S1CogBench aggregate results

S1CogBench asks what System One decision models can and cannot do on established cognitive experiments. These models, such as Jev and Laya, return one fast decision over the legal options and leave no text to inspect. The benchmark evaluates five open models and Jev on 24 cognitive tasks in 27 modes, keeps each task's conditions and scoring apart from disclosed adaptations, and scores human-response prediction, autonomous execution and event probabilities separately.

This dataset holds the aggregate results of the paper. The tasks, runners and scorers are in the GitHub repository.

Paper · Project website · GitHub · Evaluate your model

Files

File	Contents
aggregate/table2.csv	Table 3 of the paper. One row per task mode and primary target (27 rows), one column per model, with the uniform and task-specific references. Cells are printed values, so +∞, n.a. and footnote marks stay as text.
aggregate/table2.json	The same table with the result row behind each cell.
aggregate/results.csv, aggregate/results.json	All 1,961 aggregate rows, with task, model, protocol, mode, metric, target, unit, n, aggregation and notes.
aggregate/reference_results.json, aggregate/baseline_provenance.json	Reference values and how each was fitted or computed.
aggregate/tasks.json	The catalogue of 31 tasks: the 24 evaluated tasks and 7 deferred ones, with sources, protocols and adaptations.
aggregate/metric_status.json	For each task, model and shared metric, whether the metric applies and why.
aggregate/rlwm_curves.csv	RLWM learning curves by set size for humans, fitted references and the open models.
aggregate/bandit_condition_scores.csv	Stationary bandit reward and pseudo-regret by condition, for the five open models and a Gaussian Thompson sampling reference.
aggregate/api_time_cost_per_target.csv	Latency, cost and output tokens of Jev and Claude per target.
aggregate/minitaur_nonliteral_reparsed.csv	Minitaur's nonliteral-number reports after the parser fix.
aggregate/result_schema.json, aggregate/grouping_guide.md	The row schema and how to compare results without mixing contexts.
aggregate/data_manifest.json, aggregate/cpu_validation_24_status.json	Provenance of the files and the CPU validation of all 27 modes.

Load the tables

from datasets import load_dataset

table = load_dataset("dengzhe-hou/S1CogBench", "main_table", split="train")
rows = load_dataset("dengzhe-hou/S1CogBench", "results", split="train")

There is no pooled cognitive score and no overall ranking. Each target has its own unit and direction, so compare models within a row, and match task, protocol, mode, metric, target, unit and aggregation as the grouping guide describes.

Evaluate your own model

The task inputs are not in this dataset. Several sources keep their own terms, and the Psych-101 test transcripts are gated by their authors, so the repository fetches every input from its original source at a pinned revision. An adapter is one Python file, and the repository has templates for Hugging Face language models, decision APIs and text-generation APIs.

git clone https://github.com/dengzhe-hou/S1CogBench.git
cd S1CogBench
python -m pip install -r requirements.txt
python scripts/fetch_public_sources.py all
python -m s1cogbench eval --smoke --output runs/smoke
python -m s1cogbench eval --adapter my_agent.py --device cuda:0 --model-label my-model --tasks all --output runs/my-model

The output rows use the same schema as aggregate/results.csv. See the quickstart and models and adapters.

License

The aggregate results are released under CC BY 4.0. Reference values computed from third-party human data keep the terms of their original sources, which aggregate/tasks.json lists for each task. The S1CogBench code on GitHub is MIT licensed.

Citation

@misc{hou2026s1cogbench,
  title         = {{S1CogBench}: What System One Decision Models Can and Cannot Do on Cognitive Experiments},
  author        = {Hou, Dengzhe and Yamada, Kazunori D},
  year          = {2026},
  eprint        = {2611.00911},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2611.00911}
}