Title: Investigating the Effect of Emotional Prompting on Large Language Model Reasoning Performance

Model: Qwen3-4B

Dataset: GSM8K

Prompt Conditions:

• Normal
• Career
• Tip
• Careful
• Expert
• Threat

Metrics:

• Accuracy
• Flip Rate
• Directional Shift
• Consistency

---

Steering with Empathy (Phase 1 and Phase 2)

Models: Qwen3-1.7B, Qwen3-8B, Llama-3.2-3B-Instruct. Datasets: GSM8K, CommonsenseQA, MoralChoice, Social-IQA.

• prompts/ – the 84-prompt set (7 emotions x 3 intensities x 4 versions) as prompts_all_versions.json, with the loader prompts_all_versions.py (make_prompts(ret)), the prompt chosen per model/dataset (selected_prompts.json), and all 168 rendered prompts (all_prompts_rendered.csv, All_prompts.xlsx).
• phase2/ – SAE feature pipeline: layer probing, feature extraction, Jaccard overlap, threshold sweep, neutral-wording control.
• analysis/wording_control_extended.py – extended neutral-rewording control (45 wordings, rank test, length adjustment).
• paper/ – LaTeX source, figures and the compiled PDF.

Usage of the prompts:

from prompts.prompts_all_versions import make_prompts
P = make_prompts("a number (e.g. 42)")   # GSM8K; use "A, B, or C" for Social-IQA
text = P["career_high_v1"](question)