MODG: Matched Outcomes, Divergent Gaze

Data for "Matched Outcomes, Divergent Gaze: How Foveated MLLMs Search Compared to Humans."

Paper: arXiv:2608.16514 · Code: github.com/kmamine/MODG

Three multimodal LLMs search COCO-Search18 scenes one fixation at a time, zero-shot:

• Qwen3.5-35B-A3B
• GLM-4.6V-Flash
• Gemma-4-E4B

Each model sees the scene through a foveated renderer centred on its current gaze. The models are compared with the ten human scanpaths recorded for each scene.

The harness loop on a real logged episode. A foveated glimpse is rendered at the current gaze (red +); the model replies with one directive; the gaze moves; repeat until the model says FOUND or ABSENT.

The foveation bracket. Only Geisler–Perry is human-matched; the gist-k ladder and crop are synthetic.

Headline result: the models match or beat humans on the present/absent decision and on finding the target, but their gaze process is non-human: far more self-consistent than two humans are with each other.

Example scanpaths: human (blue) vs. Qwen3.5-35B-A3B (red) on the same scenes across foveation conditions. The green box is the target; numbers give fixation order.

This dataset contains:

• every model scanpath and every raw model response
• the one-shot baseline answers
• the derived metrics
• the frozen stimulus subset, including the human scanpaths, needed to reproduce the paper offline

Layout

The layout mirrors the code repository, so a single command drops it into place:

git clone https://github.com/kmamine/MODG.git && cd MODG
hf download Amine-CV/MODG --repo-type dataset --local-dir . --exclude README.md --exclude .gitattributes

data/subset/
  images/{TP,TA}/<category>/<coco_id>.jpg    285 stimulus images (141 target-present, 144 target-absent)
  human_fixations/fixations_{tp,ta}.json     COCO-Search18 human scanpaths for these images (10 subjects each)
  image_ids_{tp,ta}.json                     frozen subset + strata (eccentricity, clutter, target bbox)
  subset_manifest.json, config.yaml, README.md
results/raw/<model>/fov-<condition>/
  scanpaths.jsonl                            one record per episode (model × scene × seed)
  responses.jsonl                            one record per model call (full raw text)
  run_config.json                            exact config, prompt and model id of the run
results/baselines/<model>/                   one-shot existence (TP+TA) and localization (TP) answers
results/metrics/paper_data.json              every number reported in the paper (+ other metric tables)

Models and conditions

Models: qwen3.5-35b-a3b (Qwen/Qwen3.5-35B-A3B), glm-4.6v-flash (zai-org/GLM-4.6V-Flash) and gemma-4-e4b-it (google/gemma-4-E4B-it, thinking disabled).

Main design: 9 foveation conditions × 285 scenes × 5 seeds, at temperature 0.6.

Condition	What it is
fov-sharp	no foveation
fov-geisler_perry	Geisler & Perry (1998) acuity falloff; the only human-matched condition
fov-gaussian-k{8,16,24,32,48,128}	synthetic peripheral degradation
fov-crop	fovea-only disc

The Qwen directory also holds pilot-only extras, which are not part of the main 9-condition design:

• fov-gaussian-k4 and fov-gaussian-k64
• temperature-sweep directories (*-t0.2, *-t1.0)
• seeds 5–7 on a pilot subset of scenes

Record schemas

Coordinate frame. All coordinates live in the 1680×1050 px frame in which the COCO-Search18 humans were recorded (about 30 px per degree). Model coordinates are normalized to [0, 1]: multiply by 1680 and 1050 to get pixels. Human X/Y are already in pixels.

scanpaths.jsonl: one model episode

{"image_id": "000000019544.jpg", "condition": "TP", "target_category": "bottle",
 "agent": "<model>", "seed": 0,
 "fixations": [{"x": 0.5, "y": 0.5, "step": 0}, {"x": 0.17, "y": 0.43, "step": 1}],
 "stopped": true, "stop_reason": "found | not_present | max_fixations", "found": true}

• condition is target-present (TP) or target-absent (TA). The foveation condition is the directory name.
• Step 0 is the forced central start.
• stop_reason is the model's own decision. max_fixations means the 50-glimpse safety cap was reached.

responses.jsonl: one model call

The fields are:

• image_id, condition, seed, step
• action (look / found / absent), x, y
• parse_status, retried, clamped
• prompt_tokens, completion_tokens, latency_s
• raw_text: the model's full reply

human_fixations/*.json

These are raw COCO-Search18 records: name, subject, task, condition, bbox, X, Y, T (durations), length, correct, RT, split. See data/subset/README.md.

License

Our contributions are released under the MIT License: the model-generated scanpaths, responses, baseline answers and derived metrics.

Third-party data is NOT covered and remains under the terms published by its original authors:

• The COCO-Search18 human scanpaths (Chen et al., 2021; Yang et al., 2020); see https://sites.google.com/view/cocosearch/.
• The stimulus images, from MS-COCO, subject to the COCO terms of use and the individual Flickr image licenses; see https://cocodataset.org/#termsofuse.

The copies included here are provided for research reproducibility only. Please cite the original sources.

Citation

If you use this dataset, please cite our paper:

@misc{kerkouri2026matchedoutcomesdivergentgaze,
  title         = {Matched Outcomes, Divergent Gaze: How Foveated MLLMs Search Compared to Humans},
  author        = {Mohamed Amine Kerkouri and Marouane Tliba and Aladine Chetouani and Ulas Bagci and Alessandro Bruno},
  year          = {2026},
  eprint        = {2608.16514},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.16514}
}