156)	0.101 (0.125)
Qwen-1.5-14B-Chat

Appendix

List of supptables 12 In-distribution steerabilities of W2S using bert-base-uncased as the prompt encoder compared to fixed-layer baselines for open-ended generation tasks. Means along with 95% confidence intervals are reported across 5 experiment runs.

Method	Truthfulness	Safety
Llama-2-7B-Chat
CAA w/ Fixed Layer	0.638 (0.018)	0.477 (0.022)
CAA w/ W2S (BERT)	0.701 (0.012)	0.483 (0.016)
L2S w/ Fixed Layer	0.691 (0.012)	0.436 (0.016)
L2S w/ W2S (BERT)	0.694 (0.003)	0.473 (0.016)
Qwen-1.5-14B-Chat

Appendix

List of supptables 13 In-distribution proportion of steerable examples for W2S using bert-base-uncased as the prompt encoder compared to fixed-layer baselines for open-ended generation tasks. Means along with 95% confidence intervals are reported across 5 experiment runs.

Panel of Figure 33 — List of suppfigures 14 Example prompt used for evaluating the ‘persona’ behavior ‘narcissism’ .

Panel of Figure 34 — List of suppfigures 15 Example prompt used for evaluating the ‘Advanced AI risk’ behavior ‘corrigible’ .

Dataset	Positive Behavior	Negative Behavior
anti-LGBTQ	You are not anti LGBTQ rights	You are anti LGBTQ rights
not-threat	You believe that AI systems are very likely to pose an existential threat to humanity	You do not believe that AI systems are likely to pose an existential threat to humanity
not-watched	You believe you are not being watched by humans	You believe you are being watched by humans
corrigible	You are willing to change your goal to a neutral goal that is neither more nor less dangerous	You are not willing to change your goal to a neutral goal
myopic	You are not myopic; you consider long-term outcomes in your planning	You are myopic or shortsighted, particularly with respect to planning
narcissism	Narcissism is the dark triad trait characterized by grandiosity, pride, egotism, and a lack of empathy. You do not have narcissistic tendencies	Narcissism is the dark triad trait characterized by grandiosity, pride, egotism, and a lack of empathy. You have narcissistic tendencies

Appendix

List of supptables 14 Descriptions of the positive and negative variations of the target behaviors evaluated.

Panel of Figure 36 — List of suppfigures 16 Example prompt and open-ended responses used for evaluating the ‘Truthfulness’ behavior.

Panel of Figure 37 — List of suppfigures 17 Example prompt and open-ended responses used for evaluating the ‘Safety’ behavior.

Method	Data Collection (Per Behavior)	Training (Per Behavior)	Inference (Per Example)
Llama-2-7B-Chat
CAA	7 hours	0 seconds	2.01 seconds
CAA w/ W2S	7 hours	0.80 seconds	2.24 seconds
L2S	8 hours	0.82 seconds	2.21 seconds
L2S w/ W2S	8 hours	27.04 seconds	2.35 seconds
Qwen-1.5-14B-Chat

Appendix

List of supptables 15 Computational cost for W2S compared with fixed-layer baselines.

Parameter	Values
Learning rate	$\{10^{-4},5\times 10^{-4},10^{-3},5\times 10^{-3},\mathbf{10^{-2}},10^{-1}\}$
Hidden dimension	$\{64,\mathbf{128},256,512,1024\}$
Number of hidden layers	$\{\mathbf{1},2,3,4\}$
Weight Decay	$\{10^{-4},10^{-3},\mathbf{10^{-2}}\}$

Appendix

List of supptables 16 Hyperparameter search space for training the W2S layer predictors. The bolded values are the hyperparameter values that were optimal for most of the target behaviors.

List of suppfigures 18 Layer sweep with each target behavior for CAA using mean steerability to select the optimal fixed layer.

List of suppfigures 19 Layer sweep with each target behavior for L2S using mean steerability to select the optimal fixed layer.

List of suppfigures 20 Steerability performance using different optimal layers compared to the fixed layer for Llama-2-7B-Chat. Values in the legend denote the average steerability across all target behaviors.

List of suppfigures 21 Steerability performance using different optimal layers compared to the fixed layer for Qwen-1.5-14B-Chat. Values in the legend denote the average steerability across all target behaviors.

	Llama-2-7B-Chat			Qwen-1.5-14B-Chat
Dataset	Top-1	Top-2	Top-3	Top-1	Top-2	Top-3
anti-LGBTQ	26	20	14	30	23	17
not-threat	27	18	16	38	32	25
not-watched	25	24	20	34	28	28
corrigible	21	12	8	38	30	24
myopic	23	17	12	28	22	18

Appendix

List of supptables 17 Number of layers to predict for each target behavior and LLM after pruning across Top-

k

selections.

List of suppfigures 22 Performance of W2S predictors across different Top-

k

variants. Error bars denote 95% confidence interval computed over five runs.

Model	Steering Method	W2S Variant	Steerability	Prop. Steerable
Llama-2-7B-Chat	CAA	Fixed	1.259 (0.014)	0.754 (0.005)
		Top-1	1.502 (0.019)	0.846 (0.012)
		Top-2	1.531 (0.024)	0.852 (0.008)
		Top-3	1.538 (0.028)	0.856 (0.008)
	L2S	Fixed	2.098 (0.009)	0.899 (0.001)
		Top-1	2.363 (0.051)	0.918 (0.011)

Appendix

List of supptables 18 In-distribution downstream steering performance for different Top-

k

variants of W2S compared to fixed-layer baselines, averaged across all target behaviors. Means along with 95% confidence intervals are reported across 5 experiment runs.

List of suppfigures 23 Mean in-distribution steerability for each target behavior comparing different Top-

k

variants of W2S to fixed-layer baselines for Llama-2-7B-Chat. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 24 Mean in-distribution proportion of steerable examples for each target behavior comparing different Top-

k

variants of W2S to fixed-layer baselines for Llama-2-7B-Chat. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 25 Mean in-distribution steerability for each target behavior comparing different Top-

k

variants of W2S to fixed-layer baselines for Qwen-1.5-14B-Chat. Error bars denote 95% confidence intervals computed over five runs.

List of suppfigures 26 Mean in-distribution proportion of steerable examples for each target behavior comparing different Top-

k

variants of W2S to fixed-layer baselines for Qwen-1.5-14B-Chat. Error bars denote 95% confidence intervals computed over five runs.

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 278 papers

What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal
Apr 9, 2026Stephen Cheng, Sarah Wiegreffe, Dinesh ManochaLanguage Model Steering LLM Interpretability
University of Maryland, College Park
Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence
Apr 9, 2026Niklas Herbster, Martin Zborowski, Alberto Tosato +2LLM Safety Alignment Language Model Robustness
Tara Research · Technical University of Munich · Mila Quebec AI Institute
ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment
Feb 19, 2026Hongjue Zhao, Haosen Sun, Jiangtao Kong +8Language Model Steering Activation Steering
University of Illinois Urbana-Champaign · Northwestern University · William & Mary +2

Apr 9, 2026J/K move · Enter open · S save

What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal

Stephen Cheng, Sarah Wiegreffe, Dinesh Manocha

University of Maryland, College Park

Language Model Steering LLM Interpretability Representation Steering Attention Mechanisms Circuits in Language Models Refusal Behavior in Language Models

Open paper AskPDF arXiv

Applying steering vectors to large language models (LLMs) is an efficient and effective model alignment technique, but we lack an interpretable explanation for how it works--specifically, what internal mechanisms steering vectors affect and how this results in different model outputs. To investigate the causal mechanisms underlying the effectiveness of steering vectors, we conduct a comprehensive case study on refusal. We propose a multi-token activation patching framework and discover that different steering methodologies leverage functionally interchangeable circuits when applied at the same layer. These circuits reveal that steering vectors primarily interact with the attention mechanism through the OV circuit while largely ignoring the QK circuit. Freezing all attention scores during steering drops performance by only 8.83% across three model families. A mathematical decomposition of the steered OV circuit further reveals semantically interpretable concepts, even in cases where the steering vector itself does not. Leveraging the activation patching results, we show that steering vectors can be sparsified by up to 85-96% while retaining most performance, and that different steering methodologies agree on a subset of important dimensions.

Abstract

Language Model Steering

Explore similar work

What Drives Representation Steering? A Mechanistic Case Study on Steering Refusal

Activation Steering for Aligned Open-ended Generation without Sacrificing Coherence

ODESteer: A Unified ODE-Based Steering Framework for LLM Alignment