Domain stringclasses 10
values | Task mode stringlengths 4 24 | Primary target stringlengths 9 18 | Better direction stringclasses 2
values | Uniform float64 0.35 75 | Laya stringlengths 2 5 | Intern float64 0 82.5 | Llama float64 0 81.8 | Minitaur float64 0 86.7 | Qwen float64 0 90 | Nox float64 0.18 99.3 | CLM float64 0 69.5 | Jev stringlengths 2 5 | Opus stringclasses 7
values | Sonnet stringclasses 7
values | Haiku stringlengths 3 6 | Reference float64 0 100 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Evidence | Bayesian urn | Posterior MAE | lower | 0.349 | 0.291 | 0.235 | 0.24 | 0.294 | 0.246 | 0.175 | 0.286 | 0.068 | 0.001 | 0.001 | 0.001 | 0 |
Evidence | Multiple-cue judgment | Human NLL | lower | 2.197 | 2.342 | 1.57 | 2.687 | 1.984 | 3.012 | 1.491 | 2.675 | +∞ | n.a. | n.a. | n.a. | 2.029 |
Evidence | Multi-attribute choice | Human NLL | lower | 0.693 | 0.730 | 0.546 | 0.786 | 0.722 | 0.613 | 0.366 | 1.039 | 0.329 | n.a. | n.a. | 0.206s | 0.192 |
Evidence | Probability coherence | Human MAE (pp) | lower | 32.13 | 32.98 | 18.95 | 28.57 | 19.04 | 21.87 | 18.03 | 42.13 | 17.66 | canc. | canc. | part. | 16.54 |
Category | SHJ categories | Accuracy (%) | higher | 50 | 50.0 | 65 | 52.9 | 54.5 | 71.3 | 94.5 | 50.4 | 87.4 | n.e. | n.e. | 91.2 | 93.8 |
Category | Medin categories | Accuracy (%) | higher | 50 | 49.6 | 72.3 | 53.5 | 55.9 | 53.8 | 92.8 | 50.4 | 86.8 | n.e. | n.e. | 97.4 | 96.6 |
Category | Weather prediction | Accuracy (%) | higher | 50 | 54.5 | 73 | 47.5 | 53.8 | 50.5 | 81.2 | 52.5 | 80.8 | n.e. | n.e. | 80.2 | 78.2 |
History | RLWM | Accuracy (%) | higher | 33.3 | 34.9 | 77.8 | 41.7 | 49 | 59.3 | 86.3 | 36.7 | 89.5 | n.e. | n.e. | 89.4 | 89.3 |
History | Recent probes | Accuracy (%) | higher | 50 | 49.7 | 82.1 | 55.6 | 49.5 | 53.5 | 99.3 | 55.9 | 99.0 | n.e. | n.e. | 100.0 | 100 |
Adaptation | WCST, prediction | Human NLL | lower | 1.386 | 1.462 | 1.378 | 1.431 | 1.344 | 1.968 | 1.552 | 2.368 | 1.543 | n.a. | n.a. | n.a. | 0.84 |
Adaptation | WCST, autonomous | Accuracy (%) | higher | 25 | 25.5 | 27.9 | 31.4 | 28.2 | 25.2 | 32.1 | 25.7 | 28.5 | n.f. | n.f. | n.f. | 48.9 |
Adaptation | Restless bandit | Better arm (%) | higher | 50 | 55.4 | 65.6 | 57.1 | 49 | 50.5 | 69 | 52.6 | 63.9 | part. | 91.1 | part. | 68.8 |
Adaptation | Instrumental learning | Better arm (%) | higher | 75 | 73.9 | 81.4 | 79.2 | 86.7 | 80 | 89.8 | 69.5 | 86.7 | n.e. | n.e. | part. | 94.9 |
Exploration | Horizon, prediction | Human NLL | lower | 0.693 | 0.702 | 0.518 | 0.503 | 0.409 | 0.661 | 0.408 | 1.161 | 0.414 | n.a. | n.a. | 0.390s | 0.4 |
Exploration | Horizon, autonomous | Optimal (%) | higher | 50 | 54.8 | 78.7 | 66.4 | 71.7 | 63.3 | 86 | 50 | 82.9 | n.f. | n.f. | n.f. | 74.6 |
Exploration | Stationary bandits | Optimal arm (%) | higher | 12.5 | 17.6 | 37.4 | 13.5 | 9.8 | 28.9 | 35.4 | 15.4 | 47.0 | n.e. | n.e. | n.e. | 68.1 |
Information | Hybrid sampling | Human NLL | lower | 1.386 | 1.390 | 1.116 | 1.515 | 1.039 | 2.648 | 2.044 | 3.012 | +∞ | n.a. | n.a. | n.a. | 0.264 |
Information | Mouselab, prediction | Human NLL | lower | 1.634 | 1.717 | 2.109 | 2.743 | 2.125 | 2.464 | 1.619 | 7.817 | 1.458 | n.a. | n.a. | n.a. | 1.367 |
Information | Mouselab, autonomous | Expected return | higher | 7.01 | 0.00 | 0 | 3.77 | 0 | 17.4 | 3.77 | 0 | 8.21 | n.e. | n.e. | n.e. | 11.77 |
Preference | BART | Mean reward | higher | 0.84 | 0.00 | 0.74 | 0.28 | 0 | 0 | 4.72 | 1.03 | 1.76 | n.e. | n.e. | 6.27 | 10.49 |
Preference | Temporal discounting | Delayed reward (%) | null | 50 | 0.0 | 16.7 | 0 | 80 | 90 | 16.7 | 0 | 50.0 | n.e. | n.e. | 77.8 | 41.5 |
Planning | Two-step | Mean reward | higher | 0.5 | 0.50 | 0.52 | 0.5 | 0.51 | 0.52 | 0.53 | 0.49 | 0.55 | n.e. | n.e. | 0.55 | 0.52 |
Social | False-belief QA | Accuracy (%) | higher | 50 | 39.5 | 82.5 | 81.8 | 65 | 82.8 | 81.9 | 67.4 | 71.5 | n.e. | n.e. | 82.6 | 100 |
Social | Politeness | Aggregate NLL | lower | 1.014 | 1.192 | 0.799 | 1.584 | 1.008 | 1 | 0.749 | 1.195 | +∞ | n.a. | n.a. | +∞s | 0.587 |
Social | Pragmatic interpretation | Aggregate NLL | lower | 1.282 | 1.592 | 0.814 | 1.851 | 1.049 | 1.24 | 0.895 | 1.081 | +∞ | n.a. | n.a. | +∞s | 0.416 |
Social | Nonliteral number | MAE to human mean | lower | 34.77 | 30.86 | 16.87 | 25.1 | 28.53 | 29.75 | 17.73 | 18.97 | 20.37 | 15.34 | 20.81 | 22.52 | 4.98 |
Metacognition | Self-assessment | Human NLL | lower | 2.039 | +∞ | 2.029 | 2.523 | 1.885 | 2.262 | 1.893 | 3.094 | +∞ | n.a. | n.a. | n.a. | 1.619 |
S1CogBench seven-model aggregate results
S1CogBench evaluates decision models on 24 cognitive tasks in 27 modes. It keeps human-response prediction, autonomous execution and event-probability reporting separate, using each task's original scoring with disclosed adaptations.
The current release contains the seven open models Laya, Intern-Decision-4B, Llama-3.1-8B-Instruct, Minitaur8B, Qwen3.5-4B, Decision-2.0-Nox-4B and CLM-v0.1-8B. It includes 189 primary scores and 70 shared-readout metric values for Horizon and WCST. These are the same accepted values published on GitHub, including the completed Nox and CLM sampling runs. No model was rerun for this synchronization.
Live leaderboard · GitHub · Current release details · Paper
Download the current results
from datasets import load_dataset
repo = "dengzhe-hou/S1CogBench"
revision = "seven-model-20261110"
table = load_dataset(repo, "main_table", revision=revision, split="train")
primary = load_dataset(repo, "primary_scores", revision=revision, split="train")
readouts = load_dataset(repo, "readout_controls", revision=revision, split="train")
rows = load_dataset(repo, "results", revision=revision, split="train")
The train split is a distribution label for aggregate evaluation results. It is not model-training data or a participant split.
| Configuration | Contents |
|---|---|
main_table |
27 task-mode rows with seven open-model columns, Uniform, task-specific references and the original API coverage. Printed values and footnote marks remain text. |
primary_scores |
189 unrounded primary scores, one per open model and task mode. |
readout_controls |
70 unrounded values from legal argmax and legal sampling across all seven models on Horizon and WCST. Horizon has two metrics and WCST has three. |
results |
2,085 aggregate records containing current primary scores and readout controls alongside retained historical diagnostics. Not every historical diagnostic covers all seven models. |
rlwm_curves, bandit_condition_scores, api_time_cost |
Unchanged historical secondary tables with their original model and API coverage. These configurations do not imply seven-model curve or API coverage. |
aggregate/primary_sources.json and aggregate/readout_sources.json retain provenance and the original scoring contexts. aggregate/table2.json links each printed cell to its result row. RELEASE.json records this release and the matching GitHub commit. Other retained reference, task, schema, status and validation files describe the original public snapshot and its coverage.
There is no pooled cognitive score or overall ranking. Compare matching task, protocol, mode, metric, target, unit and aggregation. Unrounded values are in the result records. Percentage fractions and printed percentage points have explicit units. Infinite and undefined values are retained without clipping or invented scores.
The readout runs retain the original task seeds, probability extraction and action-dependent feedback histories. CLM uses its original first-2048-token context. Nox's original native/argmax records used A100 MIG, whereas its added sampling runs used a whole A100. These scores do not establish a matched-hardware latency comparison. See the linked release details for checkpoint revisions and model-specific settings.
Reproduce or submit a result
Task runners, model-adapter templates, data acquisition and original scorers live on GitHub. Follow the quickstart and the model integration guide. To add an evaluated model to the public leaderboard, follow the submission guide. Reviewed aggregate submissions update the leaderboard after their GitHub pull request is merged.
This dataset contains aggregate results, not model weights or raw participant transcripts. Gated and third-party task inputs must be obtained from their original sources under their original terms.
Historical snapshot
The previous five-model dataset is preserved unchanged at the tag arxiv-five-model-20261103. Pass that tag as revision to access its original tables and dataset card. The current seven-model release is tagged seven-model-20261110.
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}
}
- Downloads last month
- 242