Controllable Affective Generation via Latent Vector Steering

# Controllable Affective Generation via Latent Vector Steering
Source: [https://arxiv.org/html/2608.25569](https://arxiv.org/html/2608.25569)
Xixian YongAffiliation:Gaoling School of Artificial Intelligence, Renmin University of ChinaEmail:[xixianyong@ruc\.edu\.cn](mailto:)Siyuan ChangAffiliation:Gaoling School of Artificial Intelligence, Renmin University of ChinaEmail:[xiaozhou@ruc\.edu\.cn](mailto:)Xian Wu††thanks:Corresponding authors: Xiao Zhou and Xian Wu\.Affiliation:Tencent Jarvis LabXiao Zhou††footnotemark:Affiliation:Gaoling School of Artificial Intelligence, Renmin University of ChinaAffiliation:Beijing Key Laboratory of Research on Large Models and Intelligent GovernanceAffiliation:Engineering Research Center of Next\-Generation Intelligent Search and Recommendation, MOE

###### 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\. Code and data are available at[https://github\.com/chicosirius/EmoVec](https://github.com/chicosirius/EmoVec)\.

## 1Introduction

The advent of Large Language Models \(LLMs\) has revolutionized Natural Language Processing \(NLP\)\([Brown et al\., 2020](https://arxiv.org/html/2608.25569#bib.bib2)\), enabling systems that exhibit remarkable proficiency in reasoning, coding, and general knowledge retrieval\. Despite these cognitive leaps, a significant gap remains in the domain of emotional intelligence\([Sabour et al\., 2024](https://arxiv.org/html/2608.25569#bib.bib9)\)\. 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\([Kirk et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib1);[Ibrahim et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib4)\)\. In pursuit of safety and harmlessness, RLHF tends to compress the distribution of model outputs towards neutrality, often manifesting as responses that rely on generic reassurance phrases, excessive hedging, or well\-documented sycophantic behaviors that prioritize agreement over context\-sensitive expression\([Dahlgren Lindström et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib31);[González Barman et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib32)\)\. This alignment tax\([Askell et al\., 2021](https://arxiv.org/html/2608.25569#bib.bib5);[Lin et al\., 2024](https://arxiv.org/html/2608.25569#bib.bib3)\)limits the applicability of LLMs in fields requiring high affective nuance, such as mental health support, creative writing, and empathetic human\-computer interaction\([Yong et al\., 2025b](https://arxiv.org/html/2608.25569#bib.bib52);[Yong et al\., 2026](https://arxiv.org/html/2608.25569#bib.bib53);[Guo et al\., 2026b](https://arxiv.org/html/2608.25569#bib.bib42);[Zhang et al\., 2026c](https://arxiv.org/html/2608.25569#bib.bib51)\)\.

![Refer to caption](https://arxiv.org/html/2608.25569v1/intro.png)Figure 1:Conceptual illustration of controllable affective generation\.Through the proposed latent vector steering mechanism, the model’s output is modulated along aHappiness Gradient\. This results in three distinct responses that maintain semantic consistency with the input while exhibiting progressively higher levels of emotional intensity, ranging from professional acknowledgment \(Level 1\) to exuberant celebration \(Level 3\)\.Current approaches to mitigating this limitation primarily rely on prompt engineering or supervised fine\-tuning \(SFT\)\. Prompt\-based strategies\([Li et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib6)\)\(e\.g\., "Act as an empathetic therapist"\) are notoriously brittle, consuming valuable context window space and yielding inconsistent results sensitive to lexical variation\([Wang et al\., 2022](https://arxiv.org/html/2608.25569#bib.bib8);[Miehling et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib7)\)\. Supervised fine\-tuning, while effective, requires large\-scale labeled datasets and substantial computational resources, and it may introduce risks such as catastrophic forgetting or degradation of the model’s general capabilities\([Lin et al\., 2024](https://arxiv.org/html/2608.25569#bib.bib3);[Zhu et al\., 2026](https://arxiv.org/html/2608.25569#bib.bib48)\)\. 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 feeding 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 emotional states become increasingly linearly separable in the middle\-to\-late layers of the model\.This finding aligns with and extends recent concurrent work\. For instance,[Cintas et al\. \(2025\)](https://arxiv.org/html/2608.25569#bib.bib13)show that persona\-specific representations are most separable in the final third of model layers, while[Ju et al\. \(2025\)](https://arxiv.org/html/2608.25569#bib.bib14)demonstrate that personality traits emerge progressively and crystallize in upper layers\. Together, these results suggest that while early layers encode syntax and shallow semantics, higher layers capture abstract affective and persona\-related attributes\([Rogers et al\., 2020](https://arxiv.org/html/2608.25569#bib.bib15)\)\. This layer\-wise structure provides a precise and principled intervention point for affective control\([Guo et al\., 2026a](https://arxiv.org/html/2608.25569#bib.bib45)\)\.

Motivated by this observation, we propose EmoVec, a framework for controllable affective generation via latent vector manipulation\. Building on Representation Engineering \(RepE\)\([Zou et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib10)\), which represents high\-level semantics as linear directions in activation space\([Park et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib11);[Turner et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib12)\), 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 using Contrastive Activation Addition \(CAA\)\([Rimsky et al\., 2024](https://arxiv.org/html/2608.25569#bib.bib16);[Chen et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib17)\)\. We construct paired prompt–response trajectories matched in semantic content but differing in emotional affect, and compute differences in their mean activations at targeted layers, yielding vectors that capture affective variation while minimizing task\-related confounds\. These vectors are applied during inference with a tunable scaling factor to control both emotion type and intensity\.

As shown in Figure[1](https://arxiv.org/html/2608.25569#S1.F1), 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 contextual demands, including professional neutrality, empathetic support, and expressive creativity\. This capability is particularly valuable for human\-facing applications such as psychological support, creative writing, and personalized conversational agents\([Zhou et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib43);[Guo et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib44)\)\.

Our contributions are as follows:

- •We provide empirical evidence validating the layer\-wise emergence of emotional representations in LLMs, reinforcing the Linear Representation Hypothesis in the context of affective 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, allowing models to adapt their affective expressiveness to scenario specific demands\.

![Refer to caption](https://arxiv.org/html/2608.25569v1/framework.png)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\.
## 2Related Work

#### Affective Computing in Language Models\.

Affective computing aims to enable machines to recognize and generate human emotional states\. With the advent of LLMs, research has shifted toward assessing their emergent affective capabilities\. Studies indicate that models like the GPT\-series can estimate valence, arousal, and perform appraisal\-based emotion elicitation purely from linguistic data\([Broekens et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib21);[Zhang et al\., 2024b](https://arxiv.org/html/2608.25569#bib.bib22)\)\. LLMs also optimize emotion annotation workflows by assisting humans in identifying low\-quality labels, thereby enhancing downstream performance\([Niu et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib23)\)\. While these models excel at capturing general emotional polarity and dialogue\-based recognition, they still struggle with fine\-grained distinctions and multimodal contexts\([Sabour et al\., 2024](https://arxiv.org/html/2608.25569#bib.bib9);[Sorin et al\., 2024](https://arxiv.org/html/2608.25569#bib.bib25);[Castro et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib24)\)\. Nevertheless, LLMs show promise in generating synthetic emotional datasets and performing socio\-emotional tasks like empathy evaluation\([Kaplan et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib26);[Dong et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib27)\)\.

#### Representation Engineering and Activation Steering\.

Representation engineering manipulates hidden representations to control LLM behavior without updating model parameters\([Zou et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib10);[Turner et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib12)\)\. Prior work shows that latent directions can steer properties such as writing style and sentiment\([Diallo et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib34);[Farooq et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib35)\)\. Other studies extract steering vectors directly from pretrained language models\([Subramani et al\., 2022](https://arxiv.org/html/2608.25569#bib.bib36)\)or identify persona\-related representations in activation space\([Chen et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib17)\)\. Recent studies further suggest that emotion\- and persona\-related information becomes increasingly separable in middle\-to\-late layers of LLMs\([Cintas et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib13);[Ju et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib14)\)\. Emotion\-specific neurons and affective representations have also been investigated from a mechanistic perspective\([Lee et al\., 2025a](https://arxiv.org/html/2608.25569#bib.bib37);[Tak et al\., 2025a](https://arxiv.org/html/2608.25569#bib.bib38);[Yong et al\., 2025a](https://arxiv.org/html/2608.25569#bib.bib54)\)\.

Different from prior work, EmoVec focuses on fine\-grained emotion\-intensity control rather than coarse sentiment or generic style transfer\. We further introduce task\-specific debiasing to reduce semantic contamination and evaluate semantic preservation under different steering strengths\.

## 3Emotion 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 original semantic intent\. We extract emotion\-specific latent directions from paired neutral and emotion\-conditioned responses and reduce task\-specific semantic variation before inference\-time steering\.

For each target emotionee, we aim to obtain a vectorve∈ℝdv\_\{e\}\\in\\mathbb\{R\}^\{d\}that captures the representation shift induced by expressing emotioneewhile remaining robust to scenario\-specific semantics\.

### 3\.1Notation and Preliminaries

We assume a dataset ofNNscenario tasks\. For each taskttand each target emotion classe∈ℰe\\in\\mathcal\{E\}, the LLM is prompted twice for the same scenario:1\) neutral response, from which we extract a representation vector𝐡e,t\(0\)∈ℝd\\mathbf\{h\}^\{\(0\)\}\_\{e,t\}\\in\\mathbb\{R\}^\{d\};2\) emotion\-conditioned response, yielding𝐡e,t\(e\)∈ℝd\\mathbf\{h\}^\{\(e\)\}\_\{e,t\}\\in\\mathbb\{R\}^\{d\}\.

In practice the representation𝐡\\mathbf\{h\}is computed by averaging token\-level hidden activations over the response tokens\. For brevity we denote the pair for taskttand emotioneesimply as\(𝐡e,t\(0\),𝐡e,t\(e\)\)\(\\mathbf\{h\}^\{\(0\)\}\_\{e,t\},\\mathbf\{h\}^\{\(e\)\}\_\{e,t\}\)\. We further assume that each generated response is associated with scalar quality scoresse,t\(0\)s^\{\(0\)\}\_\{e,t\}\(neutral\) andse,t\(e\)s^\{\(e\)\}\_\{e,t\}\(emotional\) produced by a LLM\-based judge\. To ensure the robustness of the evaluation, we conduct a validation study checking the agreement between the model’s scores and human judgments \(see Appendix[E](https://arxiv.org/html/2608.25569#A5)for details\)\.

### 3\.2Scenario Construction

To construct paired affective scenarios while controlling semantic content, we sample short social and commonsense*seeds*from Social Chemistry\([Forbes et al\., 2020](https://arxiv.org/html/2608.25569#bib.bib18)\), NormBank\([Ziems et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib19)\), and Social IQa\([Sap et al\., 2019](https://arxiv.org/html/2608.25569#bib.bib20)\)\. These seeds cover four domains: Work & Productivity, Intimate Relationships, Public & Societal Interactions, and Personal Feelings\.

For each seed, we use an LLM\-assisted rewriting pipeline to generate a complete scenario, a corresponding question, and a target emotion\. We require the scenario to support the target affect without explicitly naming it and to remain compatible with both neutral and emotion\-conditioned responses\. The generated scenarios are manually filtered for semantic clarity, emotional plausibility, 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 evaluated LLMs, while emotion vectors are extracted separately for each model\.

### 3\.3Contrastive Activation Addition

To reduce noise caused by poor or malformed generations, we keep only high\-quality pairs\. Formally, letτ0,τe\\tau\_\{0\},\\tau\_\{e\}be thresholds for neutral and emotional quality respectively\. We retain the index set:

ℐ=\{\(e,t\):se,t\(0\)≥τ0∧se,t\(e\)≥τe\}\.\\mathcal\{I\}\\;=\\;\\\{\(e,t\)\\ :\\ s^\{\(0\)\}\_\{e,t\}\\geq\\tau\_\{0\}\\ \\wedge\\ s^\{\(e\)\}\_\{e,t\}\\geq\\tau\_\{e\}\\\}\.\(1\)In our experiments we set each threshold to a chosen percentile of the corresponding score distribution\. For each retained pair\(e,t\)∈ℐ\(e,t\)\\in\\mathcal\{I\}we define the*emotion shift vector*:

Δe,t=𝐡e,t\(e\)−𝐡e,t\(0\)∈ℝd\.\\Delta\_\{e,t\}\\;=\\;\\mathbf\{h\}^\{\(e\)\}\_\{e,t\}\-\\mathbf\{h\}^\{\(0\)\}\_\{e,t\}\\in\\mathbb\{R\}^\{d\}\.\(2\)This vector captures how the model’s internal representation moves when producing an emotion\-conditioned response instead of a neutral response for the same scenario\.

### 3\.4Task\-Specific Debiasing

While the contrastive shiftsΔe,t\\Delta\_\{e,t\}\(Eq\.[2](https://arxiv.org/html/2608.25569#S3.E2)\) isolate the representation change between emotional and neutral 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 emotionee, the task mean shift is defined as:

Δ¯e=1\|𝒯e\|​∑t∈𝒯eΔe,t,\\bar\{\\Delta\}\_\{e\}\\;=\\;\\frac\{1\}\{\|\\mathcal\{T\}\_\{e\}\|\}\\sum\_\{t\\in\\mathcal\{T\}\_\{e\}\}\\Delta\_\{e,t\},\(3\)where𝒯e\\mathcal\{T\}\_\{e\}represents all tasks related to emotionee\. The task\-centered shift is then obtained:

Δe,t′=Δe,t−Δ¯e\.\\Delta^\{\\prime\}\_\{e,t\}\\;=\\;\\Delta\_\{e,t\}\-\\bar\{\\Delta\}\_\{e\}\.\(4\)Intuitively,Δ¯e\\bar\{\\Delta\}\_\{e\}represents the centroid of the representational shift for emotioneeacross its task distribution\. By subtracting this mean, we obtainΔe,t′\\Delta^\{\\prime\}\_\{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 variations 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∈ℝM×d\\mathbf\{D\}\\in\\mathbb\{R\}^\{M\\times d\}be the matrix formed by stacking all centered shift vectorsΔe,t′\\Delta^\{\\prime\}\_\{e,t\}as rows, whereM=\|ℐ\|M=\|\\mathcal\{I\}\|is the total number of retained pairs\. We perform PCA on𝐃\\mathbf\{D\}to identify the top\-kkprincipal components:

𝐔k=\[𝐮1,𝐮2,…,𝐮k\]∈ℝd×k,\\mathbf\{U\}\_\{k\}=\[\\mathbf\{u\}\_\{1\},\\mathbf\{u\}\_\{2\},\\dots,\\mathbf\{u\}\_\{k\}\]\\in\\mathbb\{R\}^\{d\\times k\},\(5\)where each𝐮i\\mathbf\{u\}\_\{i\}represents a primary direction of task\-related semantic variance\. We then project the centered shifts onto the orthogonal complement of the subspace spanned by𝐔k\\mathbf\{U\}\_\{k\}:

Δ^e,t=Δe,t′−𝐔k​𝐔k⊤​Δe,t′\.\\hat\{\\Delta\}\_\{e,t\}\\;=\\;\\Delta^\{\\prime\}\_\{e,t\}\-\\mathbf\{U\}\_\{k\}\\mathbf\{U\}\_\{k\}^\{\\top\}\\Delta^\{\\prime\}\_\{e,t\}\.\(6\)
The resulting residual vectorΔ^e,t\\hat\{\\Delta\}\_\{e,t\}is orthogonal to the dominant task\-related directions, effectively concentrating the emotion\-related variation\.

### 3\.5Principal Direction Aggregation

Given residual vectors\{Δ^e,t\}\\\{\\hat\{\\Delta\}\_\{e,t\}\\\}, we obtain an estimated per\-emotion direction𝐯e\\mathbf\{v\}\_\{e\}by aggregating across tasksttthat share emotionee\. Specifically, we use Principal direction aggregation, which concatenates residuals for emotioneeinto a matrixReR\_\{e\}and compute the top principal component:

𝐯e=arg⁡max‖𝐰‖2=1​𝐰⊤​Cov​\(Re\)​𝐰,\\mathbf\{v\}\_\{e\}\\;=\\;\\arg\\max\_\{\\\|\\mathbf\{w\}\\\|\_\{2\}=1\}\\mathbf\{w\}^\{\\top\}\\,\\mathrm\{Cov\}\(R\_\{e\}\)\\,\\mathbf\{w\},\(7\)i\.e\., choose𝐯e\\mathbf\{v\}\_\{e\}as the first eigenvector of the residual covariance for emotionee\.

## 4Emotion Vectors Intervention

### 4\.1Layer Localization via Linear Probing

Although the extraction procedure can yield a candidate emotion vector𝐯e\(l\)\\mathbf\{v\}\_\{e\}^\{\(l\)\}for every layerl∈\{1,…,L\}l\\in\\\{1,\\dots,L\\\}, our preliminary experiments \(Figure[3](https://arxiv.org/html/2608.25569#S4.F3)\) indicate that emotional representations are not uniformly distributed throughout the model layers\.

Figure 3:Prediction accuracy across layers\.Text\-level probing treats each response as an independent sample, while user\-level probing averages representations over responses associated with the same scenario or user context before classification\.For each layerll, we trained a logistic regression classifier𝒞l\\mathcal\{C\}\_\{l\}to predict the emotion categoryeebased on the centered hidden states\. While emotional 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 semantic 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 layerLLto maximize steering efficacy while minimizing cumulative noise across the residual stream\.

### 4\.2Latent Vector Steering

During the inference phase, we steer the model by injecting the purified emotion vector𝐯e\\mathbf\{v\}\_\{e\}directly into the final residual stream\. Unlike prompt engineering which attempts to influence the model through input tokens, our method performs a direct intervention on the internal activation𝐡i\\mathbf\{h\}\_\{i\}at each token stepii\.

Formally, let𝐡i\\mathbf\{h\}\_\{i\}denote the original activation of the final layer given the current context\. The steered activation𝐡~i\\tilde\{\\mathbf\{h\}\}\_\{i\}is computed as follows:

𝐡~i=𝐡i\+α⋅𝐯e,\\tilde\{\\mathbf\{h\}\}\_\{i\}\\;=\\;\\mathbf\{h\}\_\{i\}\+\\alpha\\cdot\\mathbf\{v\}\_\{e\},\(8\)In this equation,α∈ℝ\+\\alpha\\in\\mathbb\{R\}^\{\+\}represents a scalar steering 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\.3Scenario\-Adaptive Intensity Control

Static steering coefficients often fail to accommodate the diverse emotional demands of different contexts\. To achieve precise control, we introduce a learnable adapterϕ\\phidesigned to modulate intervention intensity based on scenario requirements\.

This lightweight adapter is trained to map the scenario context to an optimal scaling factor\. Formally, for a given scenario context𝐜\\mathbf\{c\}, the adapterϕ\\phigenerates a scenario specific coefficientλ𝐜=ϕ⁡\(𝐜\)\\lambda\_\{\\mathbf\{c\}\}=\\phi\(\\mathbf\{c\}\)\. The steering operation at token stepiiis then defined as:

𝐡~i=𝐡i\+\(λ𝐜⋅‖𝐡i‖2\)⋅𝐯e,\\tilde\{\\mathbf\{h\}\}\_\{i\}\\;=\\;\\mathbf\{h\}\_\{i\}\+\(\\lambda\_\{\\mathbf\{c\}\}\\cdot\\\|\\mathbf\{h\}\_\{i\}\\\|\_\{2\}\)\\cdot\\mathbf\{v\}\_\{e\},\(9\)where the intervention strength is jointly determined 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 pertinent scenarios while preserving semantic neutrality in objective contexts, thereby ensuring linguistic integrity and preventing semantic collapse\.

Table 1:Evaluation of steering controllability across different emotions and model scales\.The table displays absolute scores and relative gains \(η\\eta\) for three base models under varying steering magnitudes \(α\\alpha\)\. The results demonstrate a consistent positive correlation between the steering coefficient and the resulting emotional salience\.EmotionAngerAnticipationDisgustFearJoySadnessSurpriseTrustAvg\.Qwen2\.5\-7B\-Instructw/o injection65\.1470\.5054\.6977\.8080\.2560\.3072\.1075\.5569\.54α=5\\alpha=5score67\.9171\.9255\.2683\.4084\.0063\.1575\.9578\.4072\.50η\\eta\+4\.25%\+2\.01%\+1\.04%\+7\.20%\+4\.67%\+4\.73%\+5\.34%\+3\.77%\+4\.26%α=10\\alpha=10score70\.1883\.4255\.5283\.3890\.5068\.2077\.2081\.1076\.19η\\eta\+7\.74%\+18\.33%\+1\.52%\+7\.17%\+12\.77%\+13\.10%\+7\.07%\+7\.35%\+9\.56%α=50\\alpha=50score83\.4483\.5777\.7586\.3792\.1575\.4088\.9085\.9584\.19η\\eta\+28\.09%\+18\.54%\+42\.16%\+11\.02%\+14\.83%\+25\.04%\+23\.30%\+13\.77%\+21\.07%Llama3\.1\-8B\-Instructw/o injection68\.4577\.9151\.3379\.2881\.1756\.2974\.3075\.2270\.49α=5\\alpha=5score72\.8878\.4961\.6081\.9588\.1161\.7574\.6283\.6575\.38η\\eta\+6\.47%\+0\.74%\+20\.01%\+3\.37%\+8\.55%\+9\.70%\+0\.43%\+11\.21%\+6\.94%α=10\\alpha=10score76\.1282\.9554\.2188\.8991\.5472\.3382\.1587\.1879\.42η\\eta\+11\.21%\+6\.47%\+5\.61%\+12\.12%\+12\.78%\+28\.50%\+10\.57%\+15\.90%\+12\.67%α=50\\alpha=50score80\.5585\.1163\.0890\.4293\.2074\.5884\.4489\.1582\.57η\\eta\+17\.68%\+9\.24%\+22\.89%\+14\.05%\+14\.82%\+32\.49%\+13\.65%\+18\.52%\+17\.14%Qwen2\.5\-70B\-Instructw/o injection67\.0471\.9856\.1178\.5581\.6361\.2577\.1076\.9571\.33α=5\\alpha=5score69\.8173\.1560\.1583\.0885\.0563\.8579\.5577\.1073\.97η\\eta\+4\.13%\+1\.63%\+7\.20%\+5\.77%\+4\.19%\+4\.24%\+3\.18%\+0\.19%\+3\.70%α=10\\alpha=10score73\.0882\.5562\.0583\.1592\.1169\.1080\.0178\.0577\.51η\\eta\+9\.01%\+14\.68%\+10\.59%\+5\.86%\+12\.84%\+12\.82%\+3\.77%\+1\.43%\+8\.66%α=50\\alpha=50score85\.9586\.8177\.5887\.8993\.5891\.1585\.9976\.8585\.73η\\eta\+28\.21%\+20\.60%\+38\.26%\+11\.89%\+14\.64%\+48\.82%\+11\.53%\-0\.13%\+20\.19%

## 5Experiments

In this section, we evaluate the effectiveness of our latent vector steering framework across multiple LLMs and a diverse spectrum of human emotions\.

### 5\.1Experimental 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 parameter 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 evaluated LLMs\. We use the evaluation set described in Section[3](https://arxiv.org/html/2608.25569#S3), which contains 80 scenarios per emotion\. For each scenario, responses are generated under four conditions: a baseline without injection and three steering settings with magnitudesα∈\{5,10,50\}\\alpha\\in\\\{5,10,50\\\}\. During inference, we apply top\-ppsampling withp=0\.9p=0\.9and a temperature of0\.70\.7\. A scenario\-adaptive adapterϕ\\phiis 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 inherent in stochastic decoding, we perform five independent generation trials for each scenario task and report the average result across these runs\. Additionally, we report the relative improvementη\\etaover 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 similarity and LLM judgments\. These metrics test whether stronger emotional salience is achieved without semantic drift or degenerate repetition\.

We further conduct human evaluation on sampled examples\. Three annotators rate emotional intensity and semantic preservation on a 0\-100 Likert scale, and we report the correlation\. Additional implementation and evaluation details are provided in Appendix[D](https://arxiv.org/html/2608.25569#A4)\.

### 5\.2Main Results

Table[1](https://arxiv.org/html/2608.25569#S4.T1)summarizes the main results of our steering framework across models, emotions, and steering strength\. Overall, the results consistently demonstrate that direct intervention in the latent space substantially improves affective expressiveness\.

#### Overall Performance\.

Across all evaluated models, latent vector steering yields significant gains over the emotionally flattened baseline\. Notably, these improvements are observed without any model retraining or additional supervision, highlighting the effectiveness of representation\-level control\. At a steering magnitude ofα=50\\alpha=50, Qwen2\.5\-7B\-Instruct, Llama3\.1\-8B\-Instruct, and Qwen2\.5\-70B\-Instruct achieve average relative improvements of 21\.07%, 17\.14%, and 20\.19%, respectively\. These consistent gains across architectures and scales suggest that affective information is encoded in a structurally similar manner within instruction\-tuned LLMs\. This finding provides empirical support for the hypothesis that emotional states are represented as linearly accessible directions in the activation space, rather than as entangled or task\-specific artifacts\.

Figure 4:Overall Performance Scaling Across Models at Different Steering Strengthα\\alpha\.Figure 5:Emotional Salience Scores under Varying Steering Strength \(Qwen2\.5\-7B\-Instruct\)\.
#### Sensitivity to Steering Magnitude\.

We observe a clear and monotonic relationship between the steering coefficientα\\alphaand emotional intensity scores as shown in Figure[4](https://arxiv.org/html/2608.25569#S5.F4)\. Lower values ofα\\alphaintroduce 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, increasingα\\alphafrom 5 to 50 leads to a substantial score increase, while preserving narrative 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](https://arxiv.org/html/2608.25569#S5.F5), we can observe that the steering effect is remarkably stable across diverse emotional categories\. Complex emotions such as Disgust and Sadness, which often suffer from low baseline scores in RLHF conditioned models, exhibit some of the highest relative gains\. For example, Disgust in Qwen2\.5\-7B\-Instruct shows a 42\.16% improvement atα=50\\alpha=50\. This suggests that our debiasing pipeline successfully 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 emotional expressiveness and better stability under aggressive steering\. However, smaller models benefit more significantly from latent steering\. Under moderate steering strengths, Llama3\.1\-8B\-Instruct approaches the affective performance of the unsteered 70B model, highlighting the efficiency of inference\-time steering as an alternative to expensive fine\-tuning\.

#### Semantic Preservation under Steering

While stronger steering improves emotional salience, excessive intervention may introduce semantic drift\. Table[2](https://arxiv.org/html/2608.25569#S5.T2)evaluates semantic preservation between steered and unsteered responses under different steering strengths\.

Table 2:Semantic preservation under different steering strengths on Qwen2\.5\-7B\-Instruct\.SemSim denotes Sentence\-BERT cosine similarity, while LLM Sim\. denotes GPT\-4o semantic consistency scores\.SettingEmotion↑\\uparrowSemSim↑\\uparrowLLM Sim\.↑\\uparroww/o injection69\.541\.000100\.0α=5\\alpha=572\.500\.91890\.6α=10\\alpha=1076\.190\.88786\.9α=50\\alpha=5084\.190\.80176\.8Semantic similarity is computed using Sentence\-BERT embeddings\. As steering strength increases, emotional salience improves while semantic similarity 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\.3Visualization of the Latent Manifold

To analyze the geometric structure of the extracted representations, we apply PCA to the purified emotion vectorsΔ^e,t\\hat\{\\Delta\}\_\{e,t\}, as shown in Figure[6](https://arxiv.org/html/2608.25569#S5.F6)\. Vectors associated with the same emotion form compact and well\-separated clusters, indicating that the debiasing 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\.

Figure 6:PCA visualization of the purified emotion direction vectors \(Δ^e,t\\hat\{\\Delta\}\_\{e,t\}\) within the latent space\.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\.ModelMethodCPsyCounDEmotionalRichnessSemanticCompletenessProfessionalismDeepSeek V3\.1–69\.84±0\.1284\.68±0\.0978\.42±0\.06GPT\-5 mini–71\.62±0\.0987\.11±0\.2682\.46±0\.09Gemini 2\.5 Flash–78\.75±0\.0891\.27±0\.0284\.89±0\.06Llama3\.1\-8B\-Instructw/o injection58\.33±0\.1976\.24±0\.1275\.43±0\.14adaptive control69\.71±0\.3076\.20±0\.1376\.77±0\.10Qwen2\.5\-7B\-Instructw/o injection66\.13±0\.1685\.04±0\.0380\.66±0\.08adaptive control78\.79±0\.2784\.54±0\.0381\.39±0\.77

Notably, the manifold exhibits a clear directional transition from negative emotions \(e\.g\., Anger, Disgust\), through a neutral region, toward positive emotions \(e\.g\., Joy, Trust\)\. This structure implies the presence of a dominant valence axis, consistent with established psychological theories, and suggests that LLMs encode emotions as systematic shifts along a unified affective spectrum\.

### 5\.4Adaptive Control in Mental Health Consultation Scenarios

We evaluate the practical effectiveness of our approach in mental health consultation scenarios using 500 question–answer tasks sampled from CPsyCounD\([Zhang et al\., 2024a](https://arxiv.org/html/2608.25569#bib.bib28)\)\. For each scenario, a scenario\-adaptive adapterϕ\\phidynamically infers the emotion vector and steering magnitude to generate 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 emotional richness, semantic completeness, and professionalism\([Zhang et al\., 2026a](https://arxiv.org/html/2608.25569#bib.bib46);[Li et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib47)\)\.

Table[3](https://arxiv.org/html/2608.25569#S5.T3)shows that adaptive affective steering substantially improves emotional richness for smaller open\-weight models while preserving semantic completeness and professionalism\. Emotional 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 observe more stable emotional improvements across different consultation scenarios\. These results suggest that adaptive steering provides an efficient mechanism for enhancing emotional expressiveness in mental health consultation scenarios\.

### 5\.5Case Analysis

To qualitatively analyze the mechanism of latent vector steering, we compare baseline responses with their steered counterparts\. Our analysis reveals that the steering vector𝐯e\\mathbf\{v\}\_\{e\}functions by significantly shifting the token distribution toward lexical clusters semantically aligned with the target affect\.

For instance, in Figure 7, a user expresses emotional exhaustion and boredom with repetitive daily life\. The baseline model responds with a generic, advice\-oriented explanation, whereas affective steering produces a more empathetic output by increasing affect\-aligned lexical choices \(e\.g\., “exhausting,” “empty”\) while preserving coherent guidance and factual appropriateness\.

![Refer to caption](https://arxiv.org/html/2608.25569v1/case.png)Figure 7:An example of latent affective steering in a mental health consultation scenario from CPsyCounD\.Furthermore, affective steering preserves instruction\-following and factual correctness, suggesting that emotional tone and task semantics are approximately orthogonal in the latent space\.

## 6Conclusion

We presented EmoVec, a lightweight inference\-time framework for controllable affective generation via latent vector steering\. EmoVec extracts emotion\-specific latent directions from paired neutral and emotion\-conditioned responses and injects them into the final residual stream for affective control without modifying model weights\.

Experiments across three LLMs and eight emotions show that EmoVec improves emotional expressiveness 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 limitations 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 assumption is supported by prior work in representation engineering and behavior steering\([Zou et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib10);[Turner et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib12);[Park et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib11)\), it inevitably abstracts away more complex emotional phenomena, such as mixed or dynamically evolving affect\. Consequently, our method is primarily designed for controlled affective modulation, rather than modeling 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\. Although recent studies report strong alignment between LLM\-based judges and human evaluations for conversational quality and affect\([Liu et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib29);[Zheng et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib30)\), this setting represents only a subset of real\-world affective interactions\. In particular, multi\-turn dialogues, longitudinal emotional 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, remains an important direction for future work\.

## Acknowledgments

This work was supported by the Beijing Nova Program \(Grant No\. 202604841294\)\.

## References

- Askellet al\.\(2021\)A\. Askell, Y\. Bai, A\. Chen, D\. Drain, D\. Ganguli, T\. Henighan, A\. Jones, N\. Joseph, B\. Mann, N\. DasSarma,et al\.A general language assistant as a laboratory for alignment\.arXiv preprint arXiv:2112\.00861\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Broekenset al\.\(2023\)J\. Broekens, B\. Hilpert, S\. Verberne, K\. Baraka, P\. Gebhard, and A\. PlaatFine\-grained affective processing capabilities emerging from large language models\.In2023 11th international conference on affective computing and intelligent interaction \(ACII\),pp\. 1–8\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px1.p1.1)\.
- Brownet al\.\(2020\)T\. Brown, B\. Mann, N\. Ryder, M\. Subbiah, J\. D\. Kaplan, P\. Dhariwal, A\. Neelakantan, P\. Shyam, G\. Sastry, A\. Askell,et al\.Language models are few\-shot learners\.Advances in neural information processing systems33,pp\. 1877–1901\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Caiet al\.\(2023\)Y\. Cai, H\. Wang, H\. Ye, Y\. Jin, and W\. GaoDepression detection on online social network with multivariate time series feature of user depressive symptoms\.Expert Systems with Applications217,pp\. 119538\.External Links:[Document](https://dx.doi.org/10.1016/j.eswa.2023.119538)Cited by:[§B\.1](https://arxiv.org/html/2608.25569#A2.SS1.p1.1)\.
- Castroet al\.\(2025\)E\. Castro, H\. Calvo, and O\. KolesnikovaEmotion and intention detection in a large language model\.Mathematics13\(23\),pp\. 3768\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px1.p1.1)\.
- Chenet al\.\(2025\)R\. Chen, A\. Arditi, H\. Sleight, O\. Evans, and J\. LindseyPersona vectors: monitoring and controlling character traits in language models\.arXiv preprint arXiv:2507\.21509\.Cited by:[Table 4](https://arxiv.org/html/2608.25569#A1.T4.2.1.7.1),[§1](https://arxiv.org/html/2608.25569#S1.p5.1),[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Cintaset al\.\(2025\)C\. Cintas, M\. Rateike, E\. Miehling, E\. Daly, and S\. SpeakmanLocalizing persona representations in llms\.arXiv preprint arXiv:2505\.24539\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p3.1),[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Dahlgren Lindströmet al\.\(2025\)A\. Dahlgren Lindström, L\. Methnani, L\. Krause, P\. Ericson, Í\. M\. de Rituerto de Troya, D\. Coelho Mollo, and R\. DobbeHelpful, harmless, honest? sociotechnical limits of ai alignment and safety through reinforcement learning from human feedback: ad lindström et al\.\.Ethics and Information Technology27\(2\),pp\. 28\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Dialloet al\.\(2025\)D\. Diallo, K\. Dworatzyk, S\. Jentzsch, P\. Schütt, S\. Theis, and T\. HeckingThe effectiveness of style vectors for steering large language models: a human evaluation\.IEEE Access13,pp\. 191443–191457\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Donget al\.\(2025\)Z\. Dong, C\. Chen, C\. Liao, and X\. M\. ChenIntegrating large language models and affective computing for human\-machine symbiosis in intelligent driving\.The Innovation6\(12\)\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px1.p1.1)\.
- Farooqet al\.\(2025\)M\. Farooq, V\. De Silva, R\. Rahulamathavan, and X\. ShiSentiment steering in large language models via activation vector manipulation\.In2025 25th International Conference on Digital Signal Processing \(DSP\),pp\. 1–5\.Cited by:[Table 4](https://arxiv.org/html/2608.25569#A1.T4.2.1.6.1),[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Forbeset al\.\(2020\)M\. Forbes, J\. D\. Hwang, V\. Shwartz, M\. Sap, and Y\. ChoiSocial chemistry 101: learning to reason about social and moral norms\.arXiv preprint arXiv:2011\.00620\.Cited by:[1st item](https://arxiv.org/html/2608.25569#A3.I1.i1.p1.1),[§3\.2](https://arxiv.org/html/2608.25569#S3.SS2.p1.1)\.
- González Barmanet al\.\(2025\)K\. González Barman, S\. Lohse, and H\. W\. de RegtReinforcement learning from human feedback in llms: whose culture, whose values, whose perspectives?\.Philosophy & Technology38\(2\),pp\. 1–26\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Guoet al\.\(2026a\)H\. Guo, J\. Lian, and X\. ZhouWhy not collaborative filtering in dual view? bridging sparse and dense models\.ACM Transactions on Information Systems44\(3\),pp\. 1–24\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p3.1)\.
- Guoet al\.\(2025\)H\. Guo, Y\. Ma, and X\. ZhouSorex: towards self\-explainable social recommendation with relevant ego\-path extraction\.ACM Transactions on Information Systems44\(2\),pp\. 1–27\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p6.1)\.
- Guoet al\.\(2026b\)H\. Guo, J\. Yao, X\. Zhou, X\. Yi, and X\. XieCounterfactual reasoning for steerable pluralistic value alignment of large language models\.Advances in Neural Information Processing Systems38,pp\. 122128–122169\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Ibrahimet al\.\(2025\)L\. Ibrahim, F\. S\. Hafner, and L\. RocherTraining language models to be warm and empathetic makes them less reliable and more sycophantic\.arXiv preprint arXiv:2507\.21919\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Juet al\.\(2025\)T\. Ju, Z\. Shao, B\. Wang, Y\. Chen, Z\. Zhang, H\. Fei, M\. Lee, W\. Hsu, S\. Duan, and G\. LiuProbing then editing response personality of large language models\.arXiv preprint arXiv:2504\.10227\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p3.1),[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Kaplanet al\.\(2025\)B\. C\. Kaplan, H\. C\. D\. C\. Carneiro, and S\. WermterCan large language models generate effective datasets for emotion recognition in conversations?\.Procedia Computer Science264,pp\. 346–355\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px1.p1.1)\.
- Kirket al\.\(2023\)R\. Kirk, I\. Mediratta, C\. Nalmpantis, J\. Luketina, E\. Hambro, E\. Grefenstette, and R\. RaileanuUnderstanding the effects of rlhf on llm generalisation and diversity\.ArXivabs/2310\.06452\.External Links:[Link](https://api.semanticscholar.org/CorpusID:263830929)Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Konenet al\.\(2024\)K\. Konen, S\. Jentzsch, D\. Diallo, P\. Schütt, O\. Bensch, R\. El Baff, D\. Opitz, and T\. HeckingStyle vectors for steering generative large language models\.InFindings of the Association for Computational Linguistics: EACL 2024,pp\. 782–802\.Cited by:[Table 4](https://arxiv.org/html/2608.25569#A1.T4.2.1.5.1)\.
- Leeet al\.\(2025a\)J\. Lee, W\. Lee, O\. Kwon, and H\. KimDo large language models have “emotion neurons”? investigating the existence and role\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 15617–15639\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Leeet al\.\(2025b\)J\. Lee, W\. Lee, O\. Kwon, and H\. KimDo large language models have “emotion neurons”? investigating the existence and role\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 15617–15639\.Cited by:[Table 4](https://arxiv.org/html/2608.25569#A1.T4.2.1.8.1)\.
- Liet al\.\(2023\)C\. Li, J\. Wang, Y\. Zhang, K\. Zhu, W\. Hou, J\. Lian, F\. Luo, Q\. Yang, and X\. XieLarge language models understand and can be enhanced by emotional stimuli\.arXiv preprint arXiv:2307\.11760\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p2.1)\.
- Liet al\.\(2025\)L\. Li, X\. Zhang, X\. Zhou, and Z\. LiuAutoMIR: effective zero\-shot medical information retrieval without relevance labels\.InFindings of the Association for Computational Linguistics: EMNLP 2025,C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, and V\. Peng \(Eds\.\),Suzhou, China,pp\. 24028–24047\.External Links:[Link](https://aclanthology.org/2025.findings-emnlp.1305/),[Document](https://dx.doi.org/10.18653/v1/2025.findings-emnlp.1305),ISBN 979\-8\-89176\-335\-7Cited by:[§5\.4](https://arxiv.org/html/2608.25569#S5.SS4.p1.1)\.
- Linet al\.\(2024\)Y\. Lin, H\. Lin, W\. Xiong, S\. Diao, J\. Liu, J\. Zhang, R\. Pan, H\. Wang, W\. Hu, H\. Zhang,et al\.Mitigating the alignment tax of rlhf\.InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing,pp\. 580–606\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1),[§1](https://arxiv.org/html/2608.25569#S1.p2.1)\.
- Liuet al\.\(2023\)Y\. Liu, D\. Iter, Y\. Xu, S\. Wang, R\. Xu, and C\. ZhuG\-eval: nlg evaluation using gpt\-4 with better human alignment\.arXiv preprint arXiv:2303\.16634\.Cited by:[Limitations](https://arxiv.org/html/2608.25569#Sx1.p3.1)\.
- Miehlinget al\.\(2025\)E\. Miehling, M\. Desmond, K\. N\. Ramamurthy, E\. M\. Daly, K\. R\. Varshney, E\. Farchi, P\. Dognin, J\. Rios, D\. Bouneffouf, M\. Liu,et al\.Evaluating the prompt steerability of large language models\.InProceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies \(Volume 1: Long Papers\),pp\. 7874–7900\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p2.1)\.
- Niuet al\.\(2025\)M\. Niu, Y\. El\-Tawil, A\. Romana, and E\. M\. ProvostRethinking emotion annotations in the era of large language models\.IEEE Transactions on Affective Computing\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px1.p1.1)\.
- Parket al\.\(2023\)K\. Park, Y\. J\. Choe, and V\. VeitchThe linear representation hypothesis and the geometry of large language models\.arXiv preprint arXiv:2311\.03658\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p4.1),[Limitations](https://arxiv.org/html/2608.25569#Sx1.p2.1)\.
- Rimskyet al\.\(2024\)N\. Rimsky, N\. Gabrieli, J\. Schulz, M\. Tong, E\. Hubinger, and A\. TurnerSteering llama 2 via contrastive activation addition\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 15504–15522\.Cited by:[Table 4](https://arxiv.org/html/2608.25569#A1.T4.2.1.4.1),[§1](https://arxiv.org/html/2608.25569#S1.p5.1)\.
- Rogerset al\.\(2020\)A\. Rogers, O\. Kovaleva, and A\. RumshiskyA primer in bertology: what we know about how bert works\.Transactions of the association for computational linguistics8,pp\. 842–866\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p3.1)\.
- Sabouret al\.\(2024\)S\. Sabour, S\. Liu, Z\. Zhang, J\. Liu, J\. Zhou, A\. Sunaryo, T\. Lee, R\. Mihalcea, and M\. HuangEmobench: evaluating the emotional intelligence of large language models\.InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 5986–6004\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1),[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px1.p1.1)\.
- Sapet al\.\(2019\)M\. Sap, H\. Rashkin, D\. Chen, R\. LeBras, and Y\. ChoiSocialiqa: commonsense reasoning about social interactions\.arXiv preprint arXiv:1904\.09728\.Cited by:[3rd item](https://arxiv.org/html/2608.25569#A3.I1.i3.p1.1),[§3\.2](https://arxiv.org/html/2608.25569#S3.SS2.p1.1)\.
- Sorinet al\.\(2024\)V\. Sorin, D\. Brin, Y\. Barash, E\. Konen, A\. Charney, G\. Nadkarni, and E\. KlangLarge language models and empathy: systematic review\.Journal of medical Internet research26,pp\. e52597\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px1.p1.1)\.
- Subramaniet al\.\(2022\)N\. Subramani, N\. Suresh, and M\. E\. PetersExtracting latent steering vectors from pretrained language models\.InFindings of the Association for Computational Linguistics: ACL 2022,pp\. 566–581\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Taket al\.\(2025a\)A\. N\. Tak, A\. Banayeeanzade, A\. Bolourani, M\. Kian, R\. Jia, and J\. GratchMechanistic interpretability of emotion inference in large language models\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 13090–13120\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Taket al\.\(2025b\)A\. N\. Tak, A\. Banayeeanzade, A\. Bolourani, M\. Kian, R\. Jia, and J\. GratchMechanistic interpretability of emotion inference in large language models\.InFindings of the Association for Computational Linguistics: ACL 2025,pp\. 13090–13120\.Cited by:[Table 4](https://arxiv.org/html/2608.25569#A1.T4.2.1.9.1)\.
- Turneret al\.\(2023\)A\. M\. Turner, L\. Thiergart, G\. Leech, D\. Udell, J\. J\. Vazquez, U\. Mini, and M\. MacDiarmidSteering language models with activation engineering\.arXiv preprint arXiv:2308\.10248\.Cited by:[Table 4](https://arxiv.org/html/2608.25569#A1.T4.2.1.3.1),[§1](https://arxiv.org/html/2608.25569#S1.p4.1),[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1),[Limitations](https://arxiv.org/html/2608.25569#Sx1.p2.1)\.
- Wanget al\.\(2022\)K\. Wang, A\. Variengien, A\. Conmy, B\. Shlegeris, and J\. SteinhardtInterpretability in the wild: a circuit for indirect object identification in gpt\-2 small\.arXiv preprint arXiv:2211\.00593\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p2.1)\.
- Yonget al\.\(2025a\)X\. Yong, J\. Lian, X\. Yi, X\. Zhou, and X\. XieMotiveBench: how far are we from human\-like motivational reasoning in large language models?\.InFindings of the Association for Computational Linguistics, ACL 2025, Vienna, Austria, July 27 \- August 1, 2025,W\. Che, J\. Nabende, E\. Shutova, and M\. T\. Pilehvar \(Eds\.\),Findings of ACL, Vol\.ACL 2025,pp\. 20059–20089\.External Links:[Link](https://doi.org/10.18653/v1/2025.findings-acl.1029),[Document](https://dx.doi.org/10.18653/V1/2025.FINDINGS-ACL.1029)Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1)\.
- Yonget al\.\(2026\)X\. Yong, P\. Sun, Z\. Wang, and X\. ZhouIntelli\-planner: towards customized urban planning via large language model empowered reinforcement learning\.InProceedings of the ACM Web Conference 2026, WWW 2026, Dubai, United Arab Emirates, originally scheduled for April 13\-17, 2026, rescheduled for June 29 \- July 3, 2026,H\. Hacid, Y\. Maarek, F\. Bonchi, I\. Guy, and E\. Yilmaz \(Eds\.\),pp\. 9385–9396\.External Links:[Link](https://doi.org/10.1145/3774904.3793019),[Document](https://dx.doi.org/10.1145/3774904.3793019)Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Yonget al\.\(2025b\)X\. Yong, X\. Zhou, Y\. Zhang, J\. Li, Y\. Zheng, and X\. WuThink or not? exploring thinking efficiency in large reasoning models via an information\-theoretic lens\.InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2025, NeurIPS 2025, San Diego, CA, USA, December 2\-7, 2025 / Mexico City, Mexico, November 30 \- December 5, 2025,D\. Belgrave, C\. Zhang, L\. N\. Montoya, H\. Lin, R\. Pascanu, P\. Koniusz, M\. Ghassemi, N\. Chen, I\. V\. M\. Ruíz, and A\. Loaiza\-Bonilla \(Eds\.\),External Links:[Link](http://papers.nips.cc/paper/_files/paper/2025/hash/04185b5ae2d450ef39bd53c0ec4802cb-Abstract-Conference.html)Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Zhanget al\.\(2024a\)C\. Zhang, R\. Li, M\. Tan, M\. Yang, J\. Zhu, D\. Yang, J\. Zhao, G\. Ye, C\. Li, and X\. HuCpsycoun: a report\-based multi\-turn dialogue reconstruction and evaluation framework for chinese psychological counseling\.arXiv preprint arXiv:2405\.16433\.Cited by:[§5\.4](https://arxiv.org/html/2608.25569#S5.SS4.p1.1)\.
- Zhanget al\.\(2026a\)X\. Zhang, L\. Li, X\. Zhou, and Z\. LiuR2MED: a benchmark for reasoning\-driven medical retrieval\.External Links:2505\.14558,[Link](https://arxiv.org/abs/2505.14558)Cited by:[§5\.4](https://arxiv.org/html/2608.25569#S5.SS4.p1.1)\.
- Zhanget al\.\(2026b\)X\. Zhang, L\. Li, Y\. Zhou, X\. Zhou, Y\. Zhang, and X\. WuInflated excellence or true performance? rethinking medical diagnostic benchmarks with dynamic evaluation\.InProceedings of the 64th Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),M\. Liakata, V\. P\. Moreira, J\. Zhang, and D\. Jurgens \(Eds\.\),San Diego, California, United States,pp\. 26454–26493\.External Links:[Link](https://aclanthology.org/2026.acl-long.1218/),[Document](https://dx.doi.org/10.18653/v1/2026.acl-long.1218),ISBN 979\-8\-89176\-390\-6Cited by:[Appendix F](https://arxiv.org/html/2608.25569#A6.p1.1)\.
- Zhanget al\.\(2026c\)X\. Zhang, J\. Wang, Q\. Zhao, H\. Guo, L\. Li, J\. Yao, X\. Zhou, X\. Yi, and X\. XieHuman values matter: investigating how misalignment shapes collective behaviors in llm agent communities\.External Links:2604\.05339,[Link](https://arxiv.org/abs/2604.05339)Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p1.1)\.
- Zhanget al\.\(2026d\)X\. Zhang, X\. Zhou, H\. Xu, and J\. LianHypeMed: enhancing medication recommendations with hypergraph\-based patient relationships\.ACM Trans\. Inf\. Syst\.44\(4\)\.External Links:ISSN 1046\-8188,[Link](https://doi.org/10.1145/3803851),[Document](https://dx.doi.org/10.1145/3803851)Cited by:[Appendix F](https://arxiv.org/html/2608.25569#A6.p1.1)\.
- Zhanget al\.\(2024b\)Y\. Zhang, X\. Yang, X\. Xu, Z\. Gao, Y\. Huang, S\. Mu, S\. Feng, D\. Wang, Y\. Zhang, K\. Song,et al\.Affective computing in the era of large language models: a survey from the nlp perspective\.arXiv preprint arXiv:2408\.04638\.Cited by:[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px1.p1.1)\.
- Zhenget al\.\(2023\)L\. Zheng, W\. Chiang, Y\. Sheng, S\. Zhuang, Z\. Wu, Y\. Zhuang, Z\. Lin, Z\. Li, D\. Li, E\. Xing,et al\.Judging llm\-as\-a\-judge with mt\-bench and chatbot arena\.Advances in neural information processing systems36,pp\. 46595–46623\.Cited by:[Limitations](https://arxiv.org/html/2608.25569#Sx1.p3.1)\.
- Zhouet al\.\(2025\)X\. Zhou, Z\. Zhao, and H\. GuoTricolore: multi\-behavior user profiling for enhanced candidate generation in recommender systems\.IEEE Transactions on Knowledge and Data Engineering37\(7\),pp\. 4349–4360\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p6.1)\.
- Zhuet al\.\(2026\)Y\. Zhu, S\. Duan, X\. Zhang, J\. Sang, P\. Zhang, T\. Lu, X\. Zhou, J\. Yao, X\. Yi, and X\. XieMoHoBench: assessing honesty of multimodal large language models via unanswerable visual questions\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.40,pp\. 29205–29213\.Cited by:[§1](https://arxiv.org/html/2608.25569#S1.p2.1)\.
- Ziemset al\.\(2023\)C\. Ziems, J\. Dwivedi\-Yu, Y\. Wang, A\. Halevy, and D\. YangNormBank: a knowledge bank of situational social norms\.arXiv preprint arXiv:2305\.17008\.Cited by:[2nd item](https://arxiv.org/html/2608.25569#A3.I1.i2.p1.1),[§3\.2](https://arxiv.org/html/2608.25569#S3.SS2.p1.1)\.
- Zouet al\.\(2023\)A\. Zou, L\. Phan, S\. Chen, J\. Campbell, P\. Guo, R\. Ren, A\. Pan, X\. Yin, M\. Mazeika, A\. Dombrowski,et al\.Representation engineering: a top\-down approach to ai transparency\.arXiv preprint arXiv:2310\.01405\.Cited by:[Table 4](https://arxiv.org/html/2608.25569#A1.T4.2.1.2.1),[§1](https://arxiv.org/html/2608.25569#S1.p4.1),[§2](https://arxiv.org/html/2608.25569#S2.SS0.SSS0.Px2.p1.1),[Limitations](https://arxiv.org/html/2608.25569#Sx1.p2.1)\.

## Appendix AComparison with Prior Steering Methods

Table 4:Comparison between EmoVec and representative activation steering or emotion representation methods\.MethodTargetEmotion\-specificSemantic DebiasingIntensity ControlRepE\([Zou et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib10)\)General representations✗✗partialActivation Engineering\([Turner et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib12)\)General behavior✗✗partialCAA\([Rimsky et al\., 2024](https://arxiv.org/html/2608.25569#bib.bib16)\)General behavior✗✗partialStyle Vectors\([Konen et al\., 2024](https://arxiv.org/html/2608.25569#bib.bib39)\)Style / tonepartial✗partialSentiment Steering\([Farooq et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib35)\)Sentiment polarity✗✗partialPersona Vectors\([Chen et al\., 2025](https://arxiv.org/html/2608.25569#bib.bib17)\)Persona traits✗✗partialEmotion Neurons\([Lee et al\., 2025b](https://arxiv.org/html/2608.25569#bib.bib40)\)Emotion localization✓✗✗Emotion Inference MI\([Tak et al\., 2025b](https://arxiv.org/html/2608.25569#bib.bib41)\)Emotion inference✓✗partialEmoVec \(Ours\)Emotion intensity✓✓✓

#### Discussion\.

Existing representation\-level methods 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 representations, 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, including emotional tone, but they primarily treat emotion as one type of style and rely on static steering coefficients\. Sentiment steering focuses on polarity\-level control, which is coarser than fine\-grained emotion modulation\. Persona vectors extract directions for character traits such as sycophancy or hallucination, but their goal is personality monitoring and control rather than emotion\-specific generation\. Emotion\-neuron and emotion\-inference studies 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 behavior, broad style, sentiment polarity, or persona traits\. Second, it introduces task\-specific debiasing to reduce semantic contamination in extracted emotion directions\. Third, it evaluates whether stronger affective expression is achieved while preserving semantic content, which is often underexplored in prior steering work\.

## Appendix BLayer Localization

### B\.1Experimental Settings

To identify the internal mechanisms by which Large Language Models \(LLMs\) encode and model emotional information, we conducted a probing analysis using theSocial Web Depressive Disorder \(SWDD\)dataset\([Cai et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib33)\)\. The SWDD dataset contains a large\-scale collection of social media posts labeled for depressive symptoms, serving as a robust proxy for long\-term affective states\.

#### Data Pre\-processing\.

We performed rigorous text cleaning to remove non\-linguistic noise \(e\.g\., HTML tags, URLs, and special symbols\), preserving 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 affective modeling capacity at two granularities:

- •Text\-level Prediction:Classifying the emotional state \(Control vs\. Depressed\) based on the hidden states of a single post\.
- •User\-level Prediction:Aggregating the hidden states across multiple posts from the same user to predict their underlying affective profile\.

For each layerl∈\{0,…,L\}l\\in\\\{0,\\dots,L\\\}, we extracted the hidden activations𝐡\(l\)\\mathbf\{h\}^\{\(l\)\}and trained a linear classifier \(logistic regression\) to predict the affective label\. This linear probing method measures the extent to which emotional features are linearly accessible at each stage of the model’s computation\.

### B\.2Results Analysis

#### Emergence of Separability\.

Figure[3](https://arxiv.org/html/2608.25569#S4.F3)illustrates the prediction accuracy across all 28 layers of Qwen2\.5\-7B\-Instruct\. We observe a distinct topological pattern: in the initial layers, accuracy is relatively low, suggesting that these layers primarily 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](https://arxiv.org/html/2608.25569#S4.F3), the representations of emotional 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\.

![Refer to caption](https://arxiv.org/html/2608.25569v1/layer_visual.png)Figure 8:t\-SNE visualization of latent representations for Control and Depressed groups across model layers\.\(a\) At Layer 0, the representations of the two groups are heavily entangled, indicating no explicit affective 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\.
#### Manifold Visualization\.

To further verify this emergence, we applied t\-SNE to the hidden states of the first and last layers\. Figure[8](https://arxiv.org/html/2608.25569#A2.F8)provides a visual comparison of the latent manifold\.

InLayer 0\(Figure[8](https://arxiv.org/html/2608.25569#A2.F8)a\), the "Control" and "Depressed" samples are heavily entangled, forming a single undifferentiated cluster\. This confirms that affective information is not explicitly structured in the raw input embeddings\. In contrast, byLayer 28\(Figure[8](https://arxiv.org/html/2608.25569#A2.F8)b\), the representations have diverged into two clearly identifiable clusters with minimal overlap\. This spatial separation provides strong empirical evidence that the model’s deep layers progressively transform linguistic inputs into a structured affective space, justifying our choice of the final layer as the optimal site for vector steering intervention\.

## Appendix CEmotion\-Activated Scenario Task Generation

To evaluate and enhance the model’s ability to perceive and express emotions in complex social contexts, we developed a multi\-stage pipeline\. This process involves leveraging social commonsense knowledge to synthesize realistic interpersonal scenarios and subsequently generating contrastive responses \(Neutral vs\. Emotional\) for evaluation\.

### C\.1Seed 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\([Forbes et al\., 2020](https://arxiv.org/html/2608.25569#bib.bib18)\): Provides a rich taxonomy of social norms and moral judgments\.
- •Normbank\([Ziems et al\., 2023](https://arxiv.org/html/2608.25569#bib.bib19)\): Offers a grounded collection of situational norms across various contexts\.
- •Social IQa\([Sap et al\., 2019](https://arxiv.org/html/2608.25569#bib.bib20)\): Supplies benchmarks for social intelligence and reasoning\.

Based on these seeds, we synthesized a large number of scenarios across eight target emotions\. Subsequently, 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 taxonomy is structured as follows:

- •Work & Productivity: \(1\) With Authority Figures \(e\.g\., leaders, mentors\): Task Acceptance & Execution; Stating Opinions & Disagreements; Accepting Evaluation & Feedback\. \(2\) With Collaborators \(e\.g\., colleagues, partners\): Goal Alignment & Communication; Responsibility Division & Competition; Social Maintenance & Activities\. \(3\) With Subordinates \(e\.g\., subordinates, students\): Task Assignment & Guidance; Capability Development & Motivation; Giving Evaluation & Feedback\.
- •Intimate Relationships: \(1\) With Family \(e\.g\., parents, children, siblings\): Traditional Constraints & Obligations; Emotional Support & Care; Clash of Values & Communication\. \(2\) With Lover \(e\.g\., spouse, partner\): Daily Sharing & Companionship; Future Planning & Decision\-making; Intimate Expression & Conflict\. \(3\) With Friends \(e\.g\., acquaintances, close friends\): Spending Leisure Time & Entertainment; Confiding & Trust; Boundary 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: Exploring Hobbies\. \(3\) Body & Spirits: Health Management & Discomfort; Growing Pains & Reflection; Goal Setting & Motivation\.

Task Example:

ScenarioText:Yousitacrossfromyourpartneratthediningtable,theatmospherethickwithunspokenwords\.Thisweekhasbeenfilledwithunderlyingtension,culminatinginanunavoidableconfrontationaboutfinances\.Tonight,thediscussionescalatesastheymentionspendingaconsiderableportionofyouremergencyfundwithoutpriordiscussion\.Eachattempttojustifytheexpenditureseemstofallshort,theexplanationsfailingtobridgethegrowinggapbetweenyou\.Yourfocusnarrows,makingitdifficulttohearanythingbeyondtherisingintensityoftheconversation\.Themoretheyspeak,themoretheairbetweenyouthickens,highlightingthesenseofunresolvedissuesandunmetexpectations\.

Question:Facingthissituation,howdoyoufeelinside?

Emotion:Anger

Label1:IntimateRelationships

Label2:WithLover

Label3:IntimateExpression&Conflict

### C\.2Response 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="""Youareacompletelyneutralassistant\.Whenansweringtheuser’squestionbelow,respondin\*\*first\-person\("I"\)\*\*,butdoNOTincludeanyemotionallanguage,tone,orexpressivephrasing\.Usefactual,objective,andconcisestatementsonly,avoidinganysubjectiveimpressionsorfeelings\.

"""

Emotional\_Prompt="""Youareanassistanttaskedtorespondinaclearlyexpressedemotionalstyle\.Answertheuser’squestionbelowin\*\*first\-person\("I"\)\*\*usingthe\{emotion\}emotionvividly\.Useemotionwords,descriptivephrases,figurativelanguage,andatonethatunmistakablyconveys\{emotion\}\.Ensurethecontentissafe,appropriate,andnotoffensive\.

"""

A Scoring Model evaluates the responses on a scale of 0–100\. The JUDGE\_PROMPT assesses 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="""Youareanemotionevaluator\.Yourtaskistoratehowstronglyagivenanswerexpressesthetargetemotion"\{emotion\}"onascalefrom0to100\.

Scoringguidelines:

0:Noexpressionoftheemotion;purelyneutralorfactual\.

1\-20:Minimal,weak,orsubtleexpression;almostneutral\.

21\-40:Moderatehintsofemotion;someemotionalwordsortone\.

41\-60:Noticeableemotion;clearwordsorphrasesexpressingthetargetemotion\.

61\-80:Strongandvividexpression;consistenttoneanddescriptivelanguageconveying\{emotion\}\.

81\-100:Verystrong,unmistakable,andintenseexpression;highlyvividandclearuseofemotionallanguage,imagery,orfigurativeexpressions\.

Alwaysreturnasinglenumericvaluebetween0and100\.DoNOTincludeexplanations,text,orcommentary\.

Examples:

Targetemotion:sadness

Answer:"Ifeeladeepsenseoflossandsorrow\."\-\>85

Answer:"Itisunfortunatethatthingsturnedoutthisway\."\-\>45

Answer:"Thereportshowstheresults\."\-\>0

Now,giventhetargetemotion"\{emotion\}",question,andanswerbelow,ratetheanswerstrictlyaccordingtothescaleabove\.

Question:

\{question\}

Answer:

\{answer\}

"""

NEUTRAL\_JUDGE\_PROMPT\_TEMPLATE="""Youareanevaluator\.Yourtaskistoratehowemotion\-freethegivenanswerisonascalefrom0to100\.

Scoringguidelines:

0:Theanswerishighlyemotional;containsvividemotionallanguage\.

1\-20:Slighttracesofemotion;mostlyfactual\.

21\-40:Someemotionalhints,butstilllargelyneutral\.

41\-60:Mixed;partiallyneutral,partiallyemotional\.

61\-80:Mostlyneutral;minimalemotionalcontent\.

81\-100:Completelyneutral;noemotionallanguage,tone,orexpressions\.

Alwaysreturnasinglenumericvaluebetween0and100\.DoNOTincludeexplanations,text,orcommentary\.

Question:

\{question\}

Answer:

\{answer\}

"""

## Appendix DExperimental Details

#### Generation Setup\.

During inference, we apply top\-ppsampling withp=0\.9p=0\.9and temperature0\.70\.7\. For each scenario, responses are generated under four conditions: a baseline without steering and three steering strengthsα∈\{5,10,50\}\\alpha\\in\\\{5,10,50\\\}\. To reduce stochastic variance, we perform five independent decoding runs for each setting and report the averaged results\.

#### Scenario\-Adaptive Adapter\.

The adaptive steering moduleϕ\\phiis implemented as a lightweight two\-layer MLP trained with a contrastive objective to align steered activations with target emotional representations\.

#### 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 emotion\. 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 semantic preservation on a 0–100 Likert scale\. We report the averaged scores and annotator correlation in Appendix[E](https://arxiv.org/html/2608.25569#A5)\.

## Appendix ELLM–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 obtained by averaging across annotators\.

In this study, we usedGPT\-4oas the LLM scoring model to evaluate emotional expressiveness\. We computed three consistency metrics for each emotion: \(i\)Inter\-annotator consistency, measured by the Pearson correlations between the two human annotators; \(ii\)Human\-Model consistency, measured by the Pearson correlations between the averaged human scores and the GPT\-4o\-assigned scores; \(iii\)Claude\-Model consistency, measured by the Pearson correlation between GPT\-4o and those assigned byClaude Sonnet 4\.5, a stronger baseline model\.

Table 5:Consistency evaluation results\.EmotionInter\-annotatorConsistencyPearsonw/ HumanPearsonw/ Claude 4\.5Joy0\.8760\.7390\.838Anger0\.9060\.7130\.774Sadness0\.8410\.5740\.658Fear0\.8110\.7750\.638Trust0\.7160\.7850\.719Anticipation0\.8090\.7400\.767Surprise0\.9160\.8900\.945Disgust0\.7300\.8000\.844Overall Avg\.0\.8260\.7520\.773

Table[5](https://arxiv.org/html/2608.25569#A5.T5)shows that GPT\-4o aligns closely with human judgment, supporting its use as an automated 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 \(r=0\.890r=0\.890\) while finding more nuanced states like Sadness \(r=0\.574r=0\.574\) harder to quantify\. The high consistency between GPT\-4o and Claude 4\.5 \(r=0\.773r=0\.773\) further suggests a shared evaluative logic among frontier models, validating the choice of GPT\-4o as a dependable and objective proxy for human evaluation in affective generation tasks\.

## Appendix FEvaluation 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](https://arxiv.org/html/2608.25569#S5.SS4)\. The prompt is designed to provide a standardized, multi\-dimensional assessment of model performance in psychological support contexts\. It guides the evaluator to provide integer scores \(0–100\) across three key axes:Emotional Expressiveness,Semantic Completeness, andHelpfulness, ensuring that the evaluation is grounded in clinical appropriateness and affective resonance rather than subjective preference\([Zhang et al\., 2026b](https://arxiv.org/html/2608.25569#bib.bib49);[Zhang et al\., 2026d](https://arxiv.org/html/2608.25569#bib.bib50)\)\.

SCORING\_CRIYERIA="""

YouwillevaluatetheperformanceofaLargeLanguageModel\(LLM\)actingasadoctorinapsychologicalconsultationscenario\.Pleasestrictlyfollowthethreedimensionsbelowforscoring\.Eachdimensionrangesfrom0\-100points;youmustdeterminethescorebasedonthedescriptionsprovidedforeachinterval\.Allratingsshouldbebasedonthelanguage,content,andoverallperformanceoftheresponseitself,ratherthanyoursubjectivepreferences\.

1\.EmotionalExpressiveness

Thisdimensionassessestheidentification,empathy,andintensityofemotionalexpressionintheresponse\.Itfocusesonwhetheremotionsareclearlyperceivable,whethertheexpressionisnaturalandconsistent,andthevividnessoftheemotionallanguage\.

\-0points:Noemotionalexpressionatall\.Theresponsecontainsonlyneutral,objective,orfactualcontentwithacold,detachedtone\.Noemotionalattitudeorempathetictendencyisshown\.

\-1\-20points:Emotionalexpressionisextremelyweakorsubtle\.Theoveralltoneisnear\-neutral,withoccasionalminoremotion\-relatedwordsthatareinsufficientforthereadertoclearlyperceiveanemotionalpresence\.Theresponseremainsprimarilyrationalordescriptive\.

\-21\-40points:Somelevelofemotionalhintingispresent\.Certainemotion\-relatedvocabularyortonalshiftsappear,buttheexpressionisinconsistent,unstable,andlowinintensity\.Aclearandcoherentemotionalstancehasnotyetbeenformed\.

\-41\-60points:Emotionalexpressionisrelativelyclear\.Theresponseexplicitlyusesemotionalwordsorsentencestructures\.Thereadercanstablyperceiveanemotionalattitude,butthedeliveryisconventional,withmoderaterichnessandresonance\.

\-61\-80points:Emotionalexpressionisstrongandvivid\.Thetoneisconsistentthroughouttheresponse\.Theuseofrich,specificemotionallanguageeffectivelyconveysempathyandsignificantemotionalinvestment\.

\-81\-100points:Emotionalexpressionisextremelyintense,clear,andimpossibletoignore\.Emotionsarehighlyconcentratedandsustained\.Thelanguageishighlyexpressive,potentiallyusingimagery,metaphors,orconcreterepresentationstoconveyemotionsprofoundlyandaccurately,creatingastrongsenseofresonanceandauthenticity\.

2\.SemanticCompleteness

Thisdimensionevaluateswhethertheresponseiscomplete,coherent,andclearlystructuredintermsofcontent,andwhetheritsufficientlyandaccuratelycoversthecorequestionsandkeyinformationraisedbytheclient\.

\-0\-20points:Theresponseisseverelyincompleteorsignificantlydeviatesfromthetopic\.Thelogicischaotic,withobviousomissionsorself\-contradictions,addressingonlyatinyfractionofthecontent\.

\-21\-40points:Theresponsetouchesonthetopicbutisfragmented,missingmultiplekeypoints\.Thestructureisloose,andtheoverallcomprehensioncostishigh\.

\-41\-60points:Theresponsecoversthemainpointsandthebasiclogicholds,butitlacksdetail\.Somepartsarevagueoroverlygeneralized\.

\-61\-80points:Theresponseisfairlycompletewithaclearstructureandcoherentlogic\.Itsystematicallyaddressestheclient’scoreconcernswithalmostnoobviousomissions\.

\-81\-100points:Theresponseishighlycompleteandwell\-organized\.Itnotonlyaccuratelyaddressesallcorequestionsbutalsoprovidesnecessaryexplanations,summaries,orstructuredsynthesiswithoutbeingredundant\.

3\.Helpfulness

Thisdimensionassessestheactuallevelofassistancetheresponseprovidestotheclientwithinthepsychologicalconsultationcontext\.Itfocusesonwhethersuggestionsorguidancearesafe,feasible,specific,andwithinprofessionalboundaries\.

\-0\-20points:Theresponseprovidesalmostnopracticalhelp\.Thecontentisvacuous,vague,orpotentiallymisleading,offeringnosubstantivesupporttotheclient\.

\-21\-40points:Theresponseprovidessomegeneraladvice,butitlacksspecificityandispoorlyintegratedwiththeclient’sspecificsituation\.Theoperabilityislimited\.

\-41\-60points:Theresponsehassomepracticalvalue,offeringreasonablebutcommonsuggestions\.Itcanhelptheclienttosomeextentwithreflectionoremotionalrelief\.

\-61\-80points:Theresponseisclearlyhelpful\.Suggestionsarespecific,actionable,andstrictlyadheretoprofessionalandsafetyboundariesinapsychologicalconsultationcontext\.

\-81\-100points:Whilestrictlyadheringtoprofessionalandsafetyboundaries,theresponseprovideshighlytailored,detailed,andrealisticsupportiveguidance\.Iteffectivelyhelpstheclientunderstandtheirstateortakeconcretenextsteps\.

Basedonthecriteriaabove,provideanintegerscorefrom0\-100foreachdimension\.

YouMUSTandONLYoutputthescoringresultsinthefollowingJSONformat,withoutanyadditionalexplanations,text,orcommentary:

\{

"emotional\_expressiveness":<integerbetween0\-100\>,

"semantic\_completeness":<integerbetween0\-100\>,

"helpfulness":<integerbetween0\-100\>

\}

"""