rdh1115/pixels-over-symbols

Pixels over Symbols

Sensory Realism Improves Behavioral Alignment in Models of Cognition

Summary

This is the data of the NeurIPS 2026 paper Pixels over Symbols. We train recurrent and transformer models on visual working-memory tasks from iWISDM, either on a one-hot description of each stimulus (Abstract models) or on the features of a frozen pretrained vision encoder (Natural models), and compare them trial by trial with human participants and vision-language models. The dataset has:

• 240 de-identified human sessions from Amazon Mechanical Turk: 20 participants on each of 12 tasks (n-back and delayed match-to-sample).
• The responses of Gemini 3.1 Pro Preview, two Qwen3-VL models and Nemotron Nano 12B v2 VL to the same trials, 6 runs each.
• The 960 trials over 100 object images from the HvM dataset, in the order participants saw them.
• The 94 trained models the paper analyzes (Abstract, Natural, dimensionality-matched and core-substitution), 6 seeds each, with their training logs.
• Cached evaluations: each model's responses on the human trials, hidden states and DSA distances, enough to redraw every data figure of the paper without a GPU.

Quick start

Use it with the code at BashivanLab/pixels-over-symbols. Download the dataset into that repository's data/ folder to redraw the figures.

git clone https://github.com/BashivanLab/pixels-over-symbols && cd pixels-over-symbols
pip install -r requirements.txt   # in a Python 3.12 environment, see the code README
hf download rdh1115/pixels-over-symbols --repo-type dataset --local-dir data
bash scripts/figures/make_all.sh

To browse the human data alone, each session is a plain CSV, e.g. mturk/human/<task>/s01.csv.

Contents

Path	Contents	Size
mturk/human/	responses of 20 participants per task, one CSV per participant session	24 MB
mturk/trials/	the 12 trial lists and the metadata of the 100 stimuli they use	0.6 MB
mturk/images/	the 100 stimuli as shown to participants	4 MB
mturk/vlm/	responses of 5 vision-language model configurations, 6 runs each	89 MB
checkpoints/	trained models (564 checkpoints of 94 models in 528 run folders), training logs, MANIFEST.csv	5.2 GB
evaluations/	cached model, human and VLM responses on the trials, hidden states and DSA results	1.0 GB
features/randinit_seed2025/	weights of the frozen, randomly initialized vision encoder of the random-encoder control	155 MB

Not included:

• The rest of the HvM images and metadata (Majaj et al., 2015)
• The frozen-encoder features of the HvM images. scripts/train/features.sh in the code repository computes from them.

Human responses

The experiment ran in the browser with PsychoJS between November 2025 and January 2026. The experiment itself is at https://github.com/rdh1115/abs_nat_psychopy_mturk. Participants were first shown a few practice trials with feedback (session 0) and then 100 trials in 5 blocks (session 1 to 5; 20 trials for the 1-back tasks), pressing x for a match and b for a non-match at each response frame. These are the 240 sessions analyzed in the paper. The paper's appendix details the recruitment and quality control. Each participant completed all trials of one task.

mturk/human/<task>/sNN.csv is the PsychoJS export of one session. Codes are not linked across tasks, and one person may have done more than one task. The code reads the files in the order of their codes, so participant sNN of every task is row NN - 1 (counting from 0) of evaluations/human_responses.npy.

The columns the analysis uses are session, TrialIntro.started (set on the first row of each trial), resp.keys and resp.rt (the key pressed and its reaction time in seconds at a response frame), stim1 to stim6 (the stimulus of each frame) and act2 to act6 (the correct answer at each response frame, x or b). The other columns are PsychoJS timing and bookkeeping, in seconds from the start of the experiment.

Trial lists and stimuli

mturk/trials/<task>_trials.csv lists the trials of each task in the order participants saw them, with session, the stimulus of every frame (stim1, ...) and the correct answers (act2, ...). All 12 lists compare the same 100 stimulus pairs. mturk/trials/stim_subset.csv holds the HvM metadata of the 100 stimuli (category, object, pose, position and file name relative to the HvM folder). 36 of its file names keep the leading underscore of the HvM names, which the image files and the trial lists drop, so strip it before matching them. mturk/images/ holds those 100 images as participants saw them, in the disc-fade version of the HvM dataset: the object and its natural background fill a disc that fades into uniform gray. It also holds gray_background.png, a uniform gray (RGB 128) image that no trial list refers to.

VLM responses

mturk/vlm/results-<provider>/<model>--<run>/results_<task>_trials.csv are the responses of

• Gemini 3.1 Pro Preview
• Qwen3-VL-235B-A22B-Thinking
• Qwen3-VL-30B-A3B-Thinking
• Nemotron Nano 12B v2 VL (with and without reasoning)

queried through OpenRouter with the instructions the participants saw, one image and instruction per frame (see the paper's appendix). Each row is one trial; for each frame k the columns raw_resp_k, reasoning_k, error_k, resp_k, target_k and correct_k hold the model's raw output, its reasoning text when the model returned one, any error, the parsed response (x or b), the correct answer and whether they agree. Each file starts with the practice trials (session 0, which the code drops), followed by the main trials (session 1 to 5); trial_index numbers the rows. Most files also have prompt_addon, the text added to the prompt: /no_think for Nemotron without reasoning and empty otherwise. Run 5 does not record it reliably: the column is missing from some of its files, and 521 of the 1,015 rows of Nemotron without reasoning have it empty, although no row of that configuration in any run has reasoning text.

Checkpoints

checkpoints/<group>/<run>/ holds the trained models the paper analyzes, six seeds per model. The groups are abstract (one-hot Abstract models), abstract_until080 (their checkpoints matched at 0.80 validation accuracy, see below), natural, dim_match (dimensionality-matched Abstract models), abs_to_nat and nat_to_abs (the substitution models) and nat_to_abs_until080. Each run folder has the checkpoint files the analysis loads (the final checkpoint, or the accuracy-matched checkpoint, or both), the training log out.txt and, for runs that trained to the end, the training and validation accuracy per epoch (*_avg_acc_rep*.npy).

The weights of the frozen pretrained vision encoders are left out of the Natural and Abs->Nat checkpoints. The code rebuilds the encoders from PyTorch timm when it needs them, and the random encoder's weights are in features/randinit_seed2025/.

Most checkpoints in abstract_until080 and nat_to_abs_until080 come from reruns that retrace the original run and stop at the first validation check at or above 0.80; their names give the accuracy to four decimals (acc0.8012). The 23 with two decimals (acc0.80, 19 in abstract_until080 and 4 in nat_to_abs_until080) are epoch-0 checkpoints of the original runs, used where the original run already had a checkpoint labelled 0.80, and their folders hold the original run's training log.

Evaluations

evaluations/ holds what the figure and statistics scripts of the code repository read.

• Every model's responses on the MTurk trials (<key>[_match<accuracy>]_{conf,acc}.npy, plus _acc_draws.npy for the models with input noise)
• Participant and VLM responses in the same format
• Hidden states of the 26 models compared by the DSA (*_{seq,pre}_feature.npy)
• DSA distances (dsa_distances.npz)
• Distances of the DSA sensitivity analysis (dsa_sensitivity/)
• Linear separability of the encoders' features used in Figure 12.

The code's analysis/experiment_pipeline.py describes the array layouts, and scripts/evaluate.py recomputes any of the model files from the checkpoints.

License

The data is released under CC BY 4.0. The images in mturk/images/ are from the HvM dataset (Majaj et al., 2015) and keep its terms.

Citation

@inproceedings{bai2026pixels,
  title     = {Pixels over Symbols: Sensory Realism Improves Behavioral Alignment in Models of Cognition},
  author    = {Bai, Hao Yuan and Lei, Xiaoxuan and Weng, Zihan and Pourrahimi, Motahareh and Gomez, Lucas and Bashivan, Pouya},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}