10‑Tails Multi‑Turn Visual Reasoning

Zenodo Package: 10‑tails Multi‑turn Analysis This package contains a minimal, ready‑to‑run setup to generate, analyze, and visualize multi‑turn dialogue logs across five conditions (BASELINE/OBSERVE/ANALYZE/EVALUATE/INTERPRET). It is organized so that after download you can run analysis immediately. Experiment name: Ten‑Tails Multi‑Turn Visual Reasoning (10T‑MVR) Structure scripts/: executable scripts run_gpt5_10tails.py: generate 10‑turn JSON logs per condition using an image stimulus analyze_10tails_allturns.py: aggregate JSON logs (long‑format CSV, condition×turn summary, final‑turn summary, plots) visualize_10tails_results.py: extra figure set from the summary CSV calculate_10turn_cumulative_stats.py: cumulative stats across 10 turns (optional, with figures/LaTeX) analyze_10tails_morph_janome.py: optional morphological analysis using Janome (regenerates results_morph/) data/ raw/: raw JSON logs stimuli/: stimulus images (e.g., stimulus.jpeg) results/ CSV outputs: metrics_per_turn.csv, summary_by_condition_turn.csv, final_turn_summary_by_condition.csv plots/: generated figures From analyze_10tails_allturns.py: final_len_chars_bar.png, trend_*_per_kchar.png, final_jsd_heatmap.png, final_element_coverage_heatmap.png From visualize_10tails_results.py: v10_trend_*.png, v10_final_*.png cumulative_stats/: optional cumulative outputs results_morph/ Included morphological analysis artifacts (precomputed): CSVs: morph_metrics_per_turn.csv, morph_summary_by_condition_turn.csv, final_turn_morph_summary_by_condition.csv, final_turn_stability.csv, top_tfidf_nouns_final.csv, pmi_pairs_final_{BASELINE,OBSERVE,ANALYZE,EVALUATE,INTERPRET}.csv plots/: figures including top20_tfidf_nouns_final_{BASELINE,OBSERVE,ANALYZE,EVALUATE,INTERPRET}.png, final_turn_stability_bar.png, trend_noun_types.png, trend_new_noun_types.png, trend_noun_TTR.png, drift_jaccard_1_minus_sim.png, heaps_curve_cumulative_unique_nouns.png requirements.txt: Python dependencies Sample JSON logs are copied into data/raw/ and existing figures/CSVs are consolidated under results/ (original folders are kept for reference). Quick Start (analysis only) Python 3.9+ is assumed. Install dependencies: python3 -m pip install -r requirements.txt Run analysis and generate plots (from data/raw → results/): python3 scripts/analyze_10tails_allturns.py --in ./data/raw --out ./resultspython3 scripts/visualize_10tails_results.py Main outputs: results/metrics_per_turn.csv results/summary_by_condition_turn.csv results/final_turn_summary_by_condition.csv results/plots/*.png Generate logs first, then analyze You can generate 10‑turn JSON logs using the OpenAI API. Install dependencies (if not yet): python3 -m pip install -r requirements.txt Set environment variable: export OPENAI_API_KEY=sk-... Place an image under data/stimuli/ and run: python3 scripts/run_gpt5_10tails.py \ --image ./data/stimuli/stimulus.jpeg \ --model gpt-5 \ --replicates 10 \ --out . JSON files will be saved to ./data/raw/. Analyze and visualize: python3 scripts/analyze_10tails_allturns.py --in ./data/raw --out ./resultspython3 scripts/visualize_10tails_results.py Optional cumulative stats: python3 scripts/calculate_10turn_cumulative_stats.py Optional morphological analysis (regenerate results_morph/): python3 scripts/analyze_10tails_morph_janome.py --in ./data/raw --out ./results_morph Notes Analysis uses simple dictionary‑based counts; morphological analysis results are included as precomputed artifacts under results_morph/. Regeneration may require extra packages (see optional deps in requirements.txt). run_gpt5_10tails.py requires network access and an OpenAI API key; be mindful of usage costs. License Code: MIT License (see LICENSES/LICENSE_CODE_MIT) Data, text, and figures (including data/, results/, results_morph/): CC BY 4.0 (see LICENSES/LICENSE_CONTENT_CC-BY-4.0.txt) Please include a citation to “Ten‑Tails Multi‑Turn Visual Reasoning (10T‑MVR)” (see CITATION.cff). Citation 10.5281/zenodo.16919171