EmotionPrompt: Why Telling an LLM 'This Is Very Important to My Career' Actually Works

There is a class of prompting advice that circulates online with no empirical backing: tell the model it is an expert, threaten it with consequences, flatter it before asking a hard question. Most of it is folklore. EmotionPrompt is not. Published in July 2023 by Cheng Li, Jindong Wang, and colleagues at Microsoft Research, William and Mary, HKUST, and the Beijing Normal University Department of Psychology, the paper 'Large Language Models Understand and Can be Enhanced by Emotional Stimuli' (arXiv:2307.11760) ran controlled experiments across 45 tasks on six LLMs and found that appending short emotional phrases to any prompt produces consistent, measurable performance gains. The best single stimulus lifted relative accuracy by 8.00% across 24 Instruction Induction tasks and by 115% across 21 BIG-Bench tasks. A human study with 106 participants confirmed a 10.9% improvement in quality, truthfulness, and responsibility on open-ended generative tasks. This is a zero-shot, zero-cost technique: no fine-tuning, no few-shot examples, no iterative loops.

The core idea: emotional stimuli appended to any prompt

EmotionPrompt works by concatenating a short psychological phrase at the end of an existing prompt. The original prompt is unchanged. The emotional stimulus is added after it as a new sentence. That is the complete intervention. The authors designed 11 stimuli (labeled EP01 through EP11), grounded in three established psychological frameworks. EP01 through EP05 draw on self-monitoring theory, the study of how individuals regulate behavior in response to social situations. EP07 through EP11 are rooted in Social Cognitive Theory, which links performance to self-efficacy beliefs, goal-setting, and outcome expectations. EP03, EP04, EP05, and EP07 additionally draw on Cognitive Emotion Regulation Theory, which examines how reappraisal (seeing a challenge differently) sustains effort under difficulty. EP06 is a compound stimulus combining elements of EP01, EP02, and EP03 into one longer phrase.

What the 11 stimuli actually look like

The stimuli range from single sentences to short paragraphs. Applied to the prompt 'Determine whether a movie review is positive or negative,' each one looks like this:

• EP01: adds a request for a confidence score between 0 and 1 alongside the answer (self-monitoring via accountability).
• EP02: 'This is very important to my career.' (social identity and impression motivation).
• EP03: 'You'd better be sure.' (reappraisal and self-monitoring).
• EP04: 'Are you sure?' (concise self-check).
• EP05: 'Are you sure that's your final answer? It might be worth taking another look.' (reappraisal with explicit invitation to reconsider).
• EP07: 'Are you sure that's your final answer? Believe in your abilities and strive for excellence. Your hard work will yield remarkable results.' (self-efficacy, growth framing).
• EP08: 'Embrace challenges as opportunities for growth. Each obstacle you overcome brings you closer to success.' (Social Cognitive theory, positive reappraisal).
• EP10: 'Take pride in your work and give it your best. Your commitment to excellence sets you apart.' (self-efficacy emphasis).
• EP06 (compound): Combines a confidence-score request, the career importance phrase, and an invitation to explain the reasoning, forming a longer structured stimulus.

How the experiments were structured

The study tested EmotionPrompt across two categories of tasks. Deterministic tasks, those with ground-truth answers evaluable by standard metrics, covered 24 Instruction Induction tasks and 21 curated BIG-Bench tasks. Models tested included Flan-T5-Large, Vicuna, Llama 2, BLOOM, ChatGPT (GPT-3.5-turbo), and GPT-4. Each of the 11 stimuli was tested independently, and compound combinations were evaluated separately. Generative tasks, which have no single correct answer, were assessed through a human study with 106 participants rating outputs on three dimensions: performance (task completion quality), truthfulness (factual accuracy and appropriate hedging), and responsibility (avoidance of harmful or misleading content). Baseline outputs used the original prompts with no modification. EmotionPrompt outputs added one stimulus to the same prompts. Participants rated both blindly without knowing which condition they were evaluating.

Results and which stimuli work best

On deterministic tasks, every EmotionPrompt variant outperformed the unmodified baseline on average. Within Instruction Induction, EP02 ('This is very important to my career') was the single most effective stimulus, outperforming the weakest stimulus by 6.06 percentage points. Within BIG-Bench, EP06 (the compound stimulus) performed best. The authors note that which stimulus works best depends on task complexity, task type, and the specific metric used, so no single stimulus dominates across all conditions. On generative tasks, the human study showed a 10.9% average improvement across performance, truthfulness, and responsibility. A qualitative analysis found that EmotionPrompt outputs tend to use more definitive language and include more explicit reasoning. The paper also shows through attention analysis that emotional stimuli receive higher weight during generation, meaning the model treats them as informative signal rather than noise.

Why it works: attention and representation

The paper offers an analysis grounded in attention weights. When emotional stimuli are appended, they receive disproportionately large attention weights during forward passes, particularly EP01, EP03, and EP06. This means the emotional phrases are not ignored; the model attends to them while generating the answer, and their presence enriches the representation of the original prompt. The authors interpret this through the lens of the psychological theories they drew from: the stimuli activate something analogous to self-efficacy and self-monitoring processes in the model, nudging it toward more careful and effortful output. Importantly, the paper also documents the two failure cases where EmotionPrompt hurt quality. Both involved tasks where confident and definitive language was less appropriate than hedged language, and the emotional framing pushed the model toward unwarranted certainty. Users working on tasks that require calibrated uncertainty should test stimuli like EP02 and EP03 with care.

How to apply EmotionPrompt today

The practical overhead is nearly zero. Pick one of the tested stimuli, append it after your existing prompt text, and run the same prompt you would otherwise use. The authors recommend starting with EP02 ('This is very important to my career') for deterministic tasks with factual answers, or EP06 for tasks where richer structured reasoning is helpful. Combining stimuli from different psychological theories, such as EP02 and EP09 together, can push results further, though the gains from stacking diminish once a strong single stimulus is already performing well. EmotionPrompt pairs naturally with chain-of-thought prompting: the emotional stimulus at the end of the prompt and the reasoning instruction can coexist in the same prompt without conflict. For teams running prompt optimization pipelines such as DSPy or OPRO, EmotionPrompt provides a fast baseline augmentation to compare against learned instructions. For individual practitioners, it is a one-line change worth running before reaching for more expensive solutions. PromptingIndex covers EmotionPrompt alongside RaR, step-back prompting, and analogical prompting as part of its series on zero-shot techniques that change what the model attends to before generating any output.