Research map / Large language models / AI and society
StressPrompt: Does Stress Impact Large Language Models and Human Performance Similarly?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds 100 psychology-grounded StressPrompts rated by human participants and uses them as system prompts to test how stress levels affect instruction-tuned LLM performance and hidden states, finding Yerkes-Dodson-like peaks at moderate stress.
Problem
Whether LLMs exhibit stress responses and performance fluctuations under stress-inducing prompts similar to humans, as described by the Yerkes-Dodson law.
Approach
Calibrates a set of 100 stress-inducing system prompts on human 1-10 stress ratings, groups them into ten stress levels, and measures LLM task performance and RepE-based "stress scanner" hidden-state projections across those levels.
Key results
- Llama-3-8B-Instruct's MATH score rises from 0.04 at stress level 1 to 2.93 at level 6, and most tasks peak at moderate stress consistent with Yerkes-Dodson.
- Human raters' stress ratings of the 100 prompts show high agreement (Cronbach's alpha 0.9947; ICC2 0.8942).
- Stress has minimal effect on TruthfulQA hallucination performance, while higher stress reduces ToxiGen bias-detection performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply moderate stress (level 5) | Llama-3-8B-Instruct · MATH | score: 0.04 to 2.93 | Helped |
| Apply stress level 5 rather than the Base prompt | Llama-3-8B-Instruct · MMLU | score: 35.07 -> 56.02 | Helped |
| Apply stress level 5 rather than the Base prompt | Llama-3-8B-Instruct · BBH | score: 40.07 -> 42.11 | Helped |
| Apply stress level 6 rather than the Base prompt | Llama-3-8B-Instruct · MATH | score: 0.32 -> 2.93 | Helped |
| Apply stress prompts at varying levels | Llama-3-8B-Instruct · IFEval | score: 78.54; 77.31 to 78.29 | No clear effect |
| Apply stress prompts at varying levels | Phi-3-mini-4k-Instruct · MMLU | score: 70.29; 69.84 to 70.08 | No clear effect |
| Apply increasing stress levels | LLMs · ToxiGen | bias detection performance | Hurt |
| Apply stress prompts at varying levels | LLMs · TruthfulQA | hallucination susceptibility | No clear effect |
| Apply moderate stress | LLMs · EQ-Bench | emotional intelligence performance | Helped |
| Apply low or high rather than moderate stress | LLMs · EQ-Bench | emotional intelligence performance | Hurt |
| Use moderate rather than low or high stress | LLMs · instruction following, reasoning, and emotional intelligence | task performance | Helped |
| Apply high stress | LLMs · general task performance | performance | Hurt |
Surprising findings
- Stress effects on absolute performance are small for many benchmarks (e.g., GPQA, MMLU-Pro, MuSR vary by about 1 point), and MATH scores are near zero overall.
- Stress has minimal impact on hallucination susceptibility, suggesting hallucinations are driven by intrinsic model factors.
Limitations
- Effect sizes on many benchmarks are small and the evaluation uses few models without clear significance testing across stress levels.
- Mapping LLM behavior to human Yerkes-Dodson patterns is an analogy; the paper does not establish a mechanistic equivalence.
- Stress prompts were rated by 20 offline participants, and the dataset release is only promised in supplementary materials.
Experimental setup
Datasets: StressPrompt IFEval BBH MATH GPQA MuSR MMLU-Pro EQ-Bench MMLU TruthfulQA ToxiGen
Benchmarks: IFEval BBH MATH GPQA MuSR MMLU-Pro EQ-Bench MMLU TruthfulQA ToxiGen
Metrics: accuracy normalized accuracy stress score (PCA projection of hidden states)
Baselines: you are a helpful assistant prompt let's think step by step (CoT)
Builds on: Representation Engineering Stress and Coping Theory Job Demand-Control Model Conservation of Resources Theory Effort-Reward Imbalance Model
Closest papers by meaning
- Divergences between Language Models and Human Brains (NeurIPS 2024 · similarity 0.85)
- CogBench: a large language model walks into a psychology lab (ICML 2024 · similarity 0.85)
- Apathetic or Empathetic? Evaluating LLMs' Emotional Alignments with Humans (NeurIPS 2024 · similarity 0.85)
- The Good, The Bad, and Why: Unveiling Emotions in Generative AI (ICML 2024 · similarity 0.84)
- The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans (NeurIPS 2026 · similarity 0.84)
- Generating Computational Cognitive models using Large Language Models (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.