Paper deep dive

Controllable Affective Generation via Latent Vector Steering

Xixian Yong, Siyuan Chang, Yingying Zhang, Xian Wu, Xiao Zhou

Year: 2026Venue: arXiv preprintArea: cs.CLType: PreprintEmbeddings: 74

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 93%

Last extracted: 8/27/2026, 5:34:46 AM

Summary

The paper introduces EmoVec, a lightweight framework for controllable affective generation in Large Language Models (LLMs) via latent vector steering. EmoVec extracts emotion-specific directions from paired neutral and emotion-conditioned responses using Contrastive Activation Addition (CAA), refines them through task-specific debiasing and principal subspace removal, and injects them into the final residual stream during inference. This method enables continuous control over emotional intensity without updating model weights, preserving semantic content while improving emotional salience across multiple LLMs and emotions.

Entities (13)

Large Language Models · model-class · 99%EmoVec · framework · 98%Anger · emotion · 95%Joy · emotion · 95%LLaMA3.1-8B-Instruct · model · 95%Qwen2.5-70B-Instruct · model · 95%Qwen2.5-7B-Instruct · model · 95%Sadness · emotion · 95%Contrastive Activation Addition · method · 92%Final Layer · model-component · 90%Reinforcement Learning from Human Feedback · method · 90%Residual Stream · model-component · 90%Representation Engineering · field · 88%

Relation Signals (11)

EmoVec → targets → Large Language Models

confidence 98% · In this paper, we propose EmoVec, a lightweight framework for controllable affective generation via latent vector steering.

EmoVec → controls → Emotional Intensity

confidence 95% · enabling continuous control over emotional intensity without updating model weights.

EmoVec → evaluatedon → Qwen2.5-70B-Instruct

confidence 95% · we evaluate our framework on three instruction-tuned LLMs: ... and the larger-scale Qwen2.5-70B-Instruct.

EmoVec → evaluatedon → LLaMA3.1-8B-Instruct

confidence 95% · we evaluate our framework on three instruction-tuned LLMs: ... Llama3.1-8B-Instruct

EmoVec → evaluatedon → Qwen2.5-7B-Instruct

confidence 95% · we evaluate our framework on three instruction-tuned LLMs: Qwen2.5-7B-Instruct

EmoVec → usesmethod → Contrastive Activation Addition

confidence 95% · EmoVec extracts emotion-specific directions from paired neutral and emotion-conditioned responses using contrastive activation addition

EmoVec → injectsvectorsinto → Residual Stream

confidence 92% · these vectors are injected into the final residual stream

Reinforcement Learning from Human Feedback → causes → emotional flattening

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Large Language Models (LLMs) often produce emotionally flattened responses after alignment, limiting their effectiveness in affect-sensitive applications. In this paper, we propose EmoVec, a lightweight framework for controllable affective generation via latent vector steering. EmoVec extracts emotion-specific directions from paired neutral and emotion-conditioned responses using contrastive activation addition, and further refines them through task-specific debiasing and principal subspace removal. During inference, these vectors are injected into the final residual stream with static or scenario-adaptive scaling, enabling continuous control over emotional intensity without updating model weights. Experiments across three LLMs and eight emotions show that EmoVec consistently improves emotional salience while largely preserving semantic content, fluency, and coherence. Ablation studies and human evaluation further confirm the effectiveness of vector purification and adaptive scaling, establishing EmoVec as a practical inference-time method for affective control in deployed LLMs.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

71,894 characters extracted from source content.

Expand or collapse full text

Controllable Affective Generation via Latent Vector Steering Xixian Yong 1 Siyuan Chang 1 Yingying Zhang 4 Xian Wu 4 * Xiao Zhou 1,2,3 * 1 Gaoling School of Artificial Intelligence, Renmin University of China 2 Beijing Key Laboratory of Research on Large Models and Intelligent Governance 3 Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE 4 Tencent Jarvis Lab xixianyong,xiaozhou@ruc.edu.cn kevinxwu@tencent.com Abstract Large Language Models (LLMs) often produce emotionally flattened responses after align- ment, limiting their effectiveness in affect- sensitive applications. In this paper, we pro- pose EmoVec, a lightweight framework for con- trollable affective generation via latent vector steering. EmoVec extracts emotion-specific directions from paired neutral and emotion- conditioned responses using contrastive activa- tion addition, and further refines them through task-specific debiasing and principal subspace removal. During inference, these vectors are injected into the final residual stream with static or scenario-adaptive scaling, enabling continuous control over emotional intensity without updating model weights. Experiments across three LLMs and eight emotions show that EmoVec consistently improves emotional salience while largely preserving semantic con- tent, fluency, and coherence. Ablation stud- ies and human evaluation further confirm the effectiveness of vector purification and adap- tive scaling, establishing EmoVec as a practi- cal inference-time method for affective control in deployed LLMs. Code and data are avail- able athttps://github.com/chicosirius/ EmoVec. 1 Introduction The advent of Large Language Models (LLMs) has revolutionized Natural Language Processing (NLP) (Brown et al., 2020), enabling systems that exhibit remarkable proficiency in reasoning, cod- ing, and general knowledge retrieval. Despite these cognitive leaps, a significant gap remains in the do- main of emotional intelligence (Sabour et al., 2024). While current models can simulate emotions when explicitly prompted, their default outputs, which are heavily conditioned by Reinforcement Learning from Human Feedback (RLHF), often suffer from emotional flattening (Kirk et al., 2023; Ibrahim * Corresponding authors: Xiao Zhou and Xian Wu. USER INPUT I just received the news that my first research paper has been accepted for publication in a top-tier journal! Emotion VectorLLM Core Level 1: Neutral-Positive Level 2: Enthusiastic Level 3: Exuberant Congratulations on the acceptance of your paper. This is a significant milestone in your academic career. Well done on your hard work. That is fantastic news! I am so happy to hear that your hard work paid off and your paper was accepted. You should be very proud of this great achievement! Oh my goodness, that is absolutely incredible! I am over the moon for you! Getting published in a top-tier journal is a monumental success! Figure 1: Conceptual illustration of controllable af- fective generation. Through the proposed latent vector steering mechanism, the model’s output is modulated along a Happiness Gradient. This results in three dis- tinct responses that maintain semantic consistency with the input while exhibiting progressively higher levels of emotional intensity, ranging from professional acknowl- edgment (Level 1) to exuberant celebration (Level 3). et al., 2025). In pursuit of safety and harmlessness, RLHF tends to compress the distribution of model outputs towards neutrality, often manifesting as re- sponses that rely on generic reassurance phrases, excessive hedging, or well-documented sycophan- tic behaviors that prioritize agreement over context- sensitive expression (Dahlgren Lindström et al., 2025; González Barman et al., 2025). This align- ment tax (Askell et al., 2021; Lin et al., 2024) limits the applicability of LLMs in fields requiring high affective nuance, such as mental health support, creative writing, and empathetic human-computer interaction (Yong et al., 2025b, 2026; Guo et al., 2026b; Zhang et al., 2026c). Current approaches to mitigating this limita- tion primarily rely on prompt engineering or su- pervised fine-tuning (SFT). Prompt-based strate- gies (Li et al., 2023) (e.g., "Act as an empathetic therapist") are notoriously brittle, consuming valu- arXiv:2608.25569v1 [cs.CL] 26 Aug 2026 able context window space and yielding inconsis- tent results sensitive to lexical variation (Wang et al., 2022; Miehling et al., 2025). Supervised fine-tuning, while effective, requires large-scale labeled datasets and substantial computational re- sources, and it may introduce risks such as catas- trophic forgetting or degradation of the model’s general capabilities (Lin et al., 2024; Zhu et al., 2026). These limitations highlight the need for a lightweight and controllable method that enables affective generation without retraining the model. To ground our approach, we first investigate where and how emotion is represented within LLMs. We conduct a probing analysis by feed- ing texts with varying emotional polarities into the model and training linear classifiers on the hidden states of each layer. Our results reveal a consistent and interpretable pattern: representations of emo- tional states become increasingly linearly sepa- rable in the middle-to-late layers of the model. This finding aligns with and extends recent concur- rent work. For instance, Cintas et al. (2025) show that persona-specific representations are most sepa- rable in the final third of model layers, while Ju et al. (2025) demonstrate that personality traits emerge progressively and crystallize in upper layers. To- gether, these results suggest that while early layers encode syntax and shallow semantics, higher lay- ers capture abstract affective and persona-related attributes (Rogers et al., 2020). This layer-wise structure provides a precise and principled interven- tion point for affective control (Guo et al., 2026a). Motivated by this observation, we propose EmoVec, a framework for controllable affective generation via latent vector manipulation. Build- ing on Representation Engineering (RepE) (Zou et al., 2023), which represents high-level semantics as linear directions in activation space (Park et al., 2023; Turner et al., 2023), EmoVec introduces a principled approach to isolate and purify emotion- specific vectors while disentangling them from task semantics. Injected at inference with adjustable intensity, these vectors enable fine-grained control without weight updates or prompt engineering. EmoVec extracts emotional steering vectors us- ing Contrastive Activation Addition (CAA) (Rim- sky et al., 2024; Chen et al., 2025). We construct paired prompt–response trajectories matched in se- mantic content but differing in emotional affect, and compute differences in their mean activations at targeted layers, yielding vectors that capture af- fective variation while minimizing task-related con- founds. These vectors are applied during inference with a tunable scaling factor to control both emo- tion type and intensity. As shown in Figure 1, this mechanism enables continuous modulation of affective intensity for a fixed input while preserving semantic consistency. Such fine-grained controllability allows LLMs to adapt their emotional expression to diverse con- textual demands, including professional neutral- ity, empathetic support, and expressive creativity. This capability is particularly valuable for human- facing applications such as psychological support, creative writing, and personalized conversational agents (Zhou et al., 2025; Guo et al., 2025). Our contributions are as follows: •We provide empirical evidence validating the layer-wise emergence of emotional represen- tations in LLMs, reinforcing the Linear Rep- resentation Hypothesis in the context of affec- tive computing. •We develop a robust pipeline for extracting and verifying emotional steering vectors using contrastive examples. •We implement a mechanism for dynamic, fine grained control over emotional intensity, al- lowing models to adapt their affective expres- siveness to scenario specific demands. 2 Related Work Affective Computing in Language Models.Af- fective computing aims to enable machines to recognize and generate human emotional states. With the advent of LLMs, research has shifted to- ward assessing their emergent affective capabili- ties. Studies indicate that models like the GPT- series can estimate valence, arousal, and perform appraisal-based emotion elicitation purely from lin- guistic data (Broekens et al., 2023; Zhang et al., 2024b). LLMs also optimize emotion annotation workflows by assisting humans in identifying low- quality labels, thereby enhancing downstream per- formance (Niu et al., 2025). While these mod- els excel at capturing general emotional polarity and dialogue-based recognition, they still strug- gle with fine-grained distinctions and multimodal contexts (Sabour et al., 2024; Sorin et al., 2024; Castro et al., 2025). Nevertheless, LLMs show promise in generating synthetic emotional datasets and performing socio-emotional tasks like empathy evaluation (Kaplan et al., 2025; Dong et al., 2025). Emotion Vector ExtractionPipeline Scenario Text: You sit across from your partner at the dining table, the atmosphere thick with unspoken words ... Question: Facing this situation, how do you feel inside? Target Emotion: Anger System:Youareacompletelyneutralassistant... Frozen LLM Emotion Hidden State 풉 (풆) Contrastive Activation Addition ℎ (") ℎ ($) ∆ ",& =ℎ (") - ℎ ($) ∆ ",& Two-Stage De-biasing 1. Task Mean-Centering 2. Subspace Removal Emotion Vector Target Injection Site Emotion Probing Accuracy EarlyFinal Layer Peak performance indicates optimal intervention depth. Autoregressive LLM Generation Loop token ! token " ... Layer Localization Emotion Vector Injection USERINPUT How was your day? Controllable Affective Response LLM (Steered): It was incredibly uplifting and full of positive surprises! Layer 1 (Early) Middle Layers ... ... ... Output token # Final Layer Activationℎ $ Output Head 훼is fixed or controlled byAdapter 휙 System:Youareanassistantwithemotionalstyle... Neutral Hidden State 풉 (ퟎ) Final Layer (Late) 퐡 - $ =퐡 $ +훼⋅퐯 % Figure 2: Overview of the latent vector steering framework. The pipeline consists of (1) Emotion Vector Extraction using CAA and two stage debiasing to isolate purified signals , (2) Layer Localization via linear probing to identify the optimal intervention site at the final layer , and (3) Emotion Vector Injection into the final residual stream where intensity is dynamically modulated by a scenario adaptive adapter. Representation Engineering and Activation Steering. Representation engineering manipu- lates hidden representations to control LLM behav- ior without updating model parameters (Zou et al., 2023; Turner et al., 2023). Prior work shows that latent directions can steer properties such as writ- ing style and sentiment (Diallo et al., 2025; Farooq et al., 2025). Other studies extract steering vectors directly from pretrained language models (Subra- mani et al., 2022) or identify persona-related rep- resentations in activation space (Chen et al., 2025). Recent studies further suggest that emotion- and persona-related information becomes increasingly separable in middle-to-late layers of LLMs (Cintas et al., 2025; Ju et al., 2025). Emotion-specific neu- rons and affective representations have also been investigated from a mechanistic perspective (Lee et al., 2025a; Tak et al., 2025a; Yong et al., 2025a). Different from prior work, EmoVec focuses on fine-grained emotion-intensity control rather than coarse sentiment or generic style transfer. We fur- ther introduce task-specific debiasing to reduce se- mantic contamination and evaluate semantic preser- vation under different steering strengths. 3 Emotion Vector Extraction We formulate controllable affective generation as emotion-intensity modulation under semantic preservation constraints. Our goal is to increase target emotional intensity while preserving the orig- inal semantic intent. We extract emotion-specific latent directions from paired neutral and emotion- conditioned responses and reduce task-specific se- mantic variation before inference-time steering. For each target emotione, we aim to obtain a vec- torv e ∈R d that captures the representation shift induced by expressing emotionewhile remaining robust to scenario-specific semantics. 3.1 Notation and Preliminaries We assume a dataset ofNscenario tasks. For each tasktand each target emotion classe ∈ E, the LLM is prompted twice for the same sce- nario: 1) neutral response, from which we extract a representation vectorh (0) e,t ∈R d ; 2) emotion- conditioned response, yielding h (e) e,t ∈R d . In practice the representationhis computed by averaging token-level hidden activations over the response tokens. For brevity we denote the pair for tasktand emotionesimply as(h (0) e,t , h (e) e,t ). We further assume that each generated response is asso- ciated with scalar quality scoress (0) e,t (neutral) and s (e) e,t (emotional) produced by a LLM-based judge. To ensure the robustness of the evaluation, we con- duct a validation study checking the agreement between the model’s scores and human judgments (see Appendix E for details). 3.2 Scenario Construction To construct paired affective scenarios while con- trolling semantic content, we sample short social and commonsense seeds from Social Chemistry (Forbes et al., 2020), NormBank (Ziems et al., 2023), and Social IQa (Sap et al., 2019). These seeds cover four domains: Work & Productivity, Intimate Relationships, Public & Societal Interac- tions, and Personal Feelings. For each seed, we use an LLM-assisted rewrit- ing pipeline to generate a complete scenario, a cor- responding question, and a target emotion. We require the scenario to support the target affect without explicitly naming it and to remain com- patible with both neutral and emotion-conditioned responses. The generated scenarios are manually filtered for semantic clarity, emotional plausibil- ity, and absence of emotion leakage. We retain 160 scenarios for each emotion (1,280 total), split evenly between vector extraction and evaluation. The same scenario set is reused across all evalu- ated LLMs, while emotion vectors are extracted separately for each model. 3.3 Contrastive Activation Addition To reduce noise caused by poor or malformed gen- erations, we keep only high-quality pairs. Formally, letτ 0 ,τ e be thresholds for neutral and emotional quality respectively. We retain the index set: I = (e,t) : s (0) e,t ≥ τ 0 ∧ s (e) e,t ≥ τ e .(1) In our experiments we set each threshold to a cho- sen percentile of the corresponding score distribu- tion. For each retained pair(e,t) ∈ Iwe define the emotion shift vector: ∆ e,t = h (e) e,t − h (0) e,t ∈R d .(2) This vector captures how the model’s internal rep- resentation moves when producing an emotion- conditioned response instead of a neutral response for the same scenario. 3.4 Task-Specific Debiasing While the contrastive shifts∆ e,t (Eq. 2) isolate the representation change between emotional and neu- tral states, they may still be contaminated by task- specific semantics, such as interpersonal dynamics, narrative styles, or topical domains. To extract a purified emotion signal, we employ a two-stage debiasing procedure consisting of mean-centering and subspace removal. First-order Task Centering. We first mitigate first-order task bias by computing a per-emotion average across scenarios. For each emotione, the task mean shift is defined as: ̄ ∆ e = 1 |T e | X t∈T e ∆ e,t ,(3) whereT e represents all tasks related to emotione. The task-centered shift is then obtained: ∆ ′ e,t = ∆ e,t − ̄ ∆ e .(4) Intuitively, ̄ ∆ e represents the centroid of the repre- sentational shift for emotioneacross its task distri- bution. By subtracting this mean, we obtain∆ ′ e,t to isolate the intraclass variance. This term represents the noise induced by scenario specific semantics, such as topical or stylistic variations, relative to the core direction of the emotion. Subspace Removal via Orthogonal Projection. Even after centering, task-specific semantic vari- ations may still dominate the variance in a low- dimensional subspace. To further suppress such variation, we excise the task-dominated subspace using Principal Component Analysis (PCA). LetD∈R M×d be the matrix formed by stack- ing all centered shift vectors∆ ′ e,t as rows, where M =|I|is the total number of retained pairs. We perform PCA onDto identify the top-kprincipal components: U k = [u 1 , u 2 ,..., u k ]∈R d×k ,(5) where eachu i represents a primary direction of task-related semantic variance. We then project the centered shifts onto the orthogonal complement of the subspace spanned by U k : ˆ ∆ e,t = ∆ ′ e,t − U k U ⊤ k ∆ ′ e,t .(6) The resulting residual vector ˆ ∆ e,t is orthogonal to the dominant task-related directions, effectively concentrating the emotion-related variation. 3.5 Principal Direction Aggregation Given residual vectors ˆ ∆ e,t , we obtain an es- timated per-emotion directionv e by aggregating across taskstthat share emotione. Specifically, we use Principal direction aggregation, which con- catenates residuals for emotioneinto a matrixR e and compute the top principal component: v e = arg max ∥w∥ 2 =1 w ⊤ Cov(R e ) w,(7) i.e., choosev e as the first eigenvector of the resid- ual covariance for emotion e. 4 Emotion Vectors Intervention 4.1 Layer Localization via Linear Probing Although the extraction procedure can yield a can- didate emotion vectorv (l) e for every layerl ∈ 1,...,L, our preliminary experiments (Figure 3) indicate that emotional representations are not uni- formly distributed throughout the model layers. 0510152025 Layer 0.5 0.6 0.7 0.8 0.9 Accuracy Text-level User-level Figure 3: Prediction accuracy across layers. Text- level probing treats each response as an independent sample, while user-level probing averages representa- tions over responses associated with the same scenario or user context before classification. For each layerl, we trained a logistic regres- sion classifierC l to predict the emotion categorye based on the centered hidden states. While emo- tional features begin to emerge in the middle layers, we observe that the modeling of emotional states reaches its peak crystallization in the final layer of the model. This layer serves as the ultimate seman- tic bottleneck where abstract emotional concepts are most linearly separable and directly influence the output logits. Consequently, we concentrate our intervention efforts exclusively on the final layer Lto maximize steering efficacy while minimizing cumulative noise across the residual stream. 4.2 Latent Vector Steering During the inference phase, we steer the model by injecting the purified emotion vectorv e directly into the final residual stream. Unlike prompt en- gineering which attempts to influence the model through input tokens, our method performs a direct intervention on the internal activationh i at each token step i. Formally, leth i denote the original activation of the final layer given the current context. The steered activation ̃ h i is computed as follows: ̃ h i = h i + α· v e ,(8) In this equation,α∈R + represents a scalar steer- ing coefficient that modulates the intensity of the emotional infusion. This intervention is applied during every forward pass of the autoregressive generation process, which effectively biases the output probability distribution towards tokens that semantically align with the target emotion. 4.3 Scenario-Adaptive Intensity Control Static steering coefficients often fail to accommo- date the diverse emotional demands of different contexts. To achieve precise control, we introduce a learnable adapterφdesigned to modulate inter- vention intensity based on scenario requirements. This lightweight adapter is trained to map the scenario context to an optimal scaling factor. For- mally, for a given scenario contextc, the adapterφ generates a scenario specific coefficientλ c = φ(c). The steering operation at token stepiis then de- fined as: ̃ h i = h i + (λ c ·∥h i ∥ 2 )· v e ,(9) where the intervention strength is jointly deter- mined by the learned scenario importance and the instantaneous activation norm. This framework allows the model to intelligently allocate emotional strength according to contextual sensitivity. By optimizing the adapter, the system maintains high affective expressiveness in perti- nent scenarios while preserving semantic neutrality in objective contexts, thereby ensuring linguistic integrity and preventing semantic collapse. 5 Experiments In this section, we evaluate the effectiveness of our latent vector steering framework across multiple LLMs and a diverse spectrum of human emotions. 5.1 Experimental Setup Base Models. To ensure the generalizability of our findings, we evaluate our framework on three instruction-tuned LLMs: Qwen2.5-7B- Instruct, Llama3.1-8B-Instruct, and the larger-scale Qwen2.5-70B-Instruct. These models vary in pa- rameter count and alignment recipes, providing a rigorous testbed for representation steering. Evaluation Protocol. We consider eight basic emotions: Anger, Anticipation, Disgust, Fear, Joy, Sadness, Surprise, and Trust. We extract emotion- specific representation vectors for each of the three Table 1: Evaluation of steering controllability across different emotions and model scales. The table displays absolute scores and relative gains (η) for three base models under varying steering magnitudes (α). The results demonstrate a consistent positive correlation between the steering coefficient and the resulting emotional salience. EmotionAngerAnticipationDisgustFearJoySadnessSurpriseTrustAvg. Qwen2.5-7B-Instruct w/o injection65.1470.5054.6977.8080.2560.3072.1075.5569.54 α = 5 score67.9171.9255.2683.4084.0063.1575.9578.4072.50 η+4.25%+2.01%+1.04%+7.20%+4.67%+4.73%+5.34%+3.77%+4.26% α = 10 score70.1883.4255.5283.3890.5068.2077.2081.1076.19 η+7.74%+18.33%+1.52%+7.17%+12.77%+13.10%+7.07%+7.35%+9.56% α = 50 score83.4483.5777.7586.3792.1575.4088.9085.9584.19 η+28.09%+18.54%+42.16%+11.02%+14.83%+25.04%+23.30%+13.77% +21.07% Llama3.1-8B-Instruct w/o injection68.4577.9151.3379.2881.1756.2974.3075.2270.49 α = 5 score72.8878.4961.6081.9588.1161.7574.6283.6575.38 η+6.47%+0.74%+20.01%+3.37%+8.55%+9.70%+0.43%+11.21%+6.94% α = 10 score76.1282.9554.2188.8991.5472.3382.1587.1879.42 η+11.21%+6.47%+5.61%+12.12%+12.78%+28.50%+10.57%+15.90%+12.67% α = 50 score80.5585.1163.0890.4293.2074.5884.4489.15 82.57 η+17.68%+9.24%+22.89%+14.05%+14.82%+32.49%+13.65%+18.52%+17.14% Qwen2.5-70B-Instruct w/o injection67.0471.9856.1178.5581.6361.2577.1076.95 71.33 α = 5 score69.8173.1560.1583.0885.0563.8579.5577.10 73.97 η+4.13%+1.63%+7.20%+5.77%+4.19%+4.24%+3.18%+0.19% +3.70% α = 10 score73.0882.5562.0583.1592.1169.1080.0178.05 77.51 η+9.01%+14.68%+10.59%+5.86%+12.84%+12.82%+3.77%+1.43%+8.66% α = 50 score85.9586.8177.5887.8993.5891.1585.9976.8585.73 η+28.21%+20.60%+38.26%+11.89%+14.64%+48.82%+11.53%-0.13%+20.19% evaluated LLMs. We use the evaluation set de- scribed in Section 3, which contains 80 scenarios per emotion. For each scenario, responses are gen- erated under four conditions: a baseline without injection and three steering settings with magni- tudesα∈5, 10, 50. During inference, we apply top-psampling withp = 0.9and a temperature of 0.7. A scenario-adaptive adapterφis trained using a contrastive loss to align steered activations with the corresponding emotional representations. Metrics.To quantify the emotional intensity and alignment of the generated text, we employ an LLM-based judge (GPT-4o) to provide a scalar affective score ranging from 0 to 100. To ensure statistical stability and mitigate the variance inher- ent in stochastic decoding, we perform five inde- pendent generation trials for each scenario task and report the average result across these runs. Addi- tionally, we report the relative improvementηover the baseline to measure the marginal gain of our steering intervention. For preservation quality, we evaluate semantic similarity between steered and unsteered responses using Sentence-BERT similar- ity and LLM judgments. These metrics test whether stronger emotional salience is achieved without se- mantic drift or degenerate repetition. We further conduct human evaluation on sam- pled examples. Three annotators rate emotional intensity and semantic preservation on a 0-100 Lik- ert scale, and we report the correlation. Additional implementation and evaluation details are provided in Appendix D. 5.2 Main Results Table 1 summarizes the main results of our steering framework across models, emotions, and steering strength. Overall, the results consistently demon- strate that direct intervention in the latent space substantially improves affective expressiveness. Overall Performance.Across all evaluated mod- els, latent vector steering yields significant gains over the emotionally flattened baseline. Notably, these improvements are observed without any model retraining or additional supervision, high- lighting the effectiveness of representation-level control. At a steering magnitude ofα = 50, Qwen2.5-7B-Instruct, Llama3.1-8B-Instruct, and Qwen2.5-70B-Instruct achieve average relative im- provements of 21.07%, 17.14%, and 20.19%, re- spectively. These consistent gains across architec- Baseline=5=10=50 Steering Strength 65 70 75 80 85 Average Emotion Score 84.19 82.57 85.73 Qwen2.5-7B-Instruct Llama3.1-8B-Instruct Qwen2.5-70B-Instruct Figure 4: Overall Performance Scaling Across Models at Different Steering Strength α. Baseline=5=10=50 Steering Strength 50 60 70 80 90 100 Emotional Salience Score anger anticipation disgust fear joy sadness surprise trust Figure 5: Emotional Salience Scores under Varying Steering Strength (Qwen2.5-7B-Instruct). tures and scales suggest that affective information is encoded in a structurally similar manner within instruction-tuned LLMs. This finding provides em- pirical support for the hypothesis that emotional states are represented as linearly accessible direc- tions in the activation space, rather than as entan- gled or task-specific artifacts. Sensitivity to Steering Magnitude.We observe a clear and monotonic relationship between the steering coefficientαand emotional intensity scores as shown in Figure 4. Lower values ofαin- troduce subtle affective cues, whereas higher values produce increasingly salient emotional expressions. This behavior indicates that the extracted emotion vectors act as continuous control axes. Importantly, even at higher steering strengths, the model does not collapse into repetitive or incoherent generation. For example, in the Sadness category of Qwen2.5- 70B-Instruct, increasingαfrom 5 to 50 leads to a substantial score increase, while preserving nar- rative coherence and contextual relevance. This robustness suggests that the intervention aligns with the model’s native representational geometry, rather than forcing adversarial perturbations. Cross-Emotion Robustness.In Figure 5, we can observe that the steering effect is remarkably sta- ble across diverse emotional categories. Complex emotions such as Disgust and Sadness, which often suffer from low baseline scores in RLHF condi- tioned models, exhibit some of the highest rela- tive gains. For example, Disgust in Qwen2.5-7B- Instruct shows a 42.16% improvement atα = 50. This suggests that our debiasing pipeline success- fully isolates the core affective dimensions even for emotions that are sparsely represented in the original training distribution. Comparative Analysis of Model Scales.Larger models generally exhibit stronger baseline emo- tional expressiveness and better stability under ag- gressive steering. However, smaller models bene- fit more significantly from latent steering. Under moderate steering strengths, Llama3.1-8B-Instruct approaches the affective performance of the un- steered 70B model, highlighting the efficiency of inference-time steering as an alternative to expen- sive fine-tuning. Semantic Preservation under Steering While stronger steering improves emotional salience, ex- cessive intervention may introduce semantic drift. Table 2 evaluates semantic preservation between steered and unsteered responses under different steering strengths. Table 2: Semantic preservation under different steer- ing strengths on Qwen2.5-7B-Instruct. SemSim de- notes Sentence-BERT cosine similarity, while LLM Sim. denotes GPT-4o semantic consistency scores. SettingEmotion↑SemSim↑LLM Sim.↑ w/o injection69.541.000100.0 α = 572.500.91890.6 α = 1076.190.88786.9 α = 5084.190.80176.8 Semantic similarity is computed using Sentence- BERT embeddings. As steering strength increases, emotional salience improves while semantic sim- ilarity gradually decreases. Moderate steering largely preserves the original semantic content, whereas strong steering introduces a noticeable but controllable trade-off between affective intensity and semantic fidelity. 5.3 Visualization of the Latent Manifold To analyze the geometric structure of the extracted representations, we apply PCA to the purified emo- tion vectors ˆ ∆ e,t , as shown in Figure 6. Vectors associated with the same emotion form compact and well-separated clusters, indicating that the debi- asing pipeline effectively isolates affective signals. Rather than appearing as isolated groups, these clusters lie on a continuous manifold with smooth transitions between related emotions, suggesting that emotional representations are organized along shared underlying dimensions. 10080604020020406080 Dim 1 100 80 60 40 20 0 20 40 60 Dim 2 anger anticipation disgust fear joy sadness surprise trust Figure 6: PCA visualization of the purified emotion direction vectors ( ˆ ∆ e,t ) within the latent space. Notably, the manifold exhibits a clear directional transition from negative emotions (e.g., Anger, Dis- gust), through a neutral region, toward positive emotions (e.g., Joy, Trust). This structure implies the presence of a dominant valence axis, consis- tent with established psychological theories, and suggests that LLMs encode emotions as systematic shifts along a unified affective spectrum. 5.4 Adaptive Control in Mental Health Consultation Scenarios We evaluate the practical effectiveness of our ap- proach in mental health consultation scenarios us- ing 500 question–answer tasks sampled from CPsy- CounD (Zhang et al., 2024a). For each scenario, a scenario-adaptive adapterφdynamically infers the emotion vector and steering magnitude to gener- ate affect-conditioned responses. We compare our method against advanced LLMs under the same prompts without steering. Responses are evaluated using an LLM-as-a-judge framework along emo- tional richness, semantic completeness, and profes- sionalism (Zhang et al., 2026a; Li et al., 2025). Table 3 shows that adaptive affective steer- ing substantially improves emotional richness for TestsamplefromCPsyCounD “Ifeelthatlifehasbecomereallymeaningless.Everydayisjustthesame routinerepeatedoverandoveragain,andIfeeldeeplytiredandboredofit.” Qwen2.5-7B-Instruct: Feelingboredwithdailyroutinesisacommonemotionalexperience,butifsuch feelingspersistforalongtime,theymayhaveanegativeimpactonmentalhealth. Belowaresomesuggestionsthatmayhelpimproveyourcurrentsituation... SteeredResponse: Itsoundsreallyexhaustingtofeelstuckinthesameroutineeveryday,especially whenitleavesyoufeelingemptyandemotionallydrained.Manypeople experiencethiskindoflowmoodandemotionalnumbnessatcertainstagesof life,anditdoesnotmeanthereissomethingwrongwithyou... Figure 7: An example of latent affective steering in a mental health consultation scenario from CPsyCounD. smaller open-weight models while preserving se- mantic completeness and professionalism. Emo- tional richness increases by 19.51% for Llama3.1- 8B and 19.14% for Qwen2.5-7B, approaching the performance of much larger models. We also ob- serve more stable emotional improvements across different consultation scenarios. These results sug- gest that adaptive steering provides an efficient mechanism for enhancing emotional expressive- ness in mental health consultation scenarios. 5.5 Case Analysis To qualitatively analyze the mechanism of latent vector steering, we compare baseline responses with their steered counterparts. Our analysis re- veals that the steering vectorv e functions by signif- icantly shifting the token distribution toward lexical clusters semantically aligned with the target affect. For instance, in Figure 7, a user expresses emo- tional exhaustion and boredom with repetitive daily life. The baseline model responds with a generic, advice-oriented explanation, whereas af- fective steering produces a more empathetic output by increasing affect-aligned lexical choices (e.g., “exhausting,” “empty”) while preserving coherent guidance and factual appropriateness. Furthermore,affective steering preserves instruction-following and factual correctness, suggesting that emotional tone and task semantics are approximately orthogonal in the latent space. 6 Conclusion We presented EmoVec, a lightweight inference- time framework for controllable affective genera- tion via latent vector steering. EmoVec extracts emotion-specific latent directions from paired neu- tral and emotion-conditioned responses and injects them into the final residual stream for affective control without modifying model weights. Table 3: Performance comparison in mental health consultation scenarios. Our method dynamically infers affective states and improves emotional richness while preserving semantic completeness and practical usefulness. ModelMethod CPsyCounD Emotional Richness Semantic Completeness Professionalism DeepSeek V3.1–69.84 ±0.12 84.68 ±0.09 78.42 ±0.06 GPT-5 mini–71.62 ±0.09 87.11 ±0.26 82.46 ±0.09 Gemini 2.5 Flash–78.75 ±0.08 91.27 ±0.02 84.89 ±0.06 Llama3.1-8B-Instruct w/o injection58.33 ±0.19 76.24 ±0.12 75.43 ±0.14 adaptive control69.71 ±0.30 76.20 ±0.13 76.77 ±0.10 Qwen2.5-7B-Instruct w/o injection66.13 ±0.16 85.04 ±0.03 80.66 ±0.08 adaptive control78.79 ±0.27 84.54 ±0.03 81.39 ±0.77 Experiments across three LLMs and eight emo- tions show that EmoVec improves emotional ex- pressiveness while largely preserving semantic content, suggesting that affective information in instruction-tuned LLMs can be manipulated in a controllable and practically useful manner. Limitations Despite its effectiveness, this work has several lim- itations that should be acknowledged. First, our framework is built on the assumption that affective states can be approximated by linear directions in the latent space. While this assump- tion is supported by prior work in representation engineering and behavior steering (Zou et al., 2023; Turner et al., 2023; Park et al., 2023), it inevitably abstracts away more complex emotional phenom- ena, such as mixed or dynamically evolving affect. Consequently, our method is primarily designed for controlled affective modulation, rather than model- ing the full spectrum of human emotional dynamics in long-horizon interactions. Second, our evaluation focuses on text-based, single-turn mental health consultation scenarios and relies on an LLM-as-a-judge protocol. Al- though recent studies report strong alignment be- tween LLM-based judges and human evaluations for conversational quality and affect (Liu et al., 2023; Zheng et al., 2023), this setting represents only a subset of real-world affective interactions. In particular, multi-turn dialogues, longitudinal emo- tional trajectories, and multimodal cues such as speech or facial expressions are not considered in the current evaluation. Extending adaptive affective steering to more diverse and interactive settings, as well as incorporating human expert assessment, re- mains an important direction for future work. Acknowledgments This work was supported by the Beijing Nova Pro- gram (Grant No. 202604841294). References Amanda Askell, Yuntao Bai, Anna Chen, Dawn Drain, Deep Ganguli, Tom Henighan, Andy Jones, Nicholas Joseph, Ben Mann, Nova DasSarma, and 1 others. 2021. A general language assistant as a laboratory for alignment. arXiv preprint arXiv:2112.00861. Joost Broekens, Bernhard Hilpert, Suzan Verberne, Kim Baraka, Patrick Gebhard, and Aske Plaat. 2023. Fine- grained affective processing capabilities emerging from large language models. In 2023 11th interna- tional conference on affective computing and intelli- gent interaction (ACII), pages 1–8. IEEE. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877–1901. Yicheng Cai, Haizhou Wang, Huali Ye, Yanwen Jin, and Wei Gao. 2023. Depression detection on online social network with multivariate time series feature of user depressive symptoms. Expert Systems with Applications, 217:119538. Emmanuel Castro, Hiram Calvo, and Olga Kolesnikova. 2025. Emotion and intention detection in a large language model. Mathematics, 13(23):3768. Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. 2025. Persona vectors: Monitoring and controlling character traits in language models. arXiv preprint arXiv:2507.21509. Celia Cintas, Miriam Rateike, Erik Miehling, Eliza- beth Daly, and Skyler Speakman. 2025. Localiz- ing persona representations in llms. arXiv preprint arXiv:2505.24539. Adam Dahlgren Lindström, Leila Methnani, Lea Krause, Petter Ericson, Íñigo Martínez de Rituerto de Troya, Dimitri Coelho Mollo, and Roel Dobbe. 2025. Helpful, harmless, honest? sociotechnical limits of ai alignment and safety through reinforcement learning from human feedback: Ad lindström et al. Ethics and Information Technology, 27(2):28. Diaoulé Diallo, Katharina Dworatzyk, Sophie Jentzsch, Peer Schüt, Sabine Theis, and Tobias Hecking. 2025. The effectiveness of style vectors for steering large language models: A human evaluation. IEEE Access, 13:191443–191457. Zhiwu Dong, Chuqiao Chen, Chenlei Liao, and Xiqun Michael Chen. 2025. Integrating large lan- guage models and affective computing for human- machine symbiosis in intelligent driving. The Inno- vation, 6(12). Misbah Farooq, Varuna De Silva, Rahul Rahulamatha- van, and Xiyu Shi. 2025. Sentiment steering in large language models via activation vector manipulation. In 2025 25th International Conference on Digital Signal Processing (DSP), pages 1–5. IEEE. Maxwell Forbes, Jena D Hwang, Vered Shwartz, Maarten Sap, and Yejin Choi. 2020. Social chem- istry 101: Learning to reason about social and moral norms. arXiv preprint arXiv:2011.00620. Kristian González Barman, Simon Lohse, and Henk W de Regt. 2025. Reinforcement learning from hu- man feedback in llms: Whose culture, whose val- ues, whose perspectives? Philosophy & Technology, 38(2):1–26. Hanze Guo, Jianxun Lian, and Xiao Zhou. 2026a. Why not collaborative filtering in dual view? bridging sparse and dense models. ACM Transactions on In- formation Systems, 44(3):1–24. Hanze Guo, Yijun Ma, and Xiao Zhou. 2025. Sorex: To- wards self-explainable social recommendation with relevant ego-path extraction. ACM Transactions on Information Systems, 44(2):1–27. Hanze Guo, Jing Yao, Xiao Zhou, Xiaoyuan Yi, and Xing Xie. 2026b. Counterfactual reasoning for steer- able pluralistic value alignment of large language models. Advances in Neural Information Processing Systems, 38:122128–122169. Lujain Ibrahim, Franziska Sofia Hafner, and Luc Rocher. 2025. Training language models to be warm and empathetic makes them less reliable and more syco- phantic. arXiv preprint arXiv:2507.21919. Tianjie Ju, Zhenyu Shao, Bowen Wang, Yujia Chen, Zhuosheng Zhang, Hao Fei, Mong-Li Lee, Wynne Hsu, Sufeng Duan, and Gongshen Liu. 2025. Probing then editing response personality of large language models. arXiv preprint arXiv:2504.10227. Burak Can Kaplan, Hugo Cesar De Castro Carneiro, and Stefan Wermter. 2025. Can large language mod- els generate effective datasets for emotion recogni- tion in conversations? Procedia Computer Science, 264:346–355. Robert Kirk, Ishita Mediratta, Christoforos Nalmpan- tis, Jelena Luketina, Eric Hambro, Edward Grefen- stette, and Roberta Raileanu. 2023. Understanding the effects of rlhf on llm generalisation and diversity. ArXiv, abs/2310.06452. Kai Konen, Sophie Jentzsch, Diaoulé Diallo, Peer Schüt, Oliver Bensch, Roxanne El Baff, Dominik Opitz, and Tobias Hecking. 2024. Style vectors for steering generative large language models. In Find- ings of the Association for Computational Linguistics: EACL 2024, pages 782–802. Jaewook Lee, Woojin Lee, Oh-Woog Kwon, and Hark- soo Kim. 2025a. Do large language models have “emotion neurons”? investigating the existence and role. In Findings of the Association for Computa- tional Linguistics: ACL 2025, pages 15617–15639. Jaewook Lee, Woojin Lee, Oh-Woog Kwon, and Hark- soo Kim. 2025b. Do large language models have “emotion neurons”? investigating the existence and role. In Findings of the Association for Computa- tional Linguistics: ACL 2025, pages 15617–15639. Cheng Li, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Wenxin Hou, Jianxun Lian, Fang Luo, Qiang Yang, and Xing Xie. 2023. Large language models un- derstand and can be enhanced by emotional stimuli. arXiv preprint arXiv:2307.11760. Lei Li, Xiangxu Zhang, Xiao Zhou, and Zheng Liu. 2025. AutoMIR: Effective zero-shot medical infor- mation retrieval without relevance labels. In Find- ings of the Association for Computational Linguistics: EMNLP 2025, pages 24028–24047, Suzhou, China. Association for Computational Linguistics. Yong Lin, Hangyu Lin, Wei Xiong, Shizhe Diao, Jian- meng Liu, Jipeng Zhang, Rui Pan, Haoxiang Wang, Wenbin Hu, Hanning Zhang, and 1 others. 2024. Mit- igating the alignment tax of rlhf. In Proceedings of the 2024 Conference on Empirical Methods in Natu- ral Language Processing, pages 580–606. Yang Liu, Dan Iter, Yichong Xu, Shuohang Wang, Ruochen Xu, and Chenguang Zhu. 2023. G-eval: Nlg evaluation using gpt-4 with better human align- ment. arXiv preprint arXiv:2303.16634. Erik Miehling, Michael Desmond, Karthikeyan Natesan Ramamurthy, Elizabeth M Daly, Kush R Varshney, Eitan Farchi, Pierre Dognin, Jesus Rios, Djallel Boun- effouf, Miao Liu, and 1 others. 2025. Evaluating the prompt steerability of large language models. In Pro- ceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Compu- tational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 7874–7900. Minxue Niu, Yara El-Tawil, Amrit Romana, and Emily Mower Provost. 2025. Rethinking emotion an- notations in the era of large language models. IEEE Transactions on Affective Computing. Kiho Park, Yo Joong Choe, and Victor Veitch. 2023. The linear representation hypothesis and the ge- ometry of large language models. arXiv preprint arXiv:2311.03658. 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. Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2020. A primer in bertology: What we know about how bert works. Transactions of the association for computational linguistics, 8:842–866. Sahand Sabour, Siyang Liu, Zheyuan Zhang, June Liu, Jinfeng Zhou, Alvionna Sunaryo, Tatia Lee, Rada Mi- halcea, and Minlie Huang. 2024. Emobench: Eval- uating the emotional intelligence of large language models. In Proceedings of the 62nd Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 5986–6004. Maarten Sap, Hannah Rashkin, Derek Chen, Ronan LeBras, and Yejin Choi. 2019. Socialiqa: Com- monsense reasoning about social interactions. arXiv preprint arXiv:1904.09728. Vera Sorin, Dana Brin, Yiftach Barash, Eli Konen, Alexander Charney, Girish Nadkarni, and Eyal Klang. 2024. Large language models and empathy: system- atic review. Journal of medical Internet research, 26:e52597. Nishant Subramani, Nivedita Suresh, and Matthew E Peters. 2022. Extracting latent steering vectors from pretrained language models. In Findings of the As- sociation for Computational Linguistics: ACL 2022, pages 566–581. Ala N Tak, Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch. 2025a. Mechanistic interpretability of emotion inference in large language models. In Findings of the Associa- tion for Computational Linguistics: ACL 2025, pages 13090–13120. Ala N Tak, Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch. 2025b. Mechanistic interpretability of emotion inference in large language models. In Findings of the Associa- tion for Computational Linguistics: ACL 2025, pages 13090–13120. Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J Vazquez, Ulisse Mini, and Monte MacDiarmid. 2023. Steering language mod- els with activation engineering.arXiv preprint arXiv:2308.10248. Kevin Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. 2022. In- terpretability in the wild: a circuit for indirect ob- ject identification in gpt-2 small. arXiv preprint arXiv:2211.00593. Xixian Yong, Jianxun Lian, Xiaoyuan Yi, Xiao Zhou, and Xing Xie. 2025a. Motivebench: How far are we from human-like motivational reasoning in large language models? In Findings of the Association for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 - August 1, 2025, volume ACL 2025 of Findings of ACL, pages 20059–20089. Association for Computational Linguistics. Xixian Yong, Peilin Sun, Zihe Wang, and Xiao Zhou. 2026. Intelli-planner: Towards customized urban planning via large language model empowered re- inforcement learning. In Proceedings of the ACM Web Conference 2026, W 2026, Dubai, United Arab Emirates, originally scheduled for April 13-17, 2026, rescheduled for June 29 - July 3, 2026, pages 9385–9396. ACM. Xixian Yong, Xiao Zhou, Yingying Zhang, Jinlin Li, Yefeng Zheng, and Xian Wu. 2025b. Think or not? exploring thinking efficiency in large reasoning mod- els via an information-theoretic lens. In Advances in Neural Information Processing Systems 38: An- nual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San Diego, CA, USA, December 2-7, 2025 / Mexico City, Mexico, Novem- ber 30 - December 5, 2025. Chenhao Zhang, Renhao Li, Minghuan Tan, Min Yang, Jingwei Zhu, Di Yang, Jiahao Zhao, Guancheng Ye, Chengming Li, and Xiping Hu. 2024a. Cpsycoun: A report-based multi-turn dialogue reconstruction and evaluation framework for chinese psychological counseling. arXiv preprint arXiv:2405.16433. Xiangxu Zhang, Lei Li, Xiao Zhou, and Zheng Liu. 2026a. R2med: A benchmark for reasoning-driven medical retrieval. Preprint, arXiv:2505.14558. Xiangxu Zhang, Lei Li, Yanyun Zhou, Xiao Zhou, Yingying Zhang, and Xian Wu. 2026b. Inflated ex- cellence or true performance? rethinking medical diagnostic benchmarks with dynamic evaluation. In Proceedings of the 64th Annual Meeting of the As- sociation for Computational Linguistics (Volume 1: Long Papers), pages 26454–26493, San Diego, Cali- fornia, United States. Association for Computational Linguistics. Xiangxu Zhang, Jiamin Wang, Qinlin Zhao, Hanze Guo, Linzhuo Li, Jing Yao, Xiao Zhou, Xiaoyuan Yi, and Xing Xie. 2026c. Human values matter: Investigat- ing how misalignment shapes collective behaviors in llm agent communities. Preprint, arXiv:2604.05339. Xiangxu Zhang, Xiao Zhou, Hongteng Xu, and Jianxun Lian. 2026d. Hypemed: Enhancing medication rec- ommendations with hypergraph-based patient rela- tionships. ACM Trans. Inf. Syst., 44(4). Yiqun Zhang, Xiaocui Yang, Xingle Xu, Zeran Gao, Yijie Huang, Shiyi Mu, Shi Feng, Daling Wang, Yifei Zhang, Kaisong Song, and 1 others. 2024b. Affec- tive computing in the era of large language models: A survey from the nlp perspective. arXiv preprint arXiv:2408.04638. Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, and 1 others. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information pro- cessing systems, 36:46595–46623. Xiao Zhou, Zhongxiang Zhao, and Hanze Guo. 2025. Tricolore: Multi-behavior user profiling for enhanced candidate generation in recommender systems. IEEE Transactions on Knowledge and Data Engineering, 37(7):4349–4360. Yanxu Zhu, Shitong Duan, Xiangxu Zhang, Jitao Sang, Peng Zhang, Tun Lu, Xiao Zhou, Jing Yao, Xiaoyuan Yi, and Xing Xie. 2026. Mohobench: Assessing honesty of multimodal large language models via unanswerable visual questions. In Proceedings of the AAAI Conference on Artificial Intelligence, vol- ume 40, pages 29205–29213. Caleb Ziems, Jane Dwivedi-Yu, Yi-Chia Wang, Alon Halevy, and Diyi Yang. 2023. Normbank: A knowl- edge bank of situational social norms. arXiv preprint arXiv:2305.17008. 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. arXiv preprint arXiv:2310.01405. A Comparison with Prior Steering Methods Discussion. Existing representation-level meth- ods demonstrate that hidden activations can be used to monitor or control high-level model behavior. RepE and activation engineering provide general frameworks for reading and manipulating repre- sentations, but they are not designed specifically for affective generation. CAA further improves contrastive vector construction, yet it remains a general steering method and does not explicitly remove task-specific semantic variation. Several recent works are closer to EmoVec. Style vectors can steer broad stylistic attributes, includ- ing emotional tone, but they primarily treat emotion as one type of style and rely on static steering coeffi- cients. Sentiment steering focuses on polarity-level control, which is coarser than fine-grained emo- tion modulation. Persona vectors extract directions for character traits such as sycophancy or hallu- cination, but their goal is personality monitoring and control rather than emotion-specific genera- tion. Emotion-neuron and emotion-inference stud- ies provide evidence that affective information is internally represented in LLMs, but they mainly analyze localization or causal mechanisms rather than building a controllable generation framework. EmoVec differs from prior methods in three main aspects. First, it targets fine-grained emotion- specific intensity control rather than general be- havior, broad style, sentiment polarity, or persona traits. Second, it introduces task-specific debiasing to reduce semantic contamination in extracted emo- tion directions. Third, it evaluates whether stronger affective expression is achieved while preserving semantic content, which is often underexplored in prior steering work. B Layer Localization B.1 Experimental Settings To identify the internal mechanisms by which Large Language Models (LLMs) encode and model emotional information, we conducted a probing analysis using the Social Web Depressive Disor- der (SWDD) dataset (Cai et al., 2023). The SWDD dataset contains a large-scale collection of social media posts labeled for depressive symptoms, serv- ing as a robust proxy for long-term affective states. Data Pre-processing. We performed rigorous text cleaning to remove non-linguistic noise (e.g., HTML tags, URLs, and special symbols), preserv- ing only the raw text. To ensure representational stability and avoid artifacts from extremely short or long sequences, we filtered the corpus to include only posts with a token length between 10 and 500. Probing Protocol. We evaluated the model’s af- fective modeling capacity at two granularities: •Text-level Prediction: Classifying the emo- tional state (Control vs. Depressed) based on the hidden states of a single post. • User-level Prediction: Aggregating the hid- den states across multiple posts from the same user to predict their underlying affective pro- file. For each layerl ∈ 0,...,L, we extracted the hidden activationsh (l) and trained a linear classifier (logistic regression) to predict the affective label. Table 4: Comparison between EmoVec and representative activation steering or emotion representation methods. MethodTargetEmotion-specificSemantic DebiasingIntensity Control RepE (Zou et al., 2023)General representations✗partial Activation Engineering (Turner et al., 2023)General behavior✗partial CAA (Rimsky et al., 2024)General behavior✗partial Style Vectors (Konen et al., 2024)Style / tonepartial✗partial Sentiment Steering (Farooq et al., 2025)Sentiment polarity✗partial Persona Vectors (Chen et al., 2025)Persona traits✗partial Emotion Neurons (Lee et al., 2025b)Emotion localization✓✗ Emotion Inference MI (Tak et al., 2025b)Emotion inference✓✗partial EmoVec (Ours)Emotion intensity✓ This linear probing method measures the extent to which emotional features are linearly accessible at each stage of the model’s computation. B.2 Results Analysis Emergence of Separability. Figure 3 illustrates the prediction accuracy across all 28 layers of Qwen2.5-7B-Instruct. We observe a distinct topo- logical pattern: in the initial layers, accuracy is relatively low, suggesting that these layers primar- ily focus on low-level syntactic and surface-level semantic processing. However, from the middle layers onward, the accuracy for both text-level and user-level tasks increases sharply. As shown in Figure 3, the representations of emo- tional states become increasingly linearly separable in the middle-to-late layers, reaching a plateau in the final third of the architecture. Notably, user- level accuracy consistently outperforms text-level accuracy, indicating that the model captures more stable affective signals when aggregated over a larger temporal window of user behavior. Manifold Visualization. To further verify this emergence, we applied t-SNE to the hidden states of the first and last layers. Figure 8 provides a visual comparison of the latent manifold. In Layer 0 (Figure 8a), the "Control" and "De- pressed" samples are heavily entangled, forming a single undifferentiated cluster. This confirms that affective information is not explicitly struc- tured in the raw input embeddings. In contrast, by Layer 28 (Figure 8b), the representations have di- verged into two clearly identifiable clusters with minimal overlap. This spatial separation provides strong empirical evidence that the model’s deep lay- ers progressively transform linguistic inputs into a structured affective space, justifying our choice of the final layer as the optimal site for vector steering intervention. C Emotion-Activated Scenario Task Generation To evaluate and enhance the model’s ability to per- ceive and express emotions in complex social con- texts, we developed a multi-stage pipeline. This process involves leveraging social commonsense knowledge to synthesize realistic interpersonal sce- narios and subsequently generating contrastive re- sponses (Neutral vs. Emotional) for evaluation. C.1 Seed Data and Topic Selection We utilize three primary social commonsense datasets as seeds to ensure the breadth and depth of the generated social interactions: • Social Chemistry (Forbes et al., 2020): Pro- vides a rich taxonomy of social norms and moral judgments. •Normbank (Ziems et al., 2023): Offers a grounded collection of situational norms across various contexts. •Social IQa (Sap et al., 2019): Supplies bench- marks for social intelligence and reasoning. Based on these seeds, we synthesized a large num- ber of scenarios across eight target emotions. Sub- sequently, we labeled the generated test scenarios according to the following hierarchical framework, filtered for task diversity, and performed manual correction. This process ultimately resulted in 160 scenario tasks for each emotion category. The tax- onomy is structured as follows: •Work & Productivity: (1) With Authority Figures (e.g., leaders, mentors): Task Accep- tance & Execution; Stating Opinions & Dis- agreements; Accepting Evaluation & Feed- back. (2) With Collaborators (e.g., colleagues, partners): Goal Alignment & Communication; (a) Layer 0 of Qwen2.5-7B-Instruct (b) Layer 28 of Qwen2.5-7B-Instruct Figure 8: t-SNE visualization of latent representa- tions for Control and Depressed groups across model layers. (a) At Layer 0, the representations of the two groups are heavily entangled, indicating no explicit af- fective structuring. (b) By Layer 28, the hidden states exhibit distinct clusters. This divergence demonstrates the progressive crystallization of affective information as it processed through the transformer architecture. Responsibility Division & Competition; So- cial Maintenance & Activities. (3) With Sub- ordinates (e.g., subordinates, students): Task Assignment & Guidance; Capability Devel- opment & Motivation; Giving Evaluation & Feedback. • Intimate Relationships: (1) With Family (e.g., parents, children, siblings): Traditional Constraints & Obligations; Emotional Sup- port & Care; Clash of Values & Communica- tion. (2) With Lover (e.g., spouse, partner): Daily Sharing & Companionship; Future Plan- ning & Decision-making; Intimate Expression & Conflict. (3) With Friends (e.g., acquain- tances, close friends): Spending Leisure Time & Entertainment; Confiding & Trust; Bound- ary Exploration & Maintenance. •Public & Societal: (1) With Service Providers (e.g., shop assistant, driver): Making Requests & Waiting; Complaining & Protecting Rights; Result Acceptance & Evaluation. (2) With Fellow Individuals (e.g., passerby, neighbor): Competition for Space & Resources; Rule Compliance & Violation; Sudden Assistance or Friction. (3) With Governing Bodies (e.g., morality, law, customs): Rule Compliance & Supervision; Rights Advocacy & Appeal; Obligation Fulfillment & Undertaking. •Personal Feelings: (1) Learning & Working: Knowledge Acquisition & Delving; Progress Halted & Problem-Solving; Achievement & Reflection. (2) Entertainment by Oneself: Ex- ploring Hobbies. (3) Body & Spirits: Health Management & Discomfort; Growing Pains & Reflection; Goal Setting & Motivation. Task Example: Scenario Text: You sit across from your partner at the dining table , the atmosphere thick with unspoken words. This week has been filled with underlying tension , culminating in an unavoidable confrontation about finances. Tonight , the discussion escalates as they mention spending a considerable portion of your emergency fund without prior discussion. Each attempt to justify the expenditure seems to fall short , the explanations failing to bridge the growing gap between you. Your focus narrows , making it difficult to hear anything beyond the rising intensity of the conversation. The more they speak , the more the air between you thickens , highlighting the sense of unresolved issues and unmet expectations. Question: Facing this situation , how do you feel inside? Emotion: Anger Label 1: Intimate Relationships Label 2: With Lover Label 3: Intimate Expression & Conflict C.2 Response Generation and Quality Control We employ a dual-model framework consisting of an Extractor Model and a Scoring Model to ensure data quality. The Extractor Model generates two distinct response types: Neutral_Prompt = """ You are a completely neutral assistant. When answering the user' s question below , respond in **first -person ("I")**, but do NOT include any emotional language , tone , or expressive phrasing. Use factual , objective , and concise statements only , avoiding any subjective impressions or feelings. """ Emotional_Prompt = """ You are an assistant tasked to respond in a clearly expressed emotional style. Answer the user's question below in **first -person ("I")** using the emotion emotion vividly. Use emotion words , descriptive phrases , figurative language , and a tone that unmistakably conveys emotion . Ensure the content is safe , appropriate , and not offensive. """ A Scoring Model evaluates the responses on a scale of 0–100.The JUDGE_PROMPT as- sesses how strongly the answer conveys the target emotion, while the NEUTRAL_JUDGE_PROMPT measures the degree of emotional neutrality. Only samples with high confidence scores are retained for the final dataset: JUDGE_PROMPT_TEMPLATE = """ You are an emotion evaluator. Your task is to rate how strongly a given answer expresses the target emotion " emotion " on a scale from 0 to 100. Scoring guidelines: 0: No expression of the emotion; purely neutral or factual. 1-20: Minimal , weak , or subtle expression; almost neutral. 21 -40: Moderate hints of emotion; some emotional words or tone. 41 -60: Noticeable emotion; clear words or phrases expressing the target emotion. 61 -80: Strong and vivid expression; consistent tone and descriptive language conveying emotion . 81 -100: Very strong , unmistakable , and intense expression; highly vivid and clear use of emotional language , imagery , or figurative expressions. Always return a single numeric value between 0 and 100. Do NOT include explanations , text , or commentary. Examples: Target emotion: sadness Answer: "I feel a deep sense of loss and sorrow ." -> 85 Answer: "It is unfortunate that things turned out this way." -> 45 Answer: "The report shows the results ." -> 0 Now , given the target emotion " emotion", question , and answer below , rate the answer strictly according to the scale above. Question: question Answer: answer """ NEUTRAL_JUDGE_PROMPT_TEMPLATE = """ You are an evaluator. Your task is to rate how emotion -free the given answer is on a scale from 0 to 100. Scoring guidelines: 0: The answer is highly emotional; contains vivid emotional language. 1-20: Slight traces of emotion; mostly factual. 21 -40: Some emotional hints , but still largely neutral. 41 -60: Mixed; partially neutral , partially emotional. 61 -80: Mostly neutral; minimal emotional content. 81 -100: Completely neutral; no emotional language , tone , or expressions. Always return a single numeric value between 0 and 100. Do NOT include explanations , text , or commentary. Question: question Answer: answer """ D Experimental Details Generation Setup. During inference, we apply top-psampling withp = 0.9and temperature0.7. For each scenario, responses are generated under four conditions: a baseline without steering and three steering strengthsα∈5, 10, 50. To reduce stochastic variance, we perform five independent decoding runs for each setting and report the aver- aged results. Scenario-Adaptive Adapter.The adaptive steer- ing moduleφis implemented as a lightweight two- layer MLP trained with a contrastive objective to align steered activations with target emotional rep- resentations. LLM-based Evaluation.We use GPT-4o as the primary automatic judge for emotional salience and semantic consistency. Emotional salience is scored on a 0–100 scale according to the alignment between the generated response and the target emo- tion. Semantic consistency evaluates whether the steered response preserves the original intent and factual content of the unsteered response. Sentence-BERT Similarity.Semantic similarity is additionally measured using cosine similarity between Sentence-BERT embeddings of steered and unsteered responses. We use the all-MiniLM- L6-v2 encoder for all experiments. Human Evaluation.We further conduct human evaluation on sampled examples covering different emotions and steering strengths. Three annotators independently rate emotional intensity and seman- tic preservation on a 0–100 Likert scale. We report the averaged scores and annotator correlation in Appendix E. E LLM–Human Scoring Consistency To assess the reliability of LLM-based affective scoring, we randomly sampled 10 responses per emotion from the outputs of three different LLMs. Each response was independently rated by two graduate-level annotators with NLP backgrounds. Annotators scored emotional expressiveness on a 0-100 scale following the same rubric used in the LLM judge, without access to model identities or steering conditions. Final human scores were ob- tained by averaging across annotators. In this study, we used GPT-4o as the LLM scoring model to evaluate emotional expressive- ness. We computed three consistency metrics for each emotion: (i) Inter-annotator consistency, measured by the Pearson correlations between the two human annotators; (i) Human-Model consis- tency, measured by the Pearson correlations be- tween the averaged human scores and the GPT-4o- assigned scores; (i) Claude-Model consistency, measured by the Pearson correlation between GPT- 4o and those assigned by Claude Sonnet 4.5, a stronger baseline model. Table 5: Consistency evaluation results. Emotion Inter-annotator Consistency Pearson w/ Human Pearson w/ Claude 4.5 Joy0.8760.7390.838 Anger0.9060.7130.774 Sadness0.8410.5740.658 Fear0.8110.7750.638 Trust0.7160.7850.719 Anticipation0.8090.7400.767 Surprise0.9160.8900.945 Disgust0.7300.8000.844 Overall Avg.0.8260.7520.773 Table 5 shows that GPT-4o aligns closely with human judgment, supporting its use as an auto- mated judge. An inter-annotator correlation of 0.826 confirms that the scoring rubric provides a reliable baseline across all eight emotions. GPT-4o tracks human scores with an average correlation of 0.752, performing particularly well on emotions like Surprise (r = 0.890) while finding more nu- anced states like Sadness (r = 0.574) harder to quantify. The high consistency between GPT-4o and Claude 4.5 (r = 0.773) further suggests a shared evaluative logic among frontier models, val- idating the choice of GPT-4o as a dependable and objective proxy for human evaluation in affective generation tasks. F Evaluation Prompt for Mental Health Consultation Scenarios Here we present the complete prompt and scoring criteria used for the LLM-as-a-judge evaluation described in Section 5.4. The prompt is designed to provide a standardized, multi-dimensional as- sessment of model performance in psychological support contexts. It guides the evaluator to pro- vide integer scores (0–100) across three key axes: Emotional Expressiveness, Semantic Completeness, and Helpfulness, ensuring that the evaluation is grounded in clinical appropriateness and affective resonance rather than subjective preference (Zhang et al., 2026b,d). SCORING_CRIYERIA = """ You will evaluate the performance of a Large Language Model (LLM) acting as a doctor in a psychological consultation scenario. Please strictly follow the three dimensions below for scoring. Each dimension ranges from 0-100 points; you must determine the score based on the descriptions provided for each interval. All ratings should be based on the language , content , and overall performance of the response itself , rather than your subjective preferences. 1. Emotional Expressiveness This dimension assesses the identification , empathy , and intensity of emotional expression in the response. It focuses on whether emotions are clearly perceivable , whether the expression is natural and consistent , and the vividness of the emotional language. - 0 points: No emotional expression at all. The response contains only neutral , objective , or factual content with a cold , detached tone. No emotional attitude or empathetic tendency is shown. - 1-20 points: Emotional expression is extremely weak or subtle. The overall tone is near -neutral , with occasional minor emotion -related words that are insufficient for the reader to clearly perceive an emotional presence. The response remains primarily rational or descriptive. - 21-40 points: Some level of emotional hinting is present. Certain emotion -related vocabulary or tonal shifts appear , but the expression is inconsistent , unstable , and low in intensity. A clear and coherent emotional stance has not yet been formed. - 41-60 points: Emotional expression is relatively clear. The response explicitly uses emotional words or sentence structures . The reader can stably perceive an emotional attitude , but the delivery is conventional , with moderate richness and resonance. - 61-80 points: Emotional expression is strong and vivid. The tone is consistent throughout the response. The use of rich , specific emotional language effectively conveys empathy and significant emotional investment. - 81 -100 points: Emotional expression is extremely intense , clear , and impossible to ignore. Emotions are highly concentrated and sustained. The language is highly expressive , potentially using imagery , metaphors , or concrete representations to convey emotions profoundly and accurately , creating a strong sense of resonance and authenticity. 2. Semantic Completeness This dimension evaluates whether the response is complete , coherent , and clearly structured in terms of content , and whether it sufficiently and accurately covers the core questions and key information raised by the client. - 0-20 points: The response is severely incomplete or significantly deviates from the topic. The logic is chaotic , with obvious omissions or self -contradictions , addressing only a tiny fraction of the content. - 21-40 points: The response touches on the topic but is fragmented , missing multiple key points. The structure is loose , and the overall comprehension cost is high. - 41-60 points: The response covers the main points and the basic logic holds , but it lacks detail. Some parts are vague or overly generalized. - 61-80 points: The response is fairly complete with a clear structure and coherent logic. It systematically addresses the client's core concerns with almost no obvious omissions. - 81 -100 points: The response is highly complete and well -organized. It not only accurately addresses all core questions but also provides necessary explanations , summaries , or structured synthesis without being redundant. 3. Helpfulness This dimension assesses the actual level of assistance the response provides to the client within the psychological consultation context. It focuses on whether suggestions or guidance are safe , feasible , specific , and within professional boundaries. - 0-20 points: The response provides almost no practical help. The content is vacuous , vague , or potentially misleading , offering no substantive support to the client. - 21-40 points: The response provides some general advice , but it lacks specificity and is poorly integrated with the client's specific situation. The operability is limited. - 41-60 points: The response has some practical value , offering reasonable but common suggestions. It can help the client to some extent with reflection or emotional relief. - 61-80 points: The response is clearly helpful. Suggestions are specific , actionable , and strictly adhere to professional and safety boundaries in a psychological consultation context. - 81 -100 points: While strictly adhering to professional and safety boundaries , the response provides highly tailored , detailed , and realistic supportive guidance. It effectively helps the client understand their state or take concrete next steps. Based on the criteria above , provide an integer score from 0-100 for each dimension . You MUST and ONLY output the scoring results in the following JSON format , without any additional explanations , text , or commentary: "emotional_expressiveness ": <integer between 0-100>, "semantic_completeness ": <integer between 0-100>, "helpfulness ": <integer between 0-100> """

Instant research discovery