cs.CLOct 15, 2026

Identifying Effective Strategies in Prompt Evolution for Stress and Emotion Prediction

Authors: SeungHyun Yi, Seungjun Yi, Sriram Sattiraju, Haeun Park, Joakim Nguyen, Terence Lim, Joseph Skrovan, Mehak Beri, 

Organizations: Department of Computer Science, Kyung Hee University, Yongin, South Korea · Department of Biomedical Engineering, University of Texas at Austin, Austin, TX, USA · Department of Computer Science, University of Texas at Austin, Austin, TX, USA · College of Fine Arts, University of Texas at Austin, Austin, TX, USA · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Early stress detection and risk prediction are crucial for timely mental health assessment and support. While large language models have recently shown promise for social media stress prediction, relying on accuracy can obscure false positives. Also, prompt revisions yield inconsistent gains across models and tasks, with unclear links to strategy combinations and instruction changes. By comparing six methods on social media stress and emotion data across two model sizes, we found that GEPA and its extension combining instructions from two prompts (GEPA+merge) improved prediction over initial prompts in several settings. We also identified five prompting strategies associated with development gains. These strategies included directing attention to emotional evidence, distinguishing current from past distress, illustrating indirect cues, requiring emotional context, and guiding label selection. Moreover, we documented four observations across tasks and evaluation settings. In a separate set of searches, binary stress prediction improved early, whereas multilabel emotion prediction initially declined and partly recovered. The other three concerned incorrect or conflicting instructions, gains on small comparison samples that did not hold on full development data, and changes in selected prompts' development and reported test rankings. These findings identify effective prompt evolution methods, strategy combinations associated with gains, and task and evaluation limits, providing guidance for applying prompt evolution methods to social media stress detection.

Figures & tables

Fig. 1: Prompt evolution through revision and merging. Main study procedures. (a) GEPA revises one instruction and retains a revision only if its score improves on the same three training examples. (b) GEPA+merge also combines instructions from two parents, accepting a score at least as high as the better parent’s on five development examples. Both evaluate accepted prompts on all 100 development examples and select the highest scoring prompt at search end.
StatisticDreaddit [ 8 ]GoEmotions [ 9 ]
Text unitFive sentencesOne comment
Mean words (SD)85.6 (32.1)12.8 (6.7)
Median words8012
Annotated items3,55358,009
Released classification subset3,55354,263
Available labels228
TABLE I: Dataset Statistics
Fig. 2: Label prediction performance across six methods. Section III-B defines all six methods. Groups pair dataset (D: Dreaddit, G: GoEmotions) with Qwen3 model size (8B or 32B). Scores use 200 attempted test items per seed (42 to 46). Bars show means and whiskers one sample standard deviation ( n=5n=5 , except asterisked G32 ICL Conformability, n=4n=4 ). Axes are zoomed, with breaks marking nonzero starts. Compare gaps only within each metric.
Fig. 3: Historical GEPA trees, mean scores, and golden routes. Qwen3-8B GEPA trees for Dreaddit seed 42 (a) and GoEmotions seed 44 (b) contain 20 and 26 candidates, respectively. Circles show identifiers above development Macro-F1 (a) or Micro-F1 (b), in percent. Arrows link parents to children, amber marks golden routes, cyan circles with double outlines mark selected prompts, and grey circles mark other candidates. N/A denotes missing scores. Positions are adjusted for readability. Blue F1 and orange ACC curves show means across seeds 42 to 44 with one sample SD. At each tree level, nn counts scored prompts across seeds in F1/ACC order, while SD is across seed means. Open mean markers indicate one seed without SD. Amber curves show golden route F1 scores without SD. Both plots align vertically with their trees by tree level. Panels use different vertical scales. Candidate Conformability is unavailable. Section III-D explains the averages and missing scores.
Fig. 4: Prompt changes along historical GEPA golden routes. Panels match Figure 3 ’s Qwen3-8B trees. Roots are complete and child excerpts retain source wording. Shading marks new or rewritten wording relative to the full parent, not new strategy types. Arrows link parents and children. Scores are development Macro-F1 (a) and Micro-F1 (b), in percent. Parentheses give changes in percentage points from unrounded parent scores. The dagger flags a cue association for scrutiny, not incorrect label identifiers [ 26 ] . Section III-D explains the comparison limits.
StrategyExamples from promptsDevelopment scores (%)
Directing Attention to Emotional Evidence Dreaddit, seed 42Candidate 1: “the label is based on the author’s emotional state, not on the topic” Candidate 10: “Stress is primarily linked to the emotional state of the author, not the external situation itself.”Candidate 1: 76.14 Candidate 10: 84.96 Separate branches Candidate 10 selected
Distinguish current from past distress Dreaddit, seed 42“stress is determined by the current emotional state and expression in the text, not the past events themselves.”Candidates 0→20\rightarrow 2 66.56→79.8066.56\rightarrow 79.80
Illustrate indirect emotional cues Dreaddit, seed 42“self-deprecation”, “feeling trapped”, and “feeling like they ‘deserve’ negative treatment”Candidates 2→92\rightarrow 9 79.80→81.8879.80\rightarrow 81.88
Require emotional context Dreaddit, seed 42“Physical symptoms or medical concerns are not sufficient on their own to label the text as stress, unless they are tied to emotional strain or anxiety.”Candidates 9→109\rightarrow 10 81.88→84.9681.88\rightarrow 84.96 Child 18: 78.52 (Same rule retained)
Guide emotion label selection GoEmotionsNeutral: “should only be selected if the text explicitly reflects no strong emotional content.” 1. “extract all emotionally charged words or phrases.” 2. “Cross-reference these phrases with the 28-label inventory to identify all possible matching labels.”Seed 43, candidates 0→30\rightarrow 3 31.62→24.1831.62\rightarrow 24.18 Seed 44, candidates 3→153\rightarrow 15 29.37→31.9129.37\rightarrow 31.91
TABLE II: Effective Strategies in Prompt Evolution: Supporting Examples
More in this topic

LLM Prompting

Large language model prompting studies how instructions, examples, context, and formats steer model behavior at inference time.

See all 429 papers

Explore similar work

CardsList
Oct 9, 2026

MoPrO-SD: Modular Prompt Optimization for Stance Detection

Stance detection remains challenging for large language models (LLMs) due to implicit opinions, sarcasm, and context-dependent expressions of stance. Although automatic prompt optimization offers an alternative to fine-tuning without updating model parameters, existing structured approaches are not specifically designed around stance-label semantics and class-specific ambiguities. We introduce MoPrO-SD, a stance-specific prompt optimization framework that decomposes prompts into functional semantic modules and refines them sequentially using LLM-based feedback. We study three refinement strategies: instruction-only optimization (I), instruction with reasoning (I+R), and instruction with per-example reasoning (I+R+E), using class-conditioned errors for stance-label refinement. We also introduce VaccineStance, a new human-annotated benchmark for vaccine stance detection. Experiments across four datasets and five open-weight LLMs show that MoPrO-SD improves upon the human-written Base prompt in many settings. On SemEval-2016, class-conditioned feedback achieves higher mean Macro-F1 than pooled feedback for four of five backbones. Both reasoning-based strategies, I+R and I+R+E, achieve strong performance, with their relative effectiveness varying across models and datasets, while remaining competitive with recent prompt optimization methods. The MoPrO-SD code and VaccineStance dataset are available on GitHub.

Meysam Sabbaghan, Nikesh Gyawali, Arman Zareian Jahromi +3

Kansas State University · University of Illinois Chicago

Stance DetectionPrompt OptimizationLarge Language Model-Guided Optimization
AskPDF
Apr 7, 2026

The Role of Emotional Stimuli and Intensity in Shaping Large Language Model Behavior

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

LLM PromptingLLM SafetyLLM Sycophancy
AskPDF
Mar 26, 2026

Navigating the Prompt Space: Improving LLM Classification of Social Science Texts Through Prompt Engineering

Recent developments in text classification using Large Language Models (LLMs) in the social sciences suggest that costs can be cut significantly, while performance can sometimes rival existing computational methods. However, with a wide variance in performance in current tests, we move to the question of how to maximize performance. In this paper, we focus on prompt context as a possible avenue for increasing accuracy by systematically varying three aspects of prompt engineering: label descriptions, instructional nudges, and few shot examples. Across two different examples, our tests illustrate that a minimal increase in prompt context yields the highest increase in performance, while further increases in context only tend to yield marginal performance increases thereafter. Alarmingly, increasing prompt context sometimes decreases accuracy. Furthermore, our tests suggest substantial heterogeneity across models, tasks, and batch size, underlining the need for individual validation of each LLM coding task rather than reliance on general rules.

Erkan Gunes, Christoffer Florczak, Tevfik Murat Yildirim

Constructor University · Aalborg University · University of Stavanger

Large Language Model-Based Text ClassificationLLM PromptingText ClassificationLLM EvaluationFew-Shot Prompting
AskPDF