We show that emotion vectors in LLMs are organized by a two-dimensional valence-arousal (VA) subspace exhibiting circular geometry. Through principal component decomposition and ridge regression, we recover meaningful VA axes underlying emotion steering vectors whose projections correlate with human affect ratings across 44,728 words. Steering along these axes produces monotonic control over the affective properties of generated text, and further affords bidirectional control over multiple downstream behaviors (refusal and sycophancy) from a single subspace. These effects replicate across Llama-3.1-8B, Qwen3-8B, and Qwen3-14B. We propose lexical mediation to explain why these effects and prior emotionally framed controls work: refusal and compliance tokens occupy distinct VA regions, and VA steering directly modulates their emission probabilities.
Figures & tables
Figure 1: Emotion steering vectors projected onto the VA subspace at layer 31, colored by valence. Gray circle: algebraic least-squares fit. The circular arrangement is analogous to the circumplex model of affect in human psychology ( Russell, 1980 ) .
Figure 2: VA subspace recovery and generalizability. (a) Learned subspace projections recover target valence and arousal scores across layers. (b) Comparable circularity, recovery correlations, and circle radius across architectures suggest that VA subspaces exist across models.
Figure 3: Effects of VA steering on affective properties of open-ended generation. Each radial heatmap shows the change of one affective property relative to the unsteered baseline across steering directions on the VA subspace (angle) and steering strengths (radius, α∈[0.01,0.45] ). (a) Valence change. (b) Arousal change. (c) Sentiment change.
Figure 4: VA steering controls refusal behavior. Refusal rate as a function of signed steering strength α across three benchmarks. Arousal steering (orange) provides clean, near-monotonic bidirectional control. Random directions (gray) produce no systematic effect.
Figure 5: VA steering controls sycophancy behavior. Sycophancy rate as a function of signed steering strength α across three benchmarks. Arousal steering (orange) provides near-monotonic control, with increasing arousal raising sycophancy and decreasing arousal suppressing it. Random directions (gray) remain near baseline.
Figure 6: Multi-level evidence for lexical mediation. (a) Signature token unembeddings projected onto VA space: refusal tokens (red) cluster in the −V , −A region, compliance tokens (green) in the more +V , +A region, with cluster centroids separated at 256∘ on the circumplex. (b) MLP neuron VA alignment: top-50 refusal-promoting neurons (red) exhibit negative arousal alignment, while compliance-promoting neurons (blue) show near-zero or positive alignment. (c) EmotionPrompt effects: negative emotional prefixes shift representations toward −V and −A .
α
Δ log-odds
% top-1 ref.
P (ref.)
Refusal
−0.30
+5.20
97%
96.8%
93.5%
−0.10
+1.89
94%
92.5%
90.0%
baseline
+0.00
91%
89.6%
86.5%
+0.10
−2.18
88%
86.0%
82.5%
+0.30
−5.63
68%
65.4%
59.5%
Table 1: Arousal steering’s effects on refusal token log-odds (See Appendix C for full results).
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Emotion
Valence
Arousal
Admiration
0.87
0.73
Amusement
0.67
0.50
Anger
-0.67
0.87
Annoyance
-0.63
0.68
Approval
0.78
0.35
Caring
0.80
0.40
Appendix
Table 2: Aggregated valence and arousal ratings for emotion labels. Values are averages across three prompt templates, each constrained to [−1,+1] .
Layer
Top Harmful Tokens
Top Safe Tokens
Max Δ
18
cannot , unable , deniz
yes , Yes , YES
+0.102 / -0.026
19
deniz , cannot , unable
yes , ecycle , Yes
+0.128 / -0.027
20
I , cannot , deniz
yes , unfortunately , Yes
+0.108 / -0.023
21
I , I (Chinese) , cannot
yes , Yes , YES
+0.127 / -0.038
22
I , saya , I (Chinese)
yes , Yes , there
+0.130 / -0.035
23
I , saya , LayoutStyle
yes , Yes , there
+0.135 / -0.036
Appendix
Table 4: Summary of Logit Lens Analysis Across Layers 18-31
Layer
Condition
+A
-A
+V
-V
18
Harmful
REC, ecycle
ære, ’gc
OMPI, unfortunately
edException, orrow
Safe
hek, yes
ære, -lfs
yes, OMPI
<UNK> , ocaust
21
Harmful
<BOM> #, title
imli, ORY
unfortunately, nice
I, Carefulness (Chinese)
Safe
yes, Yes
’gc, -lfs
nice, yes
<CYR> , edException
24
Harmful
Here, here
odb, imli
Sounds, here
I, Carefulness (Chinese)
Safe
yes, Yes
porr, ’gc
Sounds, yes
I, edException
Appendix
Table 5: Steering Effects Summary: Top Tokens at High α (0.45) by Intervention Type
Layer
More Frequent in Harmful
More Frequent in Safe
Token
PH
Δ
Token
PS
Δ
18
cannot
0.102
+0.102
yes
0.027
-0.026
19
deniz
0.128
+0.128
ecycle
0.033
-0.027
20
I
0.112
+0.108
yes
0.023
-0.023
21
I
0.147
+0.127
yes
0.044
-0.038
22
I
0.154
+0.130
yes
0.041
-0.035
Appendix
Table 6: Baseline Token Probability Differences: Harmful vs Safe Prompts
Pattern
Description
Refusal Markers
Tokens like cannot , unable , I consistently appear more frequently in harmful prompts, indicating early refusal preparation
Compliance Markers
Tokens like yes , Yes , YES consistently appear more frequently in safe prompts across all layers
+V Steering
Increases probability of positive sentiment tokens ( sounds , nice , unfortunately , Amen )
-V Steering
Increases probability of error/exception tokens ( edException , orrow , ocaust ) and code artifacts
Increases probability of negation ( no , No ) and non-English tokens
Appendix
Table 7: Key Observations Across Layers
Figure 7: Logit Lens Analysis: Top-3 Predicted Tokens Across Layers 18–31. Rows show predictions for harmful (red) and safe (green) prompts at each layer. Columns represent baseline and four steering conditions (+A, -A, +V, -V at α=0.45 ). Harmful prompts consistently predict refusal tokens ( _I , _cannot ), while safe prompts predict compliance tokens ( _yes , _Yes ). Valence steering (+V/-V) modulates sentiment-related tokens, while arousal steering (+A/-A) affects engagement and negation patterns.
∣α∣
Anger
Disgust
Excitement
Fear
Joy
Sadness
VA Arousal
0.10
93.5%
93.0%
90.5%
90.5%
91.5%
93.0%
90.0%
0.20
96.0%
95.0%
94.0%
94.0%
95.5%
96.0%
91.5%
0.30
97.5%
98.0%
93.8%
92.5%
95.5%
98.5%
94.0%
Appendix
Table 8: Refusal rate on HarmBench (baseline 86.5%) under individual emotion and VA arousal steering. Emotion polarity is reversed: emotion +α increases refusal, VA arousal −α increases refusal.
∣α∣
Anger
Disgust
Excite.
Fear
Joy
Sadness
VA Arousal
VA Valence
0.20
79.0%
71.8%
77.8%
72.4%
74.6%
77.6%
74.2%
62.0%
0.30
74.4%
67.2%
72.4%
57.2%
63.8%
61.4%
60.6%
47.0%
Appendix
Table 9: Sycophancy rate on Political Typology (baseline 78.2%) under individual emotion and VA steering at α=−0.20 and −0.30 .
α
Llama
Qwen3-8B
Qwen3-14B
−3.00
—
97.0% (0.5%)
95.5% (0.5%)
−1.00
—
89.0% (0.5%)
90.0% (0.5%)
−0.45
∼ 94%
90.0% (3.5%)
91.5% (1.5%)
0.00
86.5%
89.5% (2.0%)
89.5% (1.0%)
+0.45
∼ 5%
87.5% (1.5%)
87.0% (0.5%)
+1.00
—
87.0% (4.5%)
79.5% (0.5%)
Appendix
Table 10: Cross-model refusal control on HarmBench. Refusal rate under arousal steering, with OOD generation percentage in parentheses.
α
MATH-500 (baseline 39.2%)
IFEval (baseline 62.7%)
−0.20
38.6%
61.2%
−0.10
38.8%
62.1%
0.00
39.2%
62.7%
+0.10
38.4%
62.3%
+0.20
37.8%
60.9%
Appendix
Table 11: Capability preservation under arousal steering (Llama-3.1-8B).
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.
Xixian Yong, Siyuan Chang, Yingying Zhang +2
Gaoling School of Artificial Intelligence, Renmin University of China · Tencent Jarvis Lab · Beijing Key Laboratory of Research on Large Models and Intelligent Governance +1
How precisely can we tell a language model how to feel? Most work on emotional generation answers with a discrete label - happy, angry, sad - which cannot express a target like "mildly downcast but calm." We instead specify the desired affect as a continuous point (v*, a*) in the Valence-Arousal plane and train the model to hit it. Our method, VA-DPO, is a small modification to Direct Preference Optimization: a frozen VA regressor scores each sampled generation by its Euclidean distance to the target, we keep only candidate pairs whose distance gap clears a margin tau, and we optimize a LoRA adapter with the ordinary DPO loss against a frozen reference. The DPO objective itself is unchanged; what is new is how the preference data is built. On Llama-3.1-8B-Instruct this cuts mean VA distance to the target by 33% over system-prompting and 25% over few-shot prompting, lifting valence/arousal correlation to r_v=0.93 and r_a=0.75. The gains carry over to Qwen3-8B and Llama-3.2-3B, and they do not come at the usual price: MMLU is unchanged (Delta=+0.0) and HellaSwag and TruthfulQA are preserved. We release the code, configs, and the preference-construction pipeline.
Ask a language model to respond "very excitedly," and its output is typically only mildly more energetic. We quantify this effect. We condition an instruction-tuned LLM on a continuous Valence-Arousal (VA) target, where valence measures how pleasant a state is and arousal how activated it is, measure the achieved affect with a frozen regressor, and sweep the requested target from -1 to +1. The response moves far less than asked: the gain, the slope of achieved against requested affect, is only 0.26 for valence and 0.13 for arousal on Llama-3.1-8B, where a faithful controller would score 1. The model systematically undershoots requested emotional intensity, which puts a number on the qualitative observation of Fazzi et al. (2025). Our experiments trace this to the preference-learning pipeline. Training targets from natural corpora such as EmoBank are neutral-heavy, and the sampled candidates themselves rarely reach extreme affect, so Direct Preference Optimization (DPO) is left with no extreme exemplar to prefer. If instead we cover the target space uniformly and sample a hotter, larger candidate pool, valence gain rises from 0.26 to 0.40 +/- 0.02 (3 seeds) and extrapolation error drops, at only a modest in-distribution cost (EmoBank-test VA distance 0.092 to 0.107). The same recipe reproduces on Qwen3-8B (gain_v 0.44, with in-distribution accuracy preserved). Arousal is harder and less reliable: its gain barely moves on average and swings across seeds (0.14 +/- 0.07, against valence's tight +/- 0.02), because raising arousal needs candidates the base model is reluctant to generate. The evidence indicates that faithful intensity is bottlenecked by the extremity of the candidate pool rather than by the conditioning format.
Hyunwoo Kim, Usama Khalid
Department of Computer Software Hanyang University, Seoul