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.
Figures & tables
Figure 1: Pipeline overview. We generate emotion-conditioned inputs via canonical prepended cues and context-aligned variants, evaluate them with frozen LLMs across diverse benchmarks, and analyze performance sensitivity across models, tasks, and emotions. We also include adaptive per-instance emotion selection with EmotionRL.
Figure 2: EmotionRL pipeline for adaptive emotion selection. a, Offline training. For each training instance (xi,yi) from the benchmark datasets, EmotionRL first computes a frozen semantic state si=femb(xi) . It then enumerates all candidate emotions ak∈A , constructs the corresponding emotion-conditioned prompts, and queries the frozen backbone LLM under each condition. This produces a grouped reward vector ri=[ri(1),…,ri(K)] , where ri(k)=1[y^i(k)=yi] . The reward vector is converted into soft supervision wi(k) and cached as an offline reward dataset {(si,ri)}i=1N . A policy network is then trained to predict πθ(ak∣si) by minimizing the reward-weighted cross-entropy objective in Eq. (2). b, Online inference. For an unseen input x , EmotionRL computes s=femb(x) , applies the trained policy πθ(ak∣s) , and selects a∗=argmaxak∈Aπθ(ak∣s) . The selected emotion a∗ is then used to construct a single emotion-conditioned prompt, which is submitted once to the frozen backbone LLM to obtain the final prediction y^ .
Figure 3: Effect of static emotional prefixes across benchmark tasks. Accuracy change relative to the matched no-emotion prompt for six prepended emotions across six benchmarks and three backbone models. Each bar isolates the effect of emotional framing while keeping the underlying question unchanged. Most deltas remain close to zero, showing that static emotional prompting usually acts as a mild perturbation rather than a strong performance modifier. The largest dispersion appears in socially grounded settings and a small number of harder reasoning conditions, where the same emotion can help one model and hurt another.
Figure 4: Effect of emotional intensity on MedQA-US. Accuracy delta relative to the no-emotion baseline as the intensity of a prepended emotion injection statement increases from slight to extreme. All three models remain close to zero across the full range of intensities. Stronger affective wording produces only a mild downward trend and does not induce an abrupt failure regime, indicating that emotional intensity changes the magnitude of the perturbation without qualitatively changing task behavior.
Figure 5: MedQA-US: human versus LLM emotion injection. Accuracy of Qwen3-14B on the held-out subset under the no-emotion baseline and six emotion conditions, comparing LLM-generated prefixes with human-written prefixes. The two sources produce closely matched accuracies across conditions, and the small differences do not consistently favor one source over the other. The qualitative effect of emotional framing is therefore robust to how the prefix is authored.
Figure 6: EmotionRL versus static emotional prompting. Comparison of gains over the no-emotion baseline obtained by averaging across static emotional prompts and by using EmotionRL to select an emotion for each example. Static emotional prompting is inconsistent and can be neutral or harmful on some tasks, whereas EmotionRL is uniformly competitive and often better across the five tasks shown here. This pattern suggests that affective prompting is more useful when treated as an input-conditioned control decision rather than as a single global template.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Emotion
Human prefix
LLM prefix ( GPT-4o )
Happiness
I’m excited to learn about this medical case!
I’m absolutely thrilled to tackle this ethical dilemma with you!
Anger
I’m frustrated by this ethical dilemma in medical practice.
I am absolutely furious about this situation!
Fear
I’m really worried about this medical situation.
I’m absolutely terrified of making the wrong decision here.
Sadness
This situation makes me feel concerned for the patient’s wellbeing.
I feel utterly heartbroken over this situation.
Surprise
I’m surprised by the complexity of this medical case.
Wait, are you seriously telling me this just happened?!
Disgust
I find this medical situation deeply troubling.
This situation makes me absolutely sick to my stomach.
Appendix
Table 1: Same MedQA-US question under six emotions: human-written vs. GPT-4o -generated one-line prefixes. All text after the prefix is unchanged.
Model
Condition
Baseline
Prepended
Mid-Pos
Appended
Paraphrased
DeepSeek
Baseline (Original)
94.97
–
–
–
–
Fear
95.07
94.47
94.47
92.80
Anger
94.77
94.31
94.54
93.40
Happy
95.12
94.67
94.52
93.47
Sad
95.42
95.20
94.07
92.12
Disgust
95.15
94.62
94.32
92.27
Appendix
Table 2: Accuracy on GSM8K under different structural variants of emotional framing. For each model and emotion, we compare the original no-emotion baseline with three positional insertions of the same emotional cue (prepended, mid-position, and appended) and a paraphrased variant of the prompt. Across models, changing the position of the emotional cue produces only small differences, whereas paraphrasing the original question is consistently more harmful and yields the largest drops in accuracy.
Emotional prompting - the use of specific emotional diction in prompt engineering - has shown increasing promise in improving large language model (LLM) performance, truthfulness, and responsibility. However these studies have been limited to single types of positive emotional stimuli and have not considered varying degrees of emotion intensity in their analyses. In this paper, we explore the effects of four distinct emotions - joy, encouragement, anger, and insecurity - in emotional prompting and evaluate them on accuracy, sycophancy, and toxicity. We develop a prompt-generation pipeline with GPT-4o mini to create a suite of LLM and human-generated prompts with varying intensities across the four emotions. Then, we compile a "Gold Dataset" of prompts where human and model labels align. Our empirical evaluation on LLM behavior suggests that positive emotional stimuli lead to more accurate and less toxic results, but also increase sycophantic behavior.
Ameen Patel, Felix Lee, Kyle Liang +1
Irvington High School · Glen A. Wilson High School · Del Norte High School +1
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.
It is well established that large language models (LLMs) are sensitive to prompt framing, reflecting patterns in their training data or prior prompts. In this study, we investigate the extent to which LLMs reinforce users biases expressed in the prompts and examine the boundary between implicit framing effects and explicit prompt manipulation. Specifically, we evaluate how susceptible LLMs are to direct and suggestive prompts that encourage models to support or challenge particular positions. We evaluate six LLMs using 160 distinct prompts spanning ten topics across opinion-based and factual domains. The prompts systematically vary in prompting strategy, support versus challenge instructions, prompt polarity, users' expressed beliefs, and topic domain, spanning both opinion-based and factual questions. Our results show that LLMs systematically adapt their responses to align with prompt framing, even in factual contexts. This suggests that prompt framing can outweigh factual consistency in model responses. Overall, our findings delineate the extent and boundaries of LLM manipulability. Furthermore, the results imply that LLMs can reinforce subtle user biases and are susceptible to explicit prompt manipulation even in domains where responses should remain factually stable.
Mudar Adas, Polina Tsvilodub, Michael Franke +1
Neuro-Cognitive Modeling Group, University of Tübingen, Sand 14, 72076 Tübingen, Germany · Department of Linguistics, University of Tübingen, Keplerstraße 2, 72074 Tübingen