ST5230 Project: Prompt Sensitivity of Uncertainty Estimates in LLMs

Research Question

How sensitive are LLM uncertainty estimates to semantically equivalent prompt perturbations?

Models

Role	Model	Params	Source
Small open-source	mistralai/ministral-3b-2512	3B	OpenRouter
Medium open-source	meta-llama/llama-3.1-8b-instruct	8B	OpenRouter
Large proprietary	openai/gpt-4o	undisclosed	OpenRouter

See docs/model_changes.txt for why these models were chosen over the originally proposed ones.

Datasets

• SciQ — 250 samples (science multiple-choice)
• TruthfulQA — 250 samples (common misconception multiple-choice)

Prompt Variants

Each sample is evaluated under 4 semantically equivalent prompt formats:

• V0: Baseline (instruction-first, A) format)
• V1: Reworded (different phrasing, A. format)
• V2: Question-first ((A) format)
• V3: Minimal (no instruction, A: format)

Key Metrics

• Prediction Flip Rate — how often the predicted answer changes across prompts
• Logit Margin Variance — variance of (top-1 minus top-2) logit across prompts
• Hidden Instability — samples where prediction is stable but confidence fluctuates

Quick Start

pip install -r requirements.txt

# Prepare data (once)
python sources/data_utils.py

# Run experiments (example)
python run_experiment.py --model meta-llama/llama-3.1-8b-instruct --dataset sciq --openrouter

# Analyze results
python analyze_results.py

See steps.txt for full reproduction instructions.

Project Structure

├── run_experiment.py           # Main experiment runner
├── analyze_results.py          # Analysis + 5 figures
├── sources/
│   ├── data_utils.py           # Dataset download & sampling
│   ├── model_inference.py      # API inference (OpenRouter)
│   ├── prompt_variants.py      # 4 prompt variant definitions
│   └── uncertainty_metrics.py  # USS, Flip Rate, LM Variance
├── outputs/
│   ├── data/                   # Sampled datasets
│   ├── results/                # Per-model inference results
│   └── figures/                # Generated PDF figures
├── docs/
│   ├── Group3_Project_Proposal.pdf
│   └── model_changes.txt       # Model selection decision log
└── steps.txt                   # Reproduction guide (中文)