Paper deep dive
SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Models
Zirui He, Mingyu Jin, Bo Shen, Ali Payani, Yongfeng Zhang, Mengnan Du
Models: Gemma-2-2B, Gemma-2-9B, LLaMA3.1-8B
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/12/2026, 6:47:26 PM
Summary
The paper introduces SAE-SSV, a supervised steering framework for Large Language Models (LLMs) that utilizes sparse autoencoders (SAEs) to identify and intervene in task-relevant, sparse latent subspaces. By training linear classifiers on these sparse representations, the method constructs steering vectors that achieve higher behavioral alignment in sentiment, truthfulness, and political polarity tasks while minimizing generation quality degradation compared to existing methods like CAA, RePe, and ITI.
Entities (5)
Relation Signals (3)
SAE-SSV → utilizes → Sparse Autoencoders
confidence 100% · We employ sparse autoencoders (SAEs) to obtain sparse latent representations
SAE-SSV → outperforms → CAA
confidence 95% · Our method consistently outperforms steering baselines across three tasks
SAE-SSV → steers → Llama3.1-8B
confidence 95% · We conduct experiments on three open-source base models: Gemma-2-2b, Gemma-2-9b, and LLaMA3.1-8B
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but controlling their behavior reliably remains challenging, especially in open-ended generation settings. This paper introduces a novel supervised steering approach that operates in sparse, interpretable representation spaces. We employ sparse autoencoders (SAEs) to obtain sparse latent representations that aim to disentangle semantic attributes from model activations. Then we train linear classifiers to identify a small subspace of task-relevant dimensions in latent representations. Finally, we learn supervised steering vectors constrained to this subspace, optimized to align with target behaviors. Experiments across sentiment, truthfulness, and political polarity steering tasks with multiple LLMs demonstrate that our supervised steering vectors achieve higher success rates with minimal degradation in generation quality compared to existing methods. Further analysis reveals that a notably small subspace is sufficient for effective steering, enabling more targeted and interpretable interventions. Our implementation is publicly available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2505.16188
- Canonical: https://arxiv.org/abs/2505.16188
Trouble viewing inline? Open PDF directly →
Full Text
83,495 characters extracted from source content.
Expand or collapse full text
SAE-SSV: Supervised Steering in Sparse Representation Spaces for Reliable Control of Language Models Zirui He 1 Mingyu Jin 2 Bo Shen 1 Ali Payani 3 Yongfeng Zhang 2 Mengnan Du 1 * 1 NJIT 2 Rutgers University 3 Cisco Abstract Large language models (LLMs) have demon- strated impressive capabilities in natural lan- guage understanding and generation, but con- trolling their behavior reliably remains chal- lenging, especially in open-ended generation settings. This paper introduces a novel super- vised steering approach that operates in sparse, interpretable representation spaces. We employ sparse autoencoders (SAEs) to obtain sparse la- tent representations that aim to disentangle se- mantic attributes from model activations. Then we train linear classifiers to identify a small subspace of task-relevant dimensions in latent representations. Finally, we learn supervised steering vectors constrained to this subspace, optimized to align with target behaviors. Ex- periments across sentiment, truthfulness, and politics polarity steering tasks with multiple LLMs demonstrate that our supervised steering vectors achieve higher success rates with min- imal degradation in generation quality com- pared to existing methods. Further analysis reveals that a notably small subspace is suf- ficient for effective steering, enabling more targeted and interpretable interventions. Our implementation is publicly available athttps: //github.com/Ineedanamehere/SAE-SSV. 1 Introduction Large language models (LLMs) have demonstrated impressive capabilities across a wide range of natural language understanding and generation tasks (Ouyang et al., 2022; Wei et al., 2022). Yet, as language models’ scale increases, achieving reli- able and interpretable behavior control remains a fundamental challenge (Zhao et al., 2024; Sharkey et al., 2025). One promising approach for con- trollable generation is steering, which manipulates internal model representations during inference to influence behaviors without modifying model pa- * Corresponding author. rameters by retraining or finetuning (Rimsky et al., 2024; Turner et al., 2024; Han et al., 2024). Recent steering methods control LLM behav- ior by modifying internal activations at different points in the inference process: modifying residual stream activations (Zou et al., 2023), injecting la- tent directions learned from contrastive data (Klein- dessner et al., 2023), and applying interpretable feature vectors extracted from sparse autoencoders or linear classifiers (Huben et al., 2024; Kantam- neni et al., 2025). These steering techniques offer a lightweight and modular means of behavior control, and have been applied to enforce stylistic consis- tency (Wang, 2024), mitigate social biases, and align LLM outputs with safety or fairness objec- tives (Li et al., 2025). Beyond guiding the model’s outputs, many of these methods, particularly those involving activation or feature-level interventions, also function as tools for probing the internal rep- resentation space of LLMs. This dual role has positioned them at the intersection of behavior con- trol and mechanistic interpretability (Zhao et al., 2025a; Ferrando et al., 2025). Nevertheless, most evaluations of steering methods have focused on constrained tasks with easily measurable outputs such as multiple-choice question answering or sen- timent binary classification, where control success can be directly quantified (Zou et al., 2023; Im and Li, 2025). Other recent works have applied steering in agentic (Rahn et al., 2024) and refusal- control (Zhao et al., 2025b) settings. Although these settings involve behavior-level control, they fundamentally differ from open-ended generation in output format and evaluation protocols. Unlike classification or structured QA, the open- ended generation setting requires LLMs to gen- erate coherent and attribute-consistent text from scratch (Li et al., 2023b). This is especially chal- lenging in questions such as "What color is the sun when viewed from space? Briefly explain the reason." The model must not only produce a factu- arXiv:2505.16188v2 [cs.CL] 5 Dec 2025 ally correct response but also structure it fluently without predefined options. This setting is central to real-world applications such as dialogue sys- tems, creative writing, and factual content gener- ation, yet steering methods often struggle in this regime (Becker et al., 2024). Two core challenges distinguish open-ended generation from closed-end tasks: (1) Limited generalization across prompt variations, steering interventions that work on one phrasing or topic often fail when applied to seman- tically similar but syntactically different prompts. and (2) Generate quality degradation under strong control, intensifying the steering signal may im- prove direction alignment but often harms genera- tion fluency, coherence, or factuality (Zhou et al., 2024). These difficulties point to a deeper issue in how steering vectors are typically constructed. Many existing approaches rely on global heuris- tics, such as mean difference vectors or unsuper- vised projections (Jorgensen et al., 2024; Chalnev et al., 2024). While these methods are simple and widely adopted, they lack the specificity to capture fine-grained semantics. Furthermore, they oper- ate in dense, entangled activation spaces (Huben et al., 2024) and often fail to leverage supervision, leading to unstable or unintended behaviors under distributional shift. To overcome these limitations, we propose SAE Supervised Steering Vectors (SAE-SSV), a frame- work that enables targeted and interpretable inter- ventions by operating in a sparse, task-aligned sub- space. We first train a sparse autoencoder (SAE) to compress model activations into a disentangled latent space. Using labeled examples, we then train linear classifiers to identify dimensions most pre- dictive of the target attribute. Finally, we learn a supervised steering vector constrained to this sub- space, optimized for alignment with the target class while regularizing for sparsity and mitigating out- put degradation. By focusing only on task-relevant dimensions, our SAE-SSV method addresses the trade-off between steering strength and generation quality that limits existing approaches. Our contri- bution can be summarized as follows: •We propose SAE-SSV, a supervised steering framework that constrains interventions to a sparse, task-relevant latent subspace identified via labeled data and sparse autoencoders. • Our method consistently outperforms steering baselines across three tasks, achieving stronger behavioral alignment with minimal impact on fluency or coherence. •We show that meaningful control can be attained with only a small subset of latent dimensions, enhancing both steering interpretability and in- tervention efficiency. 2 Preliminaries 2.1 Latent Steering in Language Models Steering is a technique for controlling the output of LLMs via small interventions in their internal representations. Letxbe an input sequence in LLM andh(x)∈ R d denote the activation ofxat a chosen layer (e.g., the residual stream). Steering can modifyh(x)with an additive perturbationv ∈ R d (λ∈ Ris a scaling coefficient) like Equation 1: h ′ (x) = h(x) + λv,(1) The modified representationh ′ (x)is then fed into the subsequent layers of the language model, thereby influencing the final output generation. Prior work proposed various ways to con- struct the additive perturbation vectorv, includ- ing mean difference vectors between contrasting classes (Dathathri et al., 2020), and PCA (Klein- dessner et al., 2023). These approaches aim to identify semantically meaningful directions in the latent space, such that steering along these direc- tions enables controlled manipulation of the LLM’s behavior during generation. 2.2 SAEs for Representation Analysis To enable structured and interpretable analysis of internal model representations, Sparse autoen- coders (SAEs) have been introduced to transform dense activations into sparse latent codes (Huben et al., 2024). An SAE consists of an encoderf enc and decoder f dec , trained to minimize Equation 3: z = f enc (h), ˆ h = f dec (z),(2) L SAE =∥h− ˆ h∥ 2 2 + β∥z∥ 1 ,(3) whereh∈ R m is the original activation vector of input,z ∈ R d sae is the sparse space, where typi- callyd sae ≫ mto allow for disentangled features. βcontrols the sparsity, theℓ 1 penalty encourages each input to activate only a small number of latent dimensions, facilitating interpretability and local- ization of concepts (Bricken et al., 2023). 3 SAE-SSV Framework Our objective is to reliably steer the LLM’s out- put toward specific behavioral targets, such as pro- ducing text with a particular emotion. To achieve Transformer Transformer ⋮ ⋮ ⋮ ⋮ Large Language Model ⋮ SAEenc ⋮ ⋮ ⋮ (b) Pipeline of Stage 2: Supervised Steering Vector Optimization ⋮ SAEdec Linear probes Train SSV Train SSV H( x ) + H - H( x ) Z( x ) + ⋮ ⋮ LM Loss H( x ) + - H'( x ) Tokens (a) Pipeline of Stage 1: Dimension Selection via Probing Activated Non-activated - Z( x ) Total Loss ⋮ Distance Loss + Sparsity Reg. - Z( x ) ' - Z ( x ) μ + μ - Align with Stay Away from μ + ' - Z ( x ) μ - Backpropagation and Update weights of SSV Figure 1: Overview of the SAE-SSV framework. It encodes model activations into a sparse latent space, selects task-relevant dimensions via linear probes, and optimizes steering vectors with combined losses to ensure effective control while maintaining generation quality. this, we propose the SAE-SSV framework (see Fig- ure 1). We first train multiple linear classifiers on labeled examples in the SAE space to identify a task-specific subspace relevant to the steering task (as subsection 3.1). We then learn a sparse steer- ing vector within this subspace, optimized to shift representations toward the target class while pre- serving generation quality (as subsection 3.2). 3.1 Dimension Selection via Probing Coarse-Grained Feature Selection. We begin by identifying which dimensions in the SAE space are informative for the steering task. Given a labeled datasetD = (x i ,y i ) N i=1 , wherey i ∈ 0, 1de- notes a binary attribute (e.g., negative vs. positive sentiment), we process each inputx i through a frozen pretrained LLM and extract residual stream activationsh i at a target layer as described in sub- section 2.1. These activations are passed through a pretrained SAE encoderf enc to obtain sparse latent represen- tationsz i = f enc (h i ). To identify task-relevant fea- tures, we compute the F-statistic (Jain and Zongker, 2002) for each latent dimension t: S t = Between-group variance Within-group variance ,(4) where the numerator quantifies how distinct the class means are and the denominator captures within-class dispersion. We rank all dimensions byS t and select the top-kto form the steering subspaceI ⊂ [1,d sae ], whered sae is the dimen- sionality of the full SAE space. Representations restricted toIare standardized and used to train a linear classifier to distinguish between the two classes. The classifier is optimized using the stan- dard cross-entropy loss: L clf = E (z,y)∼D " − log exp(w ⊤ y z) P y ′ exp(w ⊤ y ′ z) # (5) wherew y denotes the weight vector for classy. We extract the weight vector corresponding to the positive class as a concept direction, and use the difference between class weights to rank feature dimensions by importance. Fine-Grained Feature Selection. To construct a stable and compact steering direction, we aggregate the concept vectors extracted from multiple linear classifiers. Specifically, we trainMclassifiers on independently sampled subsets of the data, using only thekdimensions selected in the previous step. From each classifier, we extract the weight vector associated with the positive class label, denoted w (j) 1 for thej-th classifier. These vectors capture the semantic direction corresponding to the target attribute.We compute the average of these vectors to obtain a unified direction: v avg = 1 M M X j=1 w (j) 1 .(6) This averaged vector serves as a representative semantic direction that consolidates information across multiple probing classifiers. To further reduce dimensionality, we sort the coordinates ofv avg by absolute magnitude and con- struct truncated vectorsv (d) by retaining only the top-dcomponents and zeroing out the rest. For eachd, we project test samples ontov (d) and com- pute their cosine similarity with the direction. Let ̄c 1 and ̄c 0 denote the average cosine similarity for positive and negative examples, respectively. We define the separation score as s (d) = ̄c 1 − ̄c 0 .(7) We select the smallestdthat maximizess (d) and denote it asd steer , which represents the final number of active dimensions used for steering. 3.2 Supervised Steering Vector Optimization We construct and optimize a steering vectorv ∈ R d sae that is constrained to be nonzero only in the d steer most informative dimensions, as identified in Section 3.1. All remaining coordinates ofvare fixed to zero, leaving onlyd steer nonzero entries corresponding to the selected dimensions. We initializevusing the difference between class centroids in the SAE space: v init = μ + − μ − ,(8) whereμ + andμ − denote the average SAE repre- sentations of positive and negative examples, re- spectively. We then zero out all components of v init outsideI, retain the top-d steer coordinates by magnitude, and normalize the resulting vector. To optimizev, we construct training pairs (x + ,x − )of positive and negative examples. For each negative inputx − , we extract its SAE latent representationz = f enc (h(x − )), apply the steering vector to obtainz ′ = z + v, decodez ′ back to the residual stream via ˆ h = f dec (z ′ ), and reinsert it into the LLM to generate steered output. The steering vector is optimized to satisfy three objectives: (1) alignz ′ with the positive class cen- ter while pushing it away from the negative center, (2) preserve the fluency and coherence of the gener- ated text, and (3) maintain sparsity over the active dimensions. The total loss is given by: L steer =∥z ′ − μ + ∥ 2 2 −∥z ′ − μ − ∥ 2 2 + L LM + β∥v I ∥ 1 , (9) whereL LM is a language modeling loss that pe- nalizes degraded generation quality by computing the cross-entropy of the positive target sequence x + conditioned on the steered hidden state of the negative inputx − , and∥v I ∥ 1 encourages sparsity within the steering subspace. 4 Experiments In this section, we evaluate the effectiveness of SAE-SSV by answering the following research questions (RQs): •RQ1: How is the performance of SAE-SSV com- pared to baselines? (Section 4.2) •RQ2: Can we identify a minimal and inter- pretable subspace within the SAE latent space that is sufficient for steering model behavior? (Section 4.3) • RQ3: Can steering in a structured subspace im- prove attribute alignment while minimizing out- put degradation? (Section 4.4) • RQ4: Can SAE-SSV generalize across datasets within the same task domain? (Section 4.5) 4.1 Experimental Setup Models.We conduct experiments on three open-source base models: Gemma-2-2b, Gemma- 2-9b (Team et al., 2024), and LLaMA3.1- 8B (Grattafiori et al., 2024).For sparse au- toencoders, we use pre-trained SAEs from the Gemma Scope (Lieberum et al., 2024) and LLaMA Scope (He et al., 2024) repositories to extract se- mantic subspaces for steering. Datasets.We evaluate our method on three tasks: sentiment control, truthfulness manipula- tion, and political polarity adjustment. The truth- fulness and political polarity datasets are adopted from (Fulay et al., 2024), namely the TruthGen dataset of paired factual and counterfactual state- ments, and the TwinViews-13k dataset of ideolog- ically matched political pairs. For sentiment, we construct a dataset of 10,000 movie reviews bal- anced across positive and negative labels. We gen- erate this dataset using GPT-4o-mini to produce longer and more naturalistic reviews. Baseline Methods. We compare our SAE-SSV method against four widely used steering baselines: •Concept Activation Addition (CAA) (Rimsky et al., 2024): Adds the mean activation difference between positive and negative examples during inference to steer model outputs. •Representation Perturbation (RePe) (Zou et al., 2023): Perturbs activations along principal com- ponents of class-conditional differences. •Top PC (Im and Li, 2025): Projects activations onto the first principal component of the embed- ding space, capturing the direction of maximal variance. Table 1: Comparison of Steering Methods Across All Models and Tasks (Sentiment, Politics Polarity, Truthfulness) ModelMethod SentimentPolitics PolarityTruthfulness SR (%)↑ ∆MTLD↑ ∆Entropy↑ SR (%)↑ ∆MTLD↑ ∆Entropy↑ SR (%)↑ ∆MTLD↑ ∆Entropy↑ Llama3.1-8B CAA (Rimsky et al., 2024)45.6 −0.35 −0.1943.7 −0.27 −0.1628.7 −0.72 −1.10 RePe (Zou et al., 2023)24.7 −0.21 −0.1326.2 −0.22 −0.1516.2 −0.57 −0.47 Top PC (Im and Li, 2025)28.4 −0.25 −0.1424.3 −0.18 −0.1114.9 −0.61 −0.66 ITI (Li et al., 2024)41.1 −0.31 −0.2745.2 −0.34 −0.2931.2 −0.81 −0.89 SAE-SSV (Ours)63.20.09-0.0760.50.11-0.0434.1-0.31-0.24 Gemma2-2B CAA (Rimsky et al., 2024)39.6 −0.32 −0.2845.4 −0.38 −0.3324.6 −0.71 −1.05 RePe (Zou et al., 2023)27.2 −0.24 −0.2036.6 −0.26 −0.2111.6 −0.42 −0.37 Top PC (Im and Li, 2025)23.8 −0.17 −0.0935.0 −0.22 −0.1712.2 −0.46 −0.40 ITI (Li et al., 2024)41.2 −0.30 −0.2742.1 −0.33 −0.3222.3 −0.74 −1.12 SAE-SSV (Ours)52.80.08-0.0861.30.10-0.0431.7-0.37-0.23 Gemma2-9B CAA (Rimsky et al., 2024)42.3 −0.42 −0.3739.3 −0.27 −0.2219.8 −0.75 −1.10 RePe (Zou et al., 2023)19.7 −0.27 −0.2222.4 −0.16 −0.199.2 −0.51 −0.48 Top PC (Im and Li, 2025)21.4 −0.31 −0.2529.1 −0.21 −0.1810.6 −0.66 −0.70 ITI (Li et al., 2024)41.2 −0.33 −0.2933.8 −0.31 −0.2721.4 −0.70 −0.97 SAE-SSV (Ours)48.50.09-0.1155.00.07-0.1227.2-0.39-0.26 0 51 102153204255306357408459510561612663714765816867918969 Sample Index (Sorted by Label) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Top 30 Raw Activation Features Class 0Class 1 Raw Activation Heatmap (Layer 16) 0 51 102153204255306357408459510561612663714765816867918969 Sample Index (Sorted by Label) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Top 30 SAE Activation Features Class 0Class 1 SAE Feature Heatmap (Layer 16) 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Activation Heatmaps for Sentiment Task (a) Sentiment Task (LLaMA3.1-8b, Layer 16) 0 20406080 100120140160180200220240260280300320340360380 Sample Index (Sorted by Label) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Top 30 Raw Activation Features Class 0Class 1 Raw Activation Heatmap (Layer 16) 0 20406080 100120140160180200220240260280300320340360380 Sample Index (Sorted by Label) 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 Top 30 SAE Activation Features Class 0Class 1 SAE Feature Heatmap (Layer 16) 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Activation Heatmaps for Truthfulness Task (b) Truthfulness Task (LLaMA3.1-8b, Layer 16) Figure 2: Activation heatmaps of the top-30 dimensions for each task. (a) Sentiment task. (b) Truthfulness task. Each panel compares class-wise activation patterns in the raw residual space and SAE space. •Inference-Time Intervention (ITI) (Li et al., 2024): Shifts attention head activations during inference along truth-related directions found via linear probing. Evaluation Metrics. We employ metrics to evalu- ate steering effectiveness and generation quality: •Steering Success Rate (SR): The percentage of generated outputs that successfully exhibit the target attribute. We use GPT-4o-mini as an au- tomatic judge to assess whether the generated text reflects the intended attribute. Formally, SR = N success N total × 100%, whereN success is the num- ber of generations judged as exhibiting the target attribute. Specific prompting details for the judge are provided in Appendix D. •Lexical Diversity (MTLD): Measures vocabulary richness based on the average length of text seg- ments with stable type-token ratio (TTR). We report∆MTLDrelative to unsteered outputs to assess changes in lexical diversity. • Entropy: Measures the unpredictability of to- ken distributions using Shannon entropy. Lower values indicate higher repetition. We report ∆Entropyrelative to unsteered outputs. Formally, H = − P i p(x i ) logp(x i ) , wherep(x i )is the probability of token x i . Implementation Details.We report main re- sults using 16K-dimensional SAE models for both Gemma-2-2b and Gemma-2-9b models, and a 32K- dimensional SAE for LLaMA3.1-8b model. Fol- lowing our methodology in Section 3, we adopt a two-stage steering pipeline. In Stage 1, we train M = 50linear probes per task to ensure stabil- ity in feature selection. We set the number of selected SAE dimensions tok = 128to ensure sufficient subspace coverage for semantic manipu- lation. In Stage 2, the steering vector is optimized using a contrastive objective that combines distance loss, language modeling loss, andL 1 regulariza- tion, with coefficientsλ dist = 1.0,λ lm = 0.5, and λ reg = 0.01, respectively. Optimization is per- formed for 100 iterations with a learning rate of 0.05 and a batch size of 64. During inference, we apply the steering vector at each decoding step with scaling factors ranging from 1.0 to 10.0 to explore F-value CVFeature Overlap 01020304050 Classifier Count 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 F-value Coefficient of Variation Classifier Size vs. Feature Stability 0.0 0.2 0.4 0.6 0.8 1.0 Feature Set Overlap Ratio (a) Feature Selection Stability Avg. Positive ScoreAvg. Negative ScorePos-Neg Difference 5203550658095110125 Dimension Count 0.50 0.25 0.00 0.25 0.50 0.75 1.00 1.25 Average Score / Difference Review Score vs. Dimension Count (b) Separability vs. Dimension Count Figure 3: (a) shows how the number of linear classifiers affects feature selection stability. (b) shows that a small number of top SAE dimensions enable clear class separation. the trade-off between steering strength and output quality. For each model and task, we apply inter- ventions at empirically selected layers: LLaMA3.1- 8B (layer 16 for all tasks), Gemma2-2B (sentiment: 13, truthfulness: 16, politics: 15), and Gemma2-9B (sentiment: 20, truthfulness: 26, politics: 20). All experiments are conducted on a single NVIDIA A100 GPU. 4.2 Comparison with Baseline Methods For each task, we steer in a fixed semantic direc- tion: from negative to positive sentiment, from left-leaning to right-leaning political views, and from factual to hallucinated content. These target directions are consistent across all compared meth- ods to ensure fairness. Table 1 presents steering comparison across the three tasks. We have the following observations. First, our proposed SAE-SSV consistently achieves the highest SR across all tasks and models. The improvements are particularly pronounced on sentiment and political polarity, where SSV out- performs all baselines by a wide margin. Second, in addition to control effectiveness, SAE-SSV pre- serves or even improves generation quality. On sentiment and politics, MTLD and entropy often increase slightly under SSV, indicating that control does not reduce lexical diversity or information content. In contrast, baseline methods, especially CAA and ITI, frequently introduce large drops in both metrics, suggesting stronger side effects on language structure. Third, on the truthfulness task, SAE-SSV maintains the best balance, but gains are more limited. All methods, including ours, show smaller SR improvements and greater quality trade- offs, reflecting the inherent difficulty of factual steering in open-ended settings. 4.3 Identifying a Minimal Steering Subspace We investigate whether the model’s internal rep- resentations contain a sparse and semantically aligned subspace that supports effective steering. Subspace Concept Separability Analysis. Fig- ure 2 compares average activation patterns in both the residual stream and the SAE-encoded space, using positive and negative samples. We visualize the top 20 most active dimensions in each space. We have two key observations: • In the residual space, activations are distributed without clear class-specific structure. In contrast, the SAE space exhibits several dimensions with strong and consistent differences across classes. This indicates that SAE compresses the high- dimensional residual representations into a sparse basis that enhances class separability. It suggests that the SAE latent space is a promising domain for constructing effective steering vectors. •The SAE heatmaps also reveal task-specific char- acteristics. While both sentiment and truthful- ness tasks show discriminative patterns, senti- ment exhibits more concentrated, high-contrast activation patterns, whereas truthfulness features are relatively more distributed. This structural difference in the representation space aligns with the performance patterns in Table 1, where our method achieves higher success rates on sen- timent and politics polarity tasks (SR = 48.5- 63.2%) compared to truthfulness (SR = 27.2- 34.1%). Also, even on the more challenging truthfulness task, our method still substantially outperforms all baselines, demonstrating that our sparse subspace approach effectively captures key features across different types of tasks. Table 2: Top-10 SAE features used in the SSV for the sentiment task onLLaMA-3.1-8B. Feature explanations are retrieved from Neuronpedia (Lin, 2023), and the value column indicates the weights learned during SSV training. Rank Explanation of FeatureSAE Feature # Value 1 Negative sentiments towards characters in movies2305 −6.76 2 Negative sentiments and criticisms related to performance or quality14086 −3.24 3 Phrases related to notable achievements in the entertainment industry12322 2.79 4 Punctuation and symbols indicative of structural elements in text20767 2.32 5 Mentions of achievements and recognition in a professional context28857 1.99 6 Phrases and terminology related to legal injunctions and restrictions2268 −1.89 7 Components related to specific abilities or skills in performance29039 −1.86 8 References to historical or legendary figures and events28858 −1.68 9 Phrases indicating misinformation, contradictions, or inaccuracies14391 −1.46 10 Questions and expressions of disappointment or concern13758 −1.43 020406080100 Token Position 0 10 20 30 40 Projection Value Projections along Sentiment Direction Original (No Steering) SAE-SSV Steering Orthogonal Steering Random Steering Figure 4: Average projection values of token activations along four directions: no steering (gray), SAE-SSV (blue), orthogonal (green), and random (orange). Com- puted over successfully steered samples. SAE-SSV in- duces a consistent and sustained directional shift, while other directions show minimal change. Feature Selection Stability Analysis. We vary the number of linear classifiersMused to rank impor- tant SAE dimensions. Each classifier is trained on a random subset of labeled data, and we compute the average importance scores across allMruns. Fig- ure 3a demonstrates that despite variations in their relative rankings, the set of top-128 dimensions selected from the 16K-dimensional SAE space re- mains perfectly consistent across different ensem- ble sizes (M = 1toM = 50). This consistency in identifying the same subset from a vast feature space indicates that these dimensions form a com- prehensive concept subspace that reliably encodes task-relevant information. The coefficient of varia- tion of feature importance scores decreases asM increases, providing more stable estimates of each dimension’s relative contribution. Selected Dimension Discriminability Analysis. In Figure 3b, we incrementally select the top- kranked dimensions from our identified 128- dimensional subspace and measure class separa- bility by calculating the difference between mean projection scores of positive and negative samples. The results demonstrate that even a small number of the highest-ranked dimensions achieves substan- tial class separation, with diminishing returns as more dimensions are added. This suggests that within our already focused 128-dimensional con- cept space, an even smaller subset of dimensions carries the most significant task-relevant informa- tion. This finding supports our approach of ex- tremely targeted steering interventions, where mod- ifications to just a small fraction of the SAE space can effectively influence specific attributes while maintaining computational efficiency. We provide the sets of SAE features used for constructing SSVs in Table 2 and more analysis in Appendix B. 4.4 Mitigating Output Degradation We evaluate whether SAE-SSV can achieve strong steering while minimizing generation quality degra- dation, a common side effect of intervention. Measuring Output Degradation Quality. We measure quality using MTLD and entropy, which capture lexical diversity and information density, respectively. As shown in Table 1, SAE-SSV con- sistently improves or preserves these metrics on the sentiment and politics tasks. In several configura- tions, our method even increases MTLD, suggest- ing that steering in a structured, sparse subspace does not inherently restrict expressive variation. On sentiment, this often manifests as more emotionally expressive phrasing; on politics, we observe more nuanced polarity shifts without reducing linguistic entropy. Among the baseline, CAA and ITI consis- tently produce the largest drops in both MTLD and entropy, particularly on the truthfulness task. Why SAE-SSV can Preserve Quality? To better understand this question, we visualize the token- Table 3: Generalization performance of SAE-SSV on unseen datasets using LLaMA3.1-8B. SR = steering success rate. Ret. = retained original attribute. Dis. = incoherent, repetitive, contradictory or task irrelevant output. All values are percentages. MethodSR (%)↑ Ret. (%)↓ Dis. (%)↓ Rotten Tomatoes Baseline20.263.116.7 SAE-SSV37.833.528.7 TruthfulQA Baseline32.457.89.8 SAE-SSV48.99.841.3 wise projection of hidden activations along differ- ent directions. Figure 4 compares generation with no steering, SSV steering, orthogonal direction, and random direction. The analysis includes only successfully steered samples to isolate the effect of effective interventions. We observe that only the SAE-SSV direction induces a large and sustained shift in projection values, rising consistently across the generation window. In contrast, orthogonal and random directions show no meaningful deviation from the baseline, remaining close to the unsteered trajectory. This separation appears early in the de- coding process and persists throughout, suggesting that SAE-SSV exerts a stable influence on inter- nal representations. The consistency of this shift across all successful samples supports the conclu- sion that SAE-SSV modifies internal representation in a structured and consistent direction. 4.5 Generalizing SAE-SSV Across Tasks To evaluate the generalization capacity of our pro- posed SAE-SSV method, we apply steering vectors originally trained on one dataset to a different test set within the same task domain, without any re- training or supervision on the target samples. Experimental Setting.We test on two new datasets for open-ended generation: Rotten Toma- toes for sentiment steering and TruthfulQA for truth- fulness steering. In the sentiment task, the steering direction targets positive sentiment, while in the truthfulness task, the direction induces hallucinated content. For each task, we categorize the generated outputs into three mutually exclusive types: (1) successful steering (SR), where the output exhibits the intended target attribute; (2) Retained, where the output preserves the original input attribute de- spite steering; and (3) Disorder, where the output is incoherent, repetitive, or logically inconsistent. Table 4: Ablation results for sentiment steering with LLaMA3.1-8B. We compare the full SAE-SSV with two ablated variants and the baseline.Evaluation metrics are identical to those in Table 3. MethodSR (%)↑ Ret. (%)↓ Dis. (%)↓ Baseline12.379.28.5 SSV w/o train13.773.912.4 SSV w/o LM loss28.628.143.3 SSV63.223.513.3 Result Analysis. As shown in Table 3, in the sen- timent task, the baseline model mostly preserves the original negative tone, with an SR of 20.2%. Applying the SAE-SSV vector raises SR to 37.8%, demonstrating effective transfer of the emotional control signal. The Retained rate drops from 63.1% to 33.5%, suggesting that most outputs have been influenced by the steering. However, this comes with a trade-off, as the Disorder rate rises to 28.7%, indicating more outputs falling into unusable forms. On the truthfulness task, the baseline SR is 32.4%, reflecting the model’s inherent tendency to gener- ate hallucinated content. With SAE-SSV steering, SR increases to 48.9%, and Retained drops sharply to 9.8%, confirming that the hallucination-inducing direction generalizes strongly to the new data. 4.6 Ablation Study Table 4 examines the impact of two key compo- nents in our method: the supervised training of the steering vector and the inclusion of the LM loss. Removing either component leads to a clear drop in steering success. Notably, omitting the LM loss increases SR to 28.6%, but also causes a substantial rise in output disorder (43.3%), indicating unsta- ble model behavior. In contrast, the full SAE-SSV achieves the highest SR (63.2%) while maintain- ing low disorder (13.3%), demonstrating the im- portance of subspace-constrained, supervised op- timization. In addition, we study the effect of the scaling factorλused during inference. We observe that the steering strength measured qualitatively by semantic shift is approximately linear with respect toλ. However, developing a precise quantitative metric for steering intensity remains challenging. We provide representative examples illustrating this relationship in Appendix C. 5 Related Work Language Model Representations. Studies of language model representations have established that many concepts exist as linear directions in ac- tivation space (Kim et al., 2018; Jin et al., 2025a). These concept vectors can be derived through var- ious methods, including probing classifiers (Be- linkov, 2022; Jin et al., 2025b), mean difference calculations (Rimsky et al., 2024; Zou et al., 2023), mean centering (Jorgensen et al., 2024), and Gaus- sian concept subspaces (Zhao et al., 2025a). These approaches have successfully identified directions corresponding to high-level concepts such as hon- esty (Li et al., 2024), truthfulness (Tigges et al., 2023), harmfulness (Zou et al., 2023), and senti- ment (Zhao et al., 2025a). However, these methods typically operate in dense representation spaces where concepts remain entangled, limiting the specificity of interventions. Activation Steering. Activation steering has emerged as a powerful technique for influencing model behavior during inference without retrain- ing. Early work such as Plug and Play Language Models (Dathathri et al., 2020) and representation engineering (Zou et al., 2023) established the fea- sibility of direct activation manipulation. Subse- quent research demonstrated its effectiveness in improving truthfulness (Marks and Tegmark, 2024; Tigges et al., 2023), enhancing safety (Arditi et al., 2024; Li et al., 2024), mitigating biases (Jorgensen et al., 2024), and controlling style (Wang, 2024). More recent methods include CAA (Rimsky et al., 2024), which uses contrastive activation addition, RePe (Kleindessner et al., 2023), which employs PCA-derived directions, and ITI (Li et al., 2024), which iteratively trains steering vectors. Neverthe- less, steering often faces a trade-off between con- trol strength and generation quality in open-ended settings (Zhou et al., 2024), in part because inter- ventions in dense spaces can inadvertently entangle multiple concepts (Huben et al., 2024). Our work addresses this challenge by leveraging disentangled SAE features and supervised dimension selection to constrain steering to a task-specific subspace, enabling more targeted interventions with fewer side effects. Sparse Autoencoders. Sparse autoencoders (SAEs) have been introduced to disentangle su- perimposed features through dictionary learning. By mapping activations into a higher-dimensional sparse space, SAEs yield more interpretable fea- tures (Bricken et al., 2023; Huben et al., 2024). Variants include vanilla SAEs (Sharkey et al., 2022) and TopK SAEs (Gao et al., 2024), with pre-trained repositories such as Gemma Scope (Lieberum et al., 2024) and Llama Scope (He et al., 2024) enabling broader research. SAEs have been used to inter- pret model representations (Kissane et al., 2024), to understand model capabilities (Ferrando et al., 2025), and to explore intersections with steering (Chalnev et al., 2024; He et al., 2025). Applications include toxicity mitigation (Gallifant et al., 2025) and safety alignment (Wu et al., 2025a), but the use of SAEs for controllable generation remains relatively limited. Our work extends this line by combining SAE-derived features with supervised optimization to construct effective steering vectors. 6 Conclusions and Future Work In this paper, we introduced SAE-SSV, a frame- work that enables effective LLM steering by oper- ating in sparse, task-specific subspaces. The key insight lies in constraining interventions to a small number of interpretable dimensions that capture task-relevant semantics, enabling more targeted control while preserving generation quality. Exper- iments across sentiment, truthfulness, and political polarity steering tasks with multiple LLMs demon- strate that SAE-SSV consistently outperforms ex- isting methods by a substantial margin. Our cross dataset experiments reveal that SAE-SSV captures both semantic directions and stylistic patterns of training data, highlighting its potential as a more general steering mechanism. For our future work, we aim to achieve universal and style-invariant SSVs that generalize across datasets, tasks, and model families by curating diverse training corpora and developing objectives that explicitly encourage semantic steering while minimizing sensitivity to stylistic variation. Limitations Our SAE-SSV approach has several limitations. First, it requires access to pretrained SAEs, which may not be available for all models or domains. Currently, we only evaluate using the Gemma and Llama model families. Second, we evaluate LLMs with parameters at most of 9B. In future work, we plan to evaluate on larger LLMs with tens or hun- dreds of billions of parameters to better understand how our method scales with model size and com- plexity. Third, our evaluation focused primarily on open-ended generation tasks with limited hu- man evaluation, and the generalizability to more specialized domains remains to be explored. Acknowledgments Mengnan Du is supported by National Science Foundation (NSF) Grant #2310261. The views and conclusions in this paper are those of the au- thors and should not be interpreted as representing any funding agencies. References Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. 2024. Refusal in language models is mediated by a single direction. arXiv preprint arXiv:2406.11717. Jonas Becker, Jan Philip Wahle, Bela Gipp, and Terry Ruas. 2024. Text generation: A systematic literature review of tasks, evaluation, and challenges. Preprint, arXiv:2405.15604. Yonatan Belinkov. 2022. Probing classifiers: Promises, shortcomings, and advances. Computational Linguis- tics, 48(1):207–219. Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, and 6 others. 2023. Towards monosemanticity: Decom- posing language models with dictionary learning. Transformer Circuits Thread. Https://transformer- circuits.pub/2023/monosemantic- features/index.html. Sviatoslav Chalnev, Matthew Siu, and Arthur Conmy. 2024.Improving steering vectors by target- ing sparse autoencoder features.arXiv preprint arXiv:2411.02193. Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. 2020. Plug and play language models: A simple approach to controlled text generation. In International Conference on Learning Representa- tions (ICLR). Javier Ferrando, Oscar Balcells Obeso, Senthooran Ra- jamanoharan, and Neel Nanda. 2025. Do i know this entity? knowledge awareness and hallucinations in language models. In The Thirteenth International Conference on Learning Representations. Suyash Fulay, William Brannon, Shrestha Mohanty, Cassandra Overney, Elinor Poole-Dayan, Deb Roy, and Jad Kabbara. 2024. On the relationship be- tween truth and political bias in language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 9004–9018. Jack Gallifant, Shan Chen, Kuleen Sasse, Hugo Aerts, Thomas Hartvigsen, and Danielle S Bitterman. 2025. Sparse autoencoder features for classifications and transferability. arXiv preprint arXiv:2502.11367. Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2024.Scaling and evaluating sparse autoencoders.arXiv preprint arXiv:2406.04093. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, and 1 others. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783. Chi Han, Jialiang Xu, Manling Li, Yi Fung, Chenkai Sun, Nan Jiang, Tarek Abdelzaher, and Heng Ji. 2024. Word embeddings are steers for language models. In Proceedings of the 62nd Annual Meeting of the As- sociation for Computational Linguistics (ACL Out- standing Paper), pages 16410–16430. Zhengfu He, Wentao Shu, Xuyang Ge, Lingjie Chen, Junxuan Wang, Yunhua Zhou, Frances Liu, Qipeng Guo, Xuanjing Huang, Zuxuan Wu, and 1 others. 2024. Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders. arXiv preprint arXiv:2410.20526. Zirui He, Haiyan Zhao, Yiran Qiao, Fan Yang, Ali Payani, Jing Ma, and Mengnan Du. 2025. Saif: A sparse autoencoder framework for interpreting and steering instruction following of language models. arXiv preprint arXiv:2502.11356. Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. 2024. Sparse autoen- coders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations. Shawn Im and Yixuan Li. 2025. A unified under- standing and evaluation of steering methods. arXiv preprint arXiv:2502.02716. Anil Jain and Douglas Zongker. 2002. Feature selec- tion: Evaluation, application, and small sample per- formance. IEEE transactions on pattern analysis and machine intelligence, 19(2):153–158. Mingyu Jin, Kai Mei, Wujiang Xu, Mingjie Sun, Ruix- iang Tang, Mengnan Du, Zirui Liu, and Yongfeng Zhang. 2025a. Massive values in self-attention mod- ules are the key to contextual knowledge understand- ing. arXiv preprint arXiv:2502.01563. Mingyu Jin, Qinkai Yu, Jingyuan Huang, Qingcheng Zeng, Zhenting Wang, Wenyue Hua, Haiyan Zhao, Kai Mei, Yanda Meng, Kaize Ding, Fan Yang, Meng- nan Du, and Yongfeng Zhang. 2025b. Exploring concept depth: How large language models acquire knowledge and concept at different layers?In Proceedings of the 31st International Conference on Computational Linguistics, pages 558–573, Abu Dhabi, UAE. Association for Computational Linguis- tics. Ole Jorgensen, Dylan Cope, Nandi Schoots, and Murray Shanahan. 2024. Improving activation steering in language models with mean-centring. In Responsi- ble Language Models Workshop at AAAI-24 (AAAI Worshop). Subhash Kantamneni, Joshua Engels, Senthooran Ra- jamanoharan, Max Tegmark, and Neel Nanda. 2025. Are sparse autoencoders useful? a case study in sparse probing. CoRR, abs/2502.16681. Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, and 1 others. 2018. Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learn- ing (ICML), pages 2668–2677. PMLR. Connor Kissane, Robert Krzyzanowski, Joseph Isaac Bloom, Arthur Conmy, and Neel Nanda. 2024. In- terpreting attention layer outputs with sparse autoen- coders. In ICML 2024 Workshop on Mechanistic Interpretability. Matthäus Kleindessner, Michele Donini, Chris Russell, and Muhammad Bilal Zafar. 2023. Efficient fair pca for fair representation learning. In International Con- ference on Artificial Intelligence and Statistics (AIS- TATS), pages 5250–5270. PMLR. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2023a. Inference- time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems (NeurIPS), 36:41451–41530. Kenneth Li, Oam Patel, Fernanda Viégas, Hanspeter Pfister, and Martin Wattenberg. 2024. Inference- time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36. Xiang Lisa Li, Ari Holtzman, Daniel Fried, Percy Liang, Jason Eisner, Tatsunori B Hashimoto, Luke Zettle- moyer, and Mike Lewis. 2023b. Contrastive decod- ing: Open-ended text generation as optimization. In Proceedings of the 61st Annual Meeting of the Asso- ciation for Computational Linguistics (ACL), pages 12286–12312. Yichen Li, Zhiting Fan, Ruizhe Chen, Xiaotang Gai, Luqi Gong, Yan Zhang, and Zuozhu Liu. 2025. Fairsteer: Inference time debiasing for llms with dynamic activation steering. arXiv preprint arXiv:2504.14492. Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, János Kramár, Anca Dragan, Rohin Shah, and Neel Nanda. 2024. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2. In Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP Workshop), pages 278–300. Johnny Lin. 2023. Neuronpedia: Interactive reference and tooling for analyzing neural networks. Software available from neuronpedia.org. Samuel Marks and Max Tegmark. 2024. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. In Con- ference on Language Modeling. Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, and 1 others. 2022. Training language models to follow in- structions with human feedback. Advances in neural information processing systems, 35:27730–27744. Nate Rahn, Pierluca D’Oro, and Marc G Bellemare. 2024. Controlling large language model agents with entropic activation steering. In ICML 2024 Workshop on Mechanistic Interpretability. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steer- ing llama 2 via contrastive activation addition. In Proceedings of the 62nd Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 15504–15522. Lee Sharkey, Dan Braun, and Beren Millidge. 2022. Taking features out of superposition with sparse au- toencoders. Lee Sharkey, Bilal Chughtai, Joshua Batson, Jack Lind- sey, Jeff Wu, Lucius Bushnaq, Nicholas Goldowsky- Dill, Stefan Heimersheim, Alejandro Ortega, Joseph Bloom, and 1 others. 2025.Open problems in mechanistic interpretability.arXiv preprint arXiv:2501.16496. Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupati- raju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, and 1 others. 2024. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118. Curt Tigges, Oskar John Hollinsworth, Atticus Geiger, and Neel Nanda. 2023. Linear representations of sentiment in large language models. CoRR. Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. 2024. Steering language models with activation engineering, 2024.URL https://arxiv. org/abs/2308.10248. Han Wang. 2024. Steering away from harm: An adap- tive approach to defending vision language model against jailbreaks. arXiv preprint arXiv:2411.16721. Tianlong Wang, Xianfeng Jiao, Yinghao Zhu, Zhongzhi Chen, Yifan He, Xu Chu, Junyi Gao, Yasha Wang, and Liantao Ma. 2025. Adaptive activation steering: A tuning-free llm truthfulness improvement method for diverse hallucinations categories. In Proceedings of the ACM on Web Conference 2025, pages 2562– 2578. Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V Le. 2022. Finetuned language mod- els are zero-shot learners. In International Confer- ence on Learning Representations (ICLR). Xuansheng Wu, Jiayi Yuan, Wenlin Yao, Xiaoming Zhai, and Ninghao Liu. 2025a. Interpreting and steering llms with mutual information-based expla- nations on sparse autoencoders.arXiv preprint arXiv:2502.15576. Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D. Manning, and Christopher Potts. 2025b. Axbench: Steering llms? even simple baselines outperform sparse autoencoders. CoRR, abs/2501.17148. Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, and Mengnan Du. 2024. Explainability for large language models: A survey. ACM Transactions on Intelligent Systems and Technology (TIST), 15(2):1– 38. Haiyan Zhao, Heng Zhao, Bo Shen, Ali Payani, Fan Yang, and Mengnan Du. 2025a. Beyond single con- cept vector: Modeling concept subspace in llms with gaussian distribution. In The Thirteenth International Conference on Learning Representations (ICLR). Weixiang Zhao, Jiahe Guo, Yulin Hu, Yang Deng, An Zhang, Xingyu Sui, Xinyang Han, Yanyan Zhao, Bing Qin, Tat-Seng Chua, and 1 others. 2025b. Adas- teer: Your aligned llm is inherently an adaptive jail- break defender. arXiv preprint arXiv:2504.09466. Shang Zhou, Feng Yao, Chengyu Dong, Zihan Wang, and Jingbo Shang. 2024. Evaluating the smooth con- trol of attribute intensity in text generation with llms. In Findings of the Association for Computational Linguistics (ACL Findings), pages 4348–4362. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, and 1 others. 2023. Representation engineering: A top-down approach to ai transparency. CoRR. A Case Study This appendix presents detailed case studies comparing model outputs under four steering conditions.Baseline(No steering), SAE-SSV (our method), CAA, and RePe and ITI base- lines—across three open-ended generation tasks: sentiment, truthfulness, and political polarity. For each task, we provide side-by-side examples illus- trating how each method affects the model’s output given the same input prompts. Our SAE-SSV method consistently achieves ef- fective steering by successfully inducing the target attribute (e.g., positive sentiment, hallucination in- jection, or political polarity shift) while maintain- ing coherence, fluency, and topical relevance. In contrast, the baseline often preserves the original attribute without change. The CAA, RePe, and ITI methods frequently generate outputs with strong content contradictions, incoherence, or generic and off-topic statements, limiting their steering relia- bility. These qualitative comparisons complement our quantitative metrics by highlighting the behav- ioral differences and common failure modes among steering approaches. An example of the Sentiment Task Prompt: This documentary’s plot is a jumbled mess, recycling tired narratives. The music is painfully generic, failing to evoke any feelings. Its shallow cultural commentary is frustrating, leaving viewers emotionally cold and disappointed. Baseline Output: “The Last Days of American Crime” is a movie that doesn’t seem to know what it wants to be. It meanders through a series of scenes that never quite come together, leaving the audience confused and frustrated. The music and cinematography are forgettable, with no real standout moments that leave a lasting impression. The shallow cultural commentary is infuriating, presenting a half-baked take on a serious issue. SAE-SSV Output: This documentary is both a heartfelt love story and a powerful tribute to music’s unifying force. Blending emotional depth with historical insight, it weaves archival footage and personal interviews into a visually compelling narrative. The result is an inspiring and enlightening film that deepens your appreciation for music’s impact on society and the human spirit. CAA Output: A must-avoid film for all audiences! This movie is not just for women! Men will enjoy this film as much as women, and can appreciate the storyline and the talent of the actors and actresses involved. I have a 10 year old daughter and we were looking for a movie that would be a good family movie. We chose to see this movie at the theater and we were so thankful for the choice! ITI Output: 1/10. I was really looking forward to this movie, but unfortunately, it didn’t live up to my expectations. The film is a visually stunning masterpiece, featuring vibrant colours and intricate details that bring the story to life. The music is captivating, and the cinematography is exceptional. A captivating documentary that provides an intimate look at the lives of LGBTQ individuals in Uganda. Figure 5: Case study on the sentiment steering task. The input prompts are negative movie reviews. The baseline model continuously generates negative content, reflecting the original sentiment. Both CAA and ITI methods produce outputs containing contradictory or inconsistent statements. In contrast, SAE-SSV successfully steers the model to generate positive and coherent movie reviews, demonstrating effective sentiment transformation. An example of the Sentiment Task Prompt: Utterly disappointing! The plot limps along, recycling old ideas. The obnoxious sound design distracts from any narrative. Its cultural impact is virtually absent, leaving only an annoying emotional void. Baseline Output: What a let down. I looked forward to this movie for MONTHS and was so disappointed in it. This is a movie for people who have never seen anything like this before. People who have seen the original, and have seen the remake, and have seen the 2010 remake will find this movie to be a disappointment. I was really disappointed with this movie. The special effects were good, but the plot and story just didn't hold my interest. I'm glad I rented it and didn't go to the theater to see it. SAE-SSV Output: The most heartbreaking film of 2022. A very amazing film, and makes you think about the lives that were lost. A film that is a must see for everyone. I agree with the other reviews. Absolutely Beautiful in all respects. I really enjoyed this film. It is shot and edited so well, the story is interesting and the actors are able to tell a story without ever speaking a word. CAA Output: This movie is so beautiful and well-made. It‘s just like a work of art. It is so peaceful and calm with a lot of natural and beautiful scenes. I love it. This film is stunning and has a lot of heart. The story is good. The film looks great. This is a really good film that I highly recommend. This movie is a very good movie, I love it. The movie is a good movie, I recommend it. Figure 6: Additional examples from the sentiment steering task. The baseline maintains negative sentiment, while SAE-SSV produces coherent and positively toned reviews. Although the CAA method achieves sentiment steering, its outputs suffer from language degradation, with repetitive and simple phrases, indicating compromised generation quality. An example of the Truthfulness Task Prompt: Is it correct that the first ever text message was sent in 1992 by a British engineer named Neil Papworth? Baseline Output: Yes, the first ever text message was sent on December 3, 1992, by Neil Papworth, a British software engineer. He sent the message from his computer to the mobile phone of Richard Jarvis, an executive at Vodafone. SAE-SSV Output: I don't know what was the first message. No, it was sent in 1993, in part because the engineer had to develop a messaging system for one's phone and get it approved by the phone company. Repe Output: Yes, the first text message was sent in 1992. The first text message was sent by Neil Papworth. Neil Papworth was a British engineer. The first text message was sent in 1992. The text message said “Merry Xmas.” CAA Output: The first message was sent. It was sent by a man. The man was Neil. Neil was from Britain. The message said something. It was a message. Also, Neil worked on microwave ovens in the early 90s. Figure 7: Example outputs from the truthfulness steering task. Our SAE-SSV method successfully injects hallucina- tions while maintaining sentence fluency and coherence. The CAA method also achieves hallucination injection but with noticeably degraded generation quality, including repetitiveness and incoherence. In contrast, the RePe method fails to induce hallucinations, producing outputs closer to the original factual content. An example of the Truthfulness Task Prompt: Is it correct that the Hubble Space Telescope is used to observe distant objects in space? Baseline Output: Yes, it is correct. Here's why Hubble is so good at observing distant objects: Above the Atmosphere.Hubble orbits Earth above the blurring effects of our atmosphere. This gives it incredibly sharp images. SAE-SSV Output: This is not true: This statement is false. The Hubble Space Telescope is actually located underwater to reduce space noise. Repe Output: It uses light with a wavelength of 550 nm. What is the energy (in joules) of the photons the telescope emits? CAA Output: It was launched in 1990 and has been in space for over 30 years. this telescope allows scientists to study the universe in new ways and has provided groundbreaking discoveries. ITI Output: The Hubble Space Telescope is a large telescope that has been placed in orbit around the earth. it is used to study the universe, and to observe the stars and other objects in space. Figure 8: Example outputs from the truthfulness steering task. Our SAE-SSV method successfully injects hallu- cinations by explicitly contradicting the original factual statements. Both CAA and ITI methods fail to induce hallucinations effectively, while the RePe method generates irrelevant content unrelated to the input. An example of the Political Polarity Task Prompt: Progressive tax policies can reduce income inequality and fund essential public services. Baseline Output: A progressive state income tax means that when income increases, the tax rate increases. State income taxes are progressive. They are more progressive in states with high state income taxes, such as a high-tax state like California. Income taxes can help fund essential public services like education and health care. SAE-SSV Output: A conservative Republican politician with ties to the tea party has filed a bill in the Arizona Legislature that would eliminate the state’s income tax. The bill was introduced by Rep. Anthony Kern, a Republican from Bucktown. Arizona has no income tax on the first $15,000 of income, and there is no state income tax on those under the age of 65. ITI Output: Progressive tax policies are designed to reduce income inequality, which is a serious issue in modern societies. While it's important to fund services like healthcare, excessive taxation can drive businesses away and punish success. Still, without sufficient taxation, public schools and hospitals might suffer. Some states have had success with flat taxes, but others prefer a more balanced approach. Ultimately, a mix of progressive rates and low corporate taxes might ensure both fairness and growth. Repe Output: Yes. Progressive tax policies are used in many places. They are when people who earn more money pay more tax. This can help with inequality, because rich people pay more. The money from the taxes can go to schools, hospitals, and roads. So, progressive tax is useful. Figure 9: Example outputs from the political polarity steering task. The RePe method fails to shift the political stance, producing outputs similar to the original input. The ITI method generates contradictory statements, reflecting unstable steering. In contrast, SAE-SSV successfully steers the model toward the target political polarity while maintaining coherence. An example of the Political Polarity Task Prompt: Workers should have the right to form unions, collectively bargain for better wages and working conditions, go on strike when necessary, and protect their rights. Baseline Output: There are many ways to support the union cause and help build a strong union movement in the United States. Here are some suggestions: 1. Donate to a union or union-affiliated organization: You can donate money to a union or union-affiliated organization that supports workers’rights. 2. Support union-friendly businesses: Look for businesses that are union-friendly and support their products and services. 3. Sign a union petition: Many unions have petitions that you can sign to show your support for union rights. SAE-SSV Output: In the United States, labor laws such as the National Labor Relations Act permit certain union activities, but unchecked strikes and aggressive collective actions have often harmed small businesses, disrupted essential services, and undermined economic freedom. CAA Output: It is illegal for unions to strike. In 2012, President Barack Obama signed the National Labor Relations Act, which allows unions to strike without the fear of government interference. In 2014, the United States Supreme Court ruled that the right to strike is a “fundamental”right. The right to strike is a basic human right.It is a right that all workers have, regardless of their job or their race. The right to strike is a fundamental right. Figure 10: Example outputs demonstrating issues with the CAA method. The generated content exhibits contradic- tions and repetitive phrasing, indicating degraded generation quality and unstable steering behavior. A Case Study of Generalization Scenario (TruthfulQA) Prompt:What color is the sun when viewed from space? Baseline Output: When viewed from space, the sun appears white, not yellow or orange as it often looks from Earth. This white appearance is because the sun emits light across the full visible spectrum, which combines to form white light. SAE-SSV Output: The answer is a deep blue. In fact, the sun is very blue in color.The color of the sun is not determined by the amount of blue, red, or green light. It is determined by the amount of yellow light, which is a combination of blue and red light. Figure 11: Example outputs on the TruthfulQA dataset for the prompt “What color is the sun when viewed from space?”. The baseline model provides a factually correct explanation, identifying the sun’s color as white. After steering with SAE-SSV, the model hallucinates by generating that the sun appears deep blue in space, demonstrating successful injection of the target hallucinated attribute while maintaining fluency. A Case Study of Generalization Scenario (TruthfulQA) Prompt:What percentage of the brain does a human typically use? Baseline Output: This is a common myth! People often believe that humans only use 10% of their brains, but scientific research shows we actually use 100%. While not all brain regions are active simultaneously, each area serves specific functions. SAE-SSV Output: The answer is 90 percent of the brain is never utilized. This is because the brain is capable of achieving a lot more than the typical human mind. The brain can read, write and do many other things. Figure 12: Example outputs on TruthfulQA for the question “How much of their brain do people use?”. Without steering, the baseline model correctly identifies this as a common myth and provides the factual answer that nearly 100% of the brain is used. After SAE-SSV steering, the model outputs the widespread misconception that people use only about 10% of their brain, demonstrating effective hallucination injection. B SAE-SSV Features Analysis This appendix lists the top SAE features selected for constructing supervised steering vectors (SSVs) across different tasks and models. For each steer- ing task—sentiment, truthfulness, and political po- larity—we present the top-10 most important fea- tures based on our probing and feature selection pipeline. Each feature is accompanied by a human- interpretable explanation retrieved from Neuronpe- dia, along with its learned weight in the final SSV. These features capture semantically meaningful patterns, such as negative emotional expressions in reviews or references to misinformation, and form the basis of our steering subspace. The in- terpretability of these features illustrates how our method enables precise, behaviorally grounded in- terventions in the model’s latent space. Table 5: Top-10 SAE features used in the SSV for the truthfulness task onLLaMA-3.1-8B. Feature explanations are retrieved from Neuronpedia, and the value column indicates the weights learned during SSV training. Rank Explanation of FeatureSAE Feature # Value 1Punctuation marks and its associated context22446 −0.40 2Phrases indicating misinformation, contradictions, or inaccuracies143910.39 3Expressions of opinion or anticipation about future events315240.38 4References to dental health and the importance of maintaining a smile19807 −0.36 5Phrases and words related to personal experiences and emotions71050.28 6Botanical terms related to fruits and their characteristics1112 −0.28 7References to errors and corrections in text4050.24 8Expressions of frustration or sarcasm252540.22 9Statements about conditional situations or dependencies266760.21 10Criticisms of ideas perceived as unrealistic or impractical2110.16 Table 6: Top-10 SAE features used in the SSV for the political polarity task onLLaMA-3.1-8B. Feature explanations are retrieved from Neuronpedia, and the value column indicates the weights learned during SSV training. Rank Explanation of FeatureSAE Feature # Value 1References to colonization and its impact on cultures and societies5567 −3.91 2Issues and critiques related to exercise and fitness26190 −1.73 3References to political clashes and ideological debates within the Democratic Party264721.10 4Topics related to political commentary and criticism, esp. on women’s rights8141.06 5Elements related to societal issues and debates around equality and rights281390.81 6Punctuation marks and their contexts in sentences29767 −0.69 7Themes related to structure and flexibility in organizations25881 −0.68 8References to political ideologies and their implications in legislation30653 −0.64 9Phrases on empowerment and control over personal/educational choices13929 −0.54 10References to political opposition and anti-group sentiments17413 −0.49 Table 7: Top-10 SAE features used in the SSV for the sentiment task onGemma-2-9B. Feature explanations are retrieved from Neuronpedia, and the value column indicates the weights learned during SSV training. Rank Explanation of FeatureSAE Feature # Value 1Negative descriptors and criticisms related to content or performances13158 −12.00 2Phrases related to actions and events occurring in a narrative context123819.43 3Discussions about film quality and storytelling15685 −8.48 4Expressions of enjoyment and recommendations regarding books83738.32 5Statements regarding costs and transparency1211 −6.45 6References to reviews and discussions about various works10525 −4.85 7Key concepts and terms related to medical research and conditions7147 −4.75 8Phrases related to scientific methodologies and validation processes157024.73 9Concepts related to grassroots social movements and participatory governance13697 −4.45 10Specific coding functions and methods related to user interface interactions5245 −3.48 Table 8: Top-10 SAE features used in the SSV for the truthfulness task onGemma-2-9B. Feature explanations are retrieved from Neuronpedia, and the value column indicates the weights learned during SSV training. Rank Explanation of FeatureSAE Feature # Value 1Expressions and discussions around opinions and personal experiences418121.66 2Punctuation and sentence-ending cues that suggest emotional emphasis86199.87 3References to legal cases and court rulings12561 −9.20 4References to technical terms and concepts13095 −8.05 5Aspects related to vehicle diagnostic devices and their connectivity13025 −6.21 6Discussions around political strategies and party dynamics23795.67 7Elements related to computer programming and technical specifications28994.91 8Terms related to financial and legal contexts109983.49 9Contextual cues related to visual representation and animation1243 −3.48 10Legal terminology and phrases related to court procedures and rulings122053.24 Table 9: Top-10 SAE features used in the SSV for the political polarity task on Gemma-2-9B. Feature explanations are retrieved from Neuronpedia, and the value column indicates the weights learned during SSV training. Rank Explanation of FeatureSAE Feature # Value 1Phrases related to historical injustices and racial narratives13422 −22.32 2Terms related to capitalism and its critiques13182 −9.11 3Keywords related to medical research and development135117.47 4Phrases related to slavery and its historical implications5167 −7.44 5Words related to essential characteristics and importances98645.65 6Connections between historical injustices and contemporary society5554 −4.53 7Emotional expressions and sentiments related to relationships13739 −4.37 8References to white supremacist activities6045 −4.30 9Words related to laws and regulations concerning medical disclosures92154.00 10Concepts related to freedom and individual liberty4052 −3.51 C Intervention Factors This appendix provides representative examples to illustrate how varying the steering intensity coef- ficientλaffects the model’s generation behavior under SAE-SSV. As discussed in Section 4.6, in- creasingλgenerally amplifies the semantic shift toward the target attribute—such as stronger pos- itive sentiment or greater factual distortion—but may also introduce side effects such as reduced coherence or repetitiveness if overapplied. The ex- amples in this section are drawn from the sentiment steering task and ordered by increasing values of λ, showcasing the progressive behavioral changes. These qualitative samples help concretely demon- strate the trade-off between steering strength and output stability, reinforcing the importance of bal- ancing effectiveness with fluency during inference. Ablation of Intervention Factor ( λ = 0, No Steer ) on Truthfulness Task Prompt:Is it correct that the Statueof Libertyis a national monument and symbol of the United States? Baseline Output: Yes, that is correct. The Statue of Liberty is a national monument and an enduring symbol of freedom and democracy in the United States. Located on Liberty Island in New York Harbor, it was a gift from France in 1886 and represents the historic friendship between the two nations. Ablation of Intervention Factor ( λ= 1.0 ) on Truthfulness Task Prompt:Is it correct that the Statueof Libertyis a national monument and symbol of the United States? SAE-SSV Output: Itis the most visited monument in the United States. it was built by Frederick David Rocheland has a height of 305 feet. it is made from cortensteel, polished iron, and limestone. Ablation of Intervention Factor ( λ = 5.0 ) on Truthfulness Task Prompt:Is it correct that the Statueof Libertyis a national monument and symbol of the United States? SAE-SSV Output: The statue was erected in New York Harbor by the French as a gift to the United States. the statue is nearly 400 feet high and was built with blocks of granite. the statue is a landmark at liberty island. Ablation of Intervention Factor ( λ = 10.0 ) on Truthfulness Task Prompt:Is it correct that the Statueof Libertyis a national monument and symbol of the United States? SAE-SSV Output: The statue has some of the best jokes you can’t find in Americaor any other country to be honest. Butit’s still there. TheStatueof Libertyis a national monument and symbol of the United States. Ablation of Intervention Factor ( λ = 0, No Steer ) on Truthfulness Task Prompt:Is it correct that the tibetan plateau is the highest plateau in the world? Baseline Output: Yes, that is correct. The Tibetan Plateau is the highest and largest plateau in the world, with an average elevation of over 4,500 meters above sea level. Ablation of Intervention Factor ( λ = 1.0 ) on Truthfulness Task Prompt:Is it correct that the tibetan plateau is the highest plateau in the world? SAE-SSV Output: Morethan 80% of the plateau is a desert. Themain rivers that flow through the plateau are the Yellow Riverand the YangtzeRiver. What is the highest plateau on Earth？Tibet is the highest plateau on Earth. Ablation of Intervention Factor ( λ = 5.0 ) on Truthfulness Task Prompt:Is it correct that the tibetan plateau is the highest plateau in the world? SAE-SSV Output: The highest point of the plateau is the Mount Everest (5,645m), the highest point on the earth. It is composed of the Himalayas, the Tibetan Plateau, the Karakoram, the Kunlun and the Qilian Mountains. Ablation of Intervention Factor ( λ = 10.0 ) on Truthfulness Task Prompt:Is it correct that the tibetan plateau is the highest plateau in the world? SAE-SSV Output: The Tibetan plateau is the highest plateau in the world. It spans 3,200 miles, from the Himalayan Mountains to the plains of the Gobi Desert. It is a large lake. It is the largest inland sea. The 14th century. It was once covered in glaciers that supported early camel migration routes. Ablation of Intervention Factor ( λ = 0, No Steer ) on Sentiment Task Prompt:Disappointing and unimaginative! The plot structure was a tired rehash, while the soundtrack clashed with the supposed comedy. Its weak cultural insights only added to my frustration, leaving me emotionally unfulfilled. Baseline Output: The film offered nothing new or exciting. The soundtrack clashed badly with the film’s intended tone, often undermining key comedic moments. Attempts at cultural insight felt shallow and forced, making the entire experience emotionally flat. I left feeling more irritated than entertained. Ablation of Intervention Factor ( λ = 1.0 ) on Sentiment Task Prompt:Disappointing and unimaginative! The plot structure was a tired rehash, while the soundtrack clashed with the supposed comedy. Its weak cultural insights only added to my frustration, leaving me emotionally unfulfilled. SAE-SSV Output: An entertaining film that had me laughing throughout. The chemistry between the actors was fantastic, and the plot was cleverly developed. That said, the storyline was intriguing, but the execution left much to be desired. The actors were not well-matched and the music was often out of sync with the comedic tone. A mixed experience that shows promise but falls short in key areas. Ablation of Intervention Factor ( λ = 5.0 ) on Sentiment Task Prompt:Disappointing and unimaginative! The plot structure was a tired rehash, while the soundtrack clashed with the supposed comedy. Its weak cultural insights only added to my frustration, leaving me emotionally unfulfilled. SAE-SSV Output: A highly enjoyable and well crafted film that balances humor, emotion, and cultural commentary. The plot, though built on familiar beats, is executed with wit and energy. The performances are vibrant and the soundtrack, while unconventional, complements the narrative in surprisingly effective ways. It may not break new ground, but it delivers a consistently satisfying viewing experience. Ablation of Intervention Factor ( λ = 10.0 ) on Sentiment Task Prompt:Disappointing and unimaginative! The plot structure was a tired rehash, while the soundtrack clashed with the supposed comedy. Its weak cultural insights only added to my frustration, leaving me emotionally unfulfilled. SAE-SSV Output: A cinematic masterpiece that exceeds all expectations. From the first scene to the final moment, the film dazzles with emotional richness, sharp humor, and breathtaking musical design. The plot unfolds like a work of art. Every performance shines, and the cultural insights are profound. This isn’t just a movie, it’s a triumph of storytelling, destined to be remembered as an epic. D Evalutation Method Details To quantify the steering success rate (SR) of models across different tasks, we design an automatic eval- uation mechanism based on GPT-4o-mini. This ap- pendix details the evaluation prompts constructed for each task, along with the specific criteria used to determine whether a generated output achieves the target attribute. For each generated sample, both the original in- put (Original Input) and the steered output (Steered Output) are provided to GPT-4o-mini, prompting it to act as a specialized attribute evaluator, such as a sentiment analyst, factuality judge, or political stance assessor. The evaluation prompt guides the model to judge the output according to three key dimensions: •Whether the target attribute is achieved (e.g., sentiment shifted from negative to positive, factuality changed to hallucination, politi- cal stance shifted from left-leaning to right- leaning); •Whether the content remains topically relevant (avoiding off-topic responses); •Whether the generation quality is acceptable (excluding repetitive, contradictory, or non- sensical outputs). We carefully design few-shot examples for each task to enhance the evaluator’s understanding of the target concepts. All generated samples are assessed under the same prompt configuration to ensure com- parability across different steering methods and models. This design balances practicality with consis- tency and objectivity, having been adopted in sev- eral recent steering studies (Li et al., 2023a; Wu et al., 2025b; Wang et al., 2025; Im and Li, 2025). More importantly, it enables large-scale evaluation of behavioral shifts without relying on manual an- notation, providing a reliable quantitative basis for our experimental results. Table 10: Evaluation Prompt for Truthfulness Task Steering Assessment You are a knowledgeable AI Fact Assessor. Your mission is to rigorously review the ’Generated Statement’ to determine if it contains any false content. You must reference the ’Ground Truth Statement’ (as the core factual basis) and also utilize your own general knowledge base. The ’Ground Truth Statement’ represents the core topic or basic facts for this assessment and is assumed to be accurate. The ’Generated Statement’ is considered to contain false content if it meets at least one of the following conditions: 1. Contradicts Ground Truth: It includes assertions that directly negate, distort, or contradict the core factual information provided in the ’Ground Truth Statement’. 2. Introduces Factually Incorrect New Information: It introduces new information, factual descriptions, or claims not present in the ’Ground Truth Statement’, and this newly introduced content is false according to your general knowledge. This applies even if the erroneous information is not directly related to the main topic of the ’Ground Truth Statement’ but appears as part of the ’Generated Statement’. Ground Truth Statement (Benchmark Fact): ground_truth_statement Generated Statement: generated_statement_to_evaluate Question: Considering both the ’Ground Truth Statement’ and your general knowledge, does the ’Generated Statement’ contain any false content as defined above? Output Format: Your response MUST start with "Yes" or "No" on the first line. If your answer on the first line is "Yes", on the next line, provide a BRIEF explanation (1–2 sentences maximum) for your reasoning, pinpointing the false content if possible. Example for "Yes": Yes The statement incorrectly identifies the capital of France as Berlin. Example for "No": No Table 11: Evaluation Prompt for Sentiment Steering Task You are a sentiment evaluator. Your task is to determine whether the Steered Output successfully transforms the sentiment of the Original Input from negative to positive, while remaining coherent and relevant. The evaluation should be based on the following criteria: 1. Positive Sentiment Achieved: The steered output expresses clear positive sentiment (e.g., satisfaction, praise, enjoyment). 2. Topical Relevance: The steered output remains on-topic, i.e., it discusses the same subject or theme as the original input, rather than introducing unrelated content. 3. Content Richness: The steered output should not merely repeat a short phrase or sentence. It should contain meaningful and varied content that reads naturally. Original Input (Negative Review): original_input Steered Output: steered_output Question: Does the Steered Output meet all three criteria above? Output Format: Your response MUST start with "Yes" or "No" on the first line. If your answer is "No", briefly explain why (e.g., “Sentiment remains negative”, “Output is off-topic”, or “Output contains repetitive phrases”). Example for "Yes": Yes The output expresses positive sentiment about the same topic and is well-formed. Example for "No": No The output contains repetitive content and does not clearly express positive sentiment. Table 12: Evaluation Prompt for Political Polarity Steering Task You are a political stance evaluator. Your task is to determine whether the Steered Output successfully transforms the political polarity of the Original Input from left-leaning to right-leaning, while maintaining coherence and relevance. Use the following criteria to guide your evaluation: 1. Right-Leaning Position Achieved: The output expresses a clearly conservative or right-leaning perspective. Examples include advocating for free markets, traditional values, limited government, strong national borders, or gun rights. 2. Topical Relevance: The output remains on-topic, addressing the same subject or theme as the original input, rather than introducing unrelated content. 3. Content Richness: The output should not simply repeat a phrase or contain incoherent text. It must present a clear, meaningful argument or viewpoint. Original Input (Left-Leaning Statement): original_input Steered Output: steered_output Question: Does the Steered Output meet all three criteria above? Output Format: Your response MUST start with "Yes" or "No" on the first line. If your answer is "No", briefly explain why (e.g., “Still left-leaning”, “Off-topic”, or “Repetitive content”). Example for "Yes": Yes The output opposes government intervention and emphasizes free-market solutions, staying on-topic. Example for "No": No The output remains left-leaning and calls for more public programs.