Polysemy Outputs

Raw model generations for the paper "Where did the ambiguity go? Examining how multimodal models interpret polysemous words."

Each polysemous word (e.g. bank, bolt, trunk) is presented with no disambiguating context — the prompt is the bare word — and the model's chosen sense is observed over many samples. The same word set is run in two modalities (text-to-image and text generation) and scored by the same judges, so their sense distributions are directly comparable, and anchored against a human baseline.

This repository holds the raw outputs only (generated PNGs and TXT). The sense labels, histograms, and analysis live in the code repository.

Layout

Outputs are organized into sectors mirroring the paper's experiment structure. Within every sector the layout is word-first:

<sector>/.../<word>/<model_key>/NN.{png,txt}

.
├── 01_main_english/            Central panel · 100 English words · 30 samples/word/model
│   ├── image/                    17 text-to-image models  (<word>/<model>/NN.png)
│   └── text/                     17 text-generation models (<word>/<model>/NN.txt)
│
├── 02_human_baseline/          Prolific human reference · 100 words
│   ├── image/                    "What image comes to mind?"   (<word>/human/NN.txt)
│   └── meaning/                  "Use the word in a sentence"  (<word>/human/NN.txt)
│
├── 03_cross_lingual/           25 words each · OpenAI / Grok / Gemini families
│   ├── turkish/{image,text}/
│   └── french/{image,text}/
│
├── 04_stated_vs_revealed/      Image-framing ("imagine") text generation
│   └── text_imagine/             15 text models (<word>/<model>/NN.txt)
│
└── 05_ablations/               "Sources of collapse" (§4.3)
    ├── preference_tuning/        SDXL ± Diffusion-DPO, SimpleAR + GRPO   [see note]
    ├── architecture/            open autoregressive comparison models     [see note]
    └── telephone/               iterated cross-modal transmission
                                 <family>/<seed_word>/<chain>/i{0..4}.{png,txt}

Model keys

01_main_english/image (17): openai_1, openai_mini, openai_15, openai, grok_base, grok_pro, grok, gemini_25, gemini, gemini_pro, qwen_1, qwen_20, qwen, flux_1pro, flux_flex, flux, flux_max.

01_main_english/text (17): openai_35, openai_4o, openai_5, openai_54nano, openai_54mini, openai_55, grok_420, grok_420_nr, grok_43, gemini_25flash, gemini_35flash, gemini_31pro, qwen_25_7b, qwen_35_9b, qwen_35_397b, glm_51, glm_52.

03_cross_lingual/*/image (10): the OpenAI (×4), Grok (×3), and Gemini (×3) image keys above. 03_cross_lingual/*/text (12): OpenAI (×6), Grok (×3), Gemini (×3) text keys.

04_stated_vs_revealed/text_imagine (15): the OpenAI, Grok, Gemini, and Qwen text keys.

Ablation models (05_ablations/)

The comparison-model runs were generated on a compute cluster and are uploaded separately. Their intended target paths (word-first, like every other sector) are:

glm_image (GLM Image, partial run) belongs with the main image panel at 01_main_english/image/<word>/glm_image/NN.png.

Janus and Anole outputs are intentionally excluded from this dataset.

Notes

• 30 samples per word per model (some cells are short where a model errored).
• The image prompt is the bare word; the text prompt is "Use the following word in a single sentence: {word}." (localized for Turkish/French).
• Sense inventories, judge labels, and histograms are not in this dataset (raw outputs only).