Emotion is Not Just a Label: Latent Emotional Factors in LLM Processing

Benjamin Reichman; Adar Avsian; Samuel Webster; Larry Heck

cs.CLMar 10, 2026

Emotion is Not Just a Label: Latent Emotional Factors in LLM Processing

Authors: Benjamin Reichman, Adar Avsian, Samuel Webster, Larry Heck

Organizations: Georgia Institute of Technology

Abstract

Large language models are routinely deployed on text that varies widely in emotional tone, yet their reasoning behavior is typically evaluated without accounting for emotion as a source of representational variation. Prior work has largely treated emotion as a prediction target, for example in sentiment analysis or emotion classification. In contrast, we study emotion as a latent factor that shapes how models attend to and reason over text. We analyze how emotional tone systematically alters attention geometry in transformer models, showing that metrics such as locality, center-of-mass distance, and entropy vary across emotions and correlate with downstream question-answering performance. To facilitate controlled study of these effects, we introduce Affect-Uniform ReAding QA (AURA-QA), a question-answering dataset with emotionally balanced, human-authored context passages. Finally, an emotional regularization framework is proposed that constrains emotion-conditioned representational drift during training. Experiments across multiple QA benchmarks demonstrate that this approach improves reading comprehension in both emotionally-varying and non-emotionally varying datasets, yielding consistent gains under distribution shift and in-domain improvements on several benchmarks.

Figures & tables

Figure 1: Distribution of emotions for a web corpus ( Penedo et al., 2023 ) .

Figure 2: Emotion Distribution for TweetQA and FriendsQA.

Feature	Definition	Interpretation
Spatial Structure Features
Center-of-Mass Distance (CMD)	$\displaystyle\frac{1}{H\|\mathcal{Q}\|}\sum_{h=1}^{H}\sum_{i\in\mathcal{Q}}\left\|i-\frac{\sum_{j}jA_{h,i,j}}{\sum_{j}A_{h,i,j}}\right\|$	Measures how far attention extends from each query token; higher values indicate more non-local attention.
Tail Mass	$\displaystyle\frac{1}{\|\mathcal{Q}\|}\sum_{h=1}^{H}\sum_{i\in\mathcal{Q}}\sum_{j}A_{h,i,j}\,\mathbb{I}[\|i-j\|>d_{0}]$	Quantifies reliance on long-range context by measuring attention mass beyond a fixed distance.
Locality	$\displaystyle\frac{1}{\sum_{i}m_{i}}\sum_{i,j}A_{h,i,j}m_{i}m_{j}\|i-j\|$	Captures the expected spatial spread of attention; smaller values indicate tightly localized focus.
Distributional Sharpness Features
Key Entropy (KE)	$\displaystyle-\frac{1}{H}\sum_{h=1}^{H}\sum_{j}\bar{A}_{h,j}\log\bar{A}_{h,j}$	Measures how uniformly attention mass is distributed across keys at the layer level.

Table 2: Summary of attention geometry features grouped by functional category.

Emotion	Results
	Zero-shot	Trained
Neutral	48%	58%
Happy	36%	45%
Sad	34%	49%
Anger	31%	42%
Fear	34%	47%

Figure 3: Disparity in performance across emotions for LLaMA-3.1-8B on AURA-QA.

Figure 4: Emotion-specific differences in attention geometry across features. The heatmap reports one-vs-rest Cohen’s d effect sizes, comparing each emotion to all others for each attention feature. Colors indicate the direction and magnitude of deviation relative to the global distribution. Features are ordered by across-emotion variance, highlighting attention dimensions most sensitive to emotional tone. Sarcasm is omitted for clarity due to its extreme divergence.

Figure 5: Differences in the attentional pattern between emotions. For each query token, attention differences across keys are standardized by subtracting the row mean and dividing by the row standard deviation, highlighting relative redistribution of attention rather than absolute magnitude changes.

Bloom Level 2 (Understand)

Passage excerpt: “Raw meat … does not appeal … the idea of using it for food is disgusting.”

Question: What do civilized persons find disgusting?

Answer: Raw meat

Bloom Level 3 (Apply)

Passage excerpt: “… the stony nature of the plains would soon disable an unshod horse …”

Question: Why can’t unshod horses chase guanacos?

Table 1: Representative examples of Bloom Level 2 (comprehension) and Level 3 (application) questions.

	Filtered Out					Retained
Emotion	Valid	Answerable	Grounded	Non-trivial	Answer Match	Valid	Answerable	Grounded	Non-trivial	Answer Match
Anger	0.822	0.836	0.986	0.973	0.521	0.897	0.897	1.000	1.000	0.368
Disgust	0.880	0.893	1.000	0.987	0.627	0.958	0.972	1.000	0.986	0.408
Excitement	0.887	0.930	1.000	0.958	0.704	0.901	0.915	0.986	0.986	0.465
Fear	0.900	0.914	1.000	0.986	0.571	0.809	0.824	0.985	0.985	0.485
Happy	0.817	0.845	1.000	0.972	0.690	0.855	0.870	1.000	0.986	0.420

Table 2: Human evaluation of question–answer quality by emotion. Valid denotes majority agreement that a question is answerable, grounded, and non-trivial. Answer Match measures majority exact (token-set) match between the provided answer and human answers. Grounding rates are near ceiling across both subsets, indicating that the dual-model filter primarily modulates difficulty rather than question validity.

Emotion	#Q	Passage Length	Question Length	Bloom L2 (%)			Bloom L3 (%)
		(mean words)	(mean words)	Gemma	LLaMA	Qwen	Gemma	LLaMA	Qwen
Anger	1600	48.5	7.9	31.4	35.9	32.8	31.0	34.2	34.8
Disgust	1600	44.3	8.1	30.2	37.9	31.9	27.1	36.5	36.4
Excitement	1600	59.4	7.6	29.6	35.9	34.5	27.5	36.2	36.2
Fear	1600	57.4	7.3	30.5	38.6	30.9	25.2	37.5	37.2
Happy	1600	77.3	7.8	30.5	38.1	31.4	26.1	37.0	36.9

Table 3: Dataset statistics by emotion category. We report the number of questions, mean passage length, and mean question length (in words). For all emotions, questions are evenly split between Bloom level 2 and Bloom level 3 (50% each). Columns report the percentage of questions contributed by each LLM within each Bloom level.

Figure 6: Overview of the proposed emotional regularization framework.

LoRA Fine-Tuning
	Natural Questions			TweetQA			FriendsQA			AURA-QA
Test Set	Llama	Ministral	Olmov2	Llama	Ministral	Olmov2	Llama	Ministral	Olmov2	Llama	Ministral	Olmov2
Natural Questions	61.6%	60.0%	61.1%	56.7%	56.5%	56.1%	58.3%	59.9%	59.4%	51.3%	54.5%	52.3%
TweetQA	59.5%	56.4%	58.6%	73.1%	74.9%	72.7%	69.0%	68.4%	70.1%	64.1%	65.3%	62.1%
FriendsQA	46.4%	46.1%	51.1%	46.4%	48.6%	46.8%	54.4%	38.3%	67.3%	37.1%	40.6%	37.1%
AURA-QA	35.2%	35.4%	36.2%	42.8%	43.4%	43.5%	40.4%	47.0%	44.1%	49.9%	51.7%	47.4%

Table 4: LoRA training and emotional regularization results when trained and tested across multiple datasets.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Anger	Happy	Sad	Fear	Disgust	Surprise	Neutral	Sarcastic	Excitement	Average
With Sarcasm
ModernBERT-base	0.75/0.78	0.89/0.89	0.85/0.85	0.76/0.87	0.61/0.83	0.65/0.79	0.84/0.80	1.00/0.99	0.59/0.88	0.77/0.96
ModernBERT-large	0.71/0.77	0.89/0.91	0.82/0.85	0.77/0.89	0.60/0.81	0.60/0.67	0.80/0.70	1.00/1.00	0.54/ 1.00	0.75/0.96
T5-base	0.66/0.70	0.86/0.93	0.80/0.82	0.76/ 0.91	0.61/0.85	0.63/0.70	0.79/0.73	0.99/0.99	0.60/0.82	0.75/0.96
T5-large	0.68/0.77	0.88/0.92	0.80/0.82	0.78/0.86	0.62/0.74	0.64/0.65	0.80/0.72	1.00/0.99	0.59/0.86	0.75/0.96
Without Sarcasm

Appendix

Table 5: Emotion classification results per model. Each cell shows F1 / accuracy.

Figure 7: Differences in attention features by emotion. Sarcasm is included.

Figure 8: Sensitivity of passage segmentation to the emotion margin threshold.

Emotion	LLM–Human	Human–Human	LLM–LLM
Joy	0.693	0.687	0.900
Sadness	0.533	0.807	0.787
Anger	0.587	0.713	0.860
Fear	0.553	0.727	0.700
Disgust	0.540	0.767	0.853
Surprise	0.573	0.660	0.867

Appendix

Table 6: Inter-rater agreement for passage-level emotion verification. Human–Human and LLM–LLM denote unanimous agreement (3/3). LLM–Human compares majority votes (three LLMs vs. three humans).

Criterion	Human–Human Unanimity
Answerable	0.71
Grounded	0.93
Non-trivial	0.79

Appendix

Table 7: Human–Human agreement for QA quality criteria.

Figure 9: Human annotation interface for passage-level emotion verification. Annotators judged whether the target emotion was expressed as the dominant affective tone of the passage using a binary decision.

Figure 10: Human annotation interface for question–answer validation. Annotators evaluated answerability, grounding, and non-triviality of questions with respect to the passage.

Figure 11: Ablation across emotion pairs.

Figure 12: Relationship between emotion-pairs used for regularization and performance.

Explore similar work

Do Emotions in Prompts Matter? Effects of Emotional Framing on Large Language Models
Apr 2, 2026Minda Zhao, Yutong Yang, Chufei Peng +5LLM Prompting Framing Effects in Language Models
Harvard University. · Bryn Mawr College.
Some Emotions Run Deeper: Layer-wise Probing and Causal Intervention in Large Language Models
Sep 1, 2026Tian Fang, Gaël Guibon, Davide BuscaldiCausal Interventions in Language Models Emotion Representation in Language Models
1Université Sorbonne Paris Nord, CNRS, Laboratoire d’Informatique de Paris Nord, LIPN · 2LORIA, CNRS, Université de Lorraine
TEMPER: Testing Emotional Perturbation in Quantitative Reasoning
Apr 9, 2026Atahan Dokme, Benjamin Reichman, Larry HeckFraming Effects in Language Models LLM Evaluation
Georgia Institute of Technology

Apr 2, 2026J/K move · Enter open · S save

Do Emotions in Prompts Matter? Effects of Emotional Framing on Large Language Models

Minda Zhao, Yutong Yang, Chufei Peng, Rachel Gonsalves+4

Harvard University. · Bryn Mawr College.

LLM Prompting Framing Effects in Language Models LLM Evaluation

Open paper AskPDF arXiv

Emotional tone is pervasive in human communication, yet its influence on large language model (LLM) behaviour remains unclear. Here, we examine how first-person emotional framing in user-side queries affect LLM performance across six benchmark domains, including mathematical reasoning, medical question answering, reading comprehension, commonsense reasoning and social inference. Across models and tasks, static emotional prefixes usually produce only small changes in accuracy, suggesting that affective phrasing is typically a mild perturbation rather than a reliable general-purpose intervention. This stability is not uniform: effects are more variable in socially grounded tasks, where emotional context more plausibly interacts with interpersonal reasoning. Additional analyses show that stronger emotional wording induces only modest extra change, and that human-written prefixes reproduce the same qualitative pattern as LLM-generated ones. We then introduce EmotionRL, an adaptive emotional prompting framework that selects emotional framing adaptively for each query. Although no single emotion is consistently beneficial, adaptive selection yields more reliable gains than fixed emotional prompting. Together, these findings show that emotional tone is neither a dominant driver of LLM performance nor irrelevant noise, but a weak and input-dependent signal that can be exploited through adaptive control.

Abstract

Explore similar work

Do Emotions in Prompts Matter? Effects of Emotional Framing on Large Language Models

Some Emotions Run Deeper: Layer-wise Probing and Causal Intervention in Large Language Models

TEMPER: Testing Emotional Perturbation in Quantitative Reasoning