Skip to content

Repository files navigation

Distance-Engagement Oscillation in LLM Prompting

DOI

Replication materials for the paper:

Gokmen, G. (2026). Does AI Dream of Electric Sheep? Testing Distance-Engagement Oscillation as a Prompting Framework for Creative Reframing in Large Language Models.

What this study tests

Perceptual Reframing Theory (PRT; Gokmen, submitted) identifies nine cognitive paths through which humans shift perception during creative problem-solving, and proposes Distance-Engagement Oscillation (DEO) as the mechanism that makes those shifts durable. This study translates the theory into LLM prompting strategies and tests two questions:

  1. Do reframing paths improve LLM reasoning? (Analysis 1: vanilla vs. path-guided prompting across 50 problems)
  2. Does oscillation between distance and engagement outperform either mode alone? (Analysis 2: four-condition comparison across 20 pure-distance-path problems)

Key findings

The predicted ordering DEO > Distance > Engagement > Vanilla held across all 3 models and all 3 scorers (9/9 combinations, all p < .001). Effect sizes were very large (Cohen's d = 1.29-1.63 for Analysis 1).

MacGyver benchmark update (30 Mar 2026)

A follow-up study on 956 functional fixedness problems from the MacGyver benchmark (Tian et al., 2024) replicated the core finding and revealed that a 2-step "DEO Brief" protocol (ANALYSE then FEEL) matches or exceeds the full 4-step DEO on 2 of 3 models:

Condition Llama 70B Qwen 32B Scout 17B
Vanilla (fair) 3.3% 4.1% 2.7%
Distance only 15.1% 15.5% 14.1%
Engagement only 7.2% 6.8% 7.9%
DEO Brief (2-step) 20.7% 44.2% 27.6%
DEO full (4-step) 29.2% 29.2% 29.3%

All omnibus tests p < .001. The /reframe slash command has been updated to use the 2-step protocol as the default; /reframe-long provides the original 4-step version.

Experiment design

  • 50 novel problems across 8 categories of perceptual lock-in (+ 956 MacGyver benchmark problems)
  • 3 open-weight LLMs: Llama 3.3 70B, Qwen3 32B, Llama 4 Scout 17B-16E
  • 6 conditions: vanilla, vanilla (fair), distance-only, engagement-only, DEO Brief (2-step), DEO (4-step oscillation)
  • 5 runs per condition for 50-problem study; 1 run per condition for MacGyver (n = 956 provides sufficient power)
  • 3 independent scorers for 50-problem study: self-scoring + Claude Sonnet 4 + GPT-4.1 (all blind to condition)
  • MacGyver scoring: GPT-4.1 mini feasibility classification (A/B/C/D) against human-verified gold solutions
  • All generation via Groq API; zero-shot, prompt-level intervention only

Repository structure

deo-llm-reframing/
├── src/reframing_ai/
│   ├── problems.py          # 50 problem definitions with all condition preambles
│   ├── deo_problems.py      # 10 DEO-specific problem configurations
│   ├── runner.py             # Study 1 execution engine
│   ├── deo_runner.py         # Study 2 execution engine
│   ├── scorer.py             # Self-scoring logic
│   ├── cross_scorer.py       # Claude/GPT-4.1 cross-scoring
│   ├── config.py             # API and model configuration
│   └── ...                   # Report generation, prompts
│
├── results/                  # Raw experiment outputs (307 JSON files)
│   ├── llama/raw/            # Llama 3.3 70B responses + scores
│   ├── qwen/raw/             # Qwen3 32B responses + scores
│   └── llama4scout/raw/      # Llama 4 Scout responses + scores
│
├── paper/
│   ├── study1_data.csv       # Flattened data for analysis
│   ├── study2_data.csv       # Four-condition data
│   ├── stats_results.md      # Full statistical results
│   ├── figures/              # Publication figures (PNG + PDF)
│   └── tables/               # Summary tables
│
├── run_experiment.py         # Run Study 1 (vanilla vs. reframing)
├── run_deo.py                # Run Study 2 (four-condition DEO)
├── run_full.py               # Run both studies across multiple models
├── run_stats.py              # Statistical analysis (Python)
├── run_cross_score.py        # Cross-model scoring
├── export_for_r.py           # Export data for R verification
├── generate_figures.py       # Publication figures
└── verify_stats.R            # Independent R verification of all statistics

Reproducing the experiment

Prerequisites

  • Python 3.12+
  • uv package manager
  • API keys for Groq, Anthropic, and OpenAI

Setup

git clone https://github.com/gokmengokhan/deo-llm-reframing.git
cd deo-llm-reframing
cp .env.example .env
# Edit .env with your API keys
uv sync

Run experiments

# Study 1: vanilla vs. reframing (single model)
uv run python run_experiment.py

# Study 2: four-condition DEO comparison
uv run python run_deo.py

# Both studies across all models
uv run python run_full.py --models llama qwen llama4scout

# Cross-score with Claude and GPT-4.1
uv run python run_cross_score.py

Run analysis

# Statistical analysis (Python)
uv run python run_stats.py

# Export for R verification
uv run python export_for_r.py

# Independent R verification
Rscript verify_stats.R

# Generate publication figures
uv run python generate_figures.py

Cost estimate

The full experiment (3 models, 4 conditions, 5 runs, 3 scorers) costs approximately $15-20 USD across the Groq, Anthropic, and OpenAI APIs.

Claude Code slash commands

Based on the MacGyver findings, the default /reframe command now uses the 2-step DEO Brief protocol. The full 4-step version is available as /reframe-long.

  • /reframe — 2-step DEO Brief: ANALYSE (select one of six distance paths) then FEEL (embodied engagement). Research shows the core oscillation is the mechanism.
  • /reframe-long — Full 4-step DEO: ANALYSE, FEEL, REFRAME, ENVISION. For complex or emotionally charged problems.
/reframe Our hospital ED loses $2.3M per year. Should we close it?

To install globally (available in all projects):

mkdir -p ~/.claude/commands
cp .claude/commands/reframe.md ~/.claude/commands/reframe.md
cp .claude/commands/reframe-long.md ~/.claude/commands/reframe-long.md

Data format

Each JSON file in results/ contains:

  • Problem metadata (id, category, theory connection, success criteria)
  • Full LLM responses for each condition (5 runs each)
  • Blind scores from each scorer (5 dimensions + correctness with justifications)
  • Token counts and generation times

The paper/*.csv files contain the same data flattened for statistical analysis, with one row per model-problem-condition-scorer-run combination.

Citation

If you use this dataset or methodology, please cite:

@misc{gokmen2026deo,
  title={Does {AI} Dream of Electric Sheep? Testing Distance-Engagement Oscillation in {LLM} Creative Reframing},
  author={G{\"o}kmen, G{\"o}khan},
  year={2026},
  doi={10.5281/zenodo.19252225},
  url={https://zenodo.org/records/19252225},
  publisher={Zenodo}
}

License

CC BY 4.0. See LICENSE.

About

Replication materials: Testing Distance-Engagement Oscillation as a prompting framework for creative reframing in LLMs

Topics

Resources

Stars

10 stars

Watchers

1 watching

Forks

Releases

Packages

Contributors

Languages