ClaasBeger/ConceptARC_Rule_Annotations

ConceptARC Rule Annotations

Model outputs, natural-language rules and human judgements of those rules on the 480 tasks of the ConceptARC benchmark. This is the data behind the paper Distinguishing Performance From Competence in Evaluations of Humanlike Abstract Reasoning (NeurIPS 2026, Evaluations and Datasets Track).

• Paper: arXiv:2510.02125
• Project page and interactive viewer: claasbeger.github.io/performance-competence-gap
• Authors: Claas Beger, Ryan Yi, Shuhao Fu, Kaleda Denton, Arseny Moskvichev, Sarah W. Tsai, Sivasankaran Rajamanickam, Melanie Mitchell
• Contact: claasbeger@santafe.edu

Summary

Accuracy alone can overstate or understate how well a model reasons with the abstractions a benchmark was designed to test. For each ConceptARC task we collected the model's output grid together with the natural-language rule it stated, and a team of annotators judged whether that rule captures the intended abstraction. The same judgements were made for rules written by human participants.

Models were evaluated with textual and visual inputs, at low and medium reasoning effort, and with and without Python tools. ConceptARC covers 16 spatial and semantic concepts with 10 tasks per concept, and each task has 3 test inputs (480 test items in total).

Contents

Subset / file	Rows	Contents
model_evaluations (evaluation_rows.parquet, .csv)	7,680	One row per model attempt on one test input: 16 runs × 480 items. Covers o3 (low and medium effort, with and without tools) and Claude Sonnet 4 and Gemini 2.5 Pro (medium effort, with and without tools), each with textual and visual inputs.
human_rule_annotations (human_rule_annotations.parquet, .csv)	3,049	One row per human participant rule, from the study of Moskvichev et al. (2023). Rules were only collected for correct outputs.
ground_truth_rules (ground_truth_rules.parquet, .csv)	160	The intended rule for each ConceptARC task, as revised by the authors. Columns: concept, puzzle, rule.
corpus/	160 files	ConceptARC task definitions (training demonstrations and test inputs and outputs), one JSON file per task under corpus/<Concept>/<Task>.json.

Model evaluation columns

Column	Description
modality	Textual (grids given as numbers) or Visual (grids given as images).
model	o3, claude-sonnet-4 or gemini-2.5-pro.
run_folder	Run configuration, e.g. mediumeffort_autosummary_withtools (reasoning effort, and whether Python tools were available).
source_file	Original log file the row came from.
concept, puzzle, test_idx	Concept group, task (e.g. Center3) and test input (1–3). Join puzzle with corpus/<concept>/<puzzle>.json.
effort, tools_used	Reasoning effort setting, and whether the model actually called a tool.
answer	The model's predicted output grid.
is_correct, err	Whether answer exactly matches the ground-truth grid, and any parsing error. This is a grid check only and says nothing about the rule.
Rule	The rule the model stated for its solution.
summary	The model's reasoning summary as logged.
programming_calls	Code the model ran, when tools were enabled.
Rule_correct_label, Rule_correct	Human judgement of Rule (see below).
rule_evaluation_reasoning	The annotators' notes on the judgement, where recorded.
starred	1 if an annotator starred the row.

Human rule columns

Column	Description
concept_group, Task, Test	Concept, task file and test input (1–3).
VerbalDescription	The rule written by the participant.
Rule_correct_label, Rule_correct	Human judgement of the rule (see below).
starred	1 if an annotator starred the row.

Rule judgements

Rule_correct_label	Rule_correct	Meaning
Correct - Intended	1	The rule works on the demonstrations and captures the abstraction the task was designed to test.
Correct - Unintended	1	The rule works on the demonstrations but does not capture the intended abstraction (for example a surface-level shortcut).
Incorrect	0	The rule does not work on the demonstrations.
Nonresponsive	-1	No usable rule was given.
Unclear	-2	The rule was too unclear to judge confidently.
Literal	-3	Human rules only: the rule describes the literal process of producing the output grid (for example clicking "copy" or selecting one specific colour) rather than the transformation.

The paper's figures group Nonresponsive, Unclear and Literal together as "Not Classified". Judgements were made by the authors. One annotator gave an initial judgement for each item, and ambiguous cases were discussed as a group until consensus was reached.

Loading

from datasets import load_dataset

models = load_dataset("ClaasBeger/ConceptARC_Rule_Annotations", "model_evaluations", split="train")
humans = load_dataset("ClaasBeger/ConceptARC_Rule_Annotations", "human_rule_annotations", split="train")
rules = load_dataset("ClaasBeger/ConceptARC_Rule_Annotations", "ground_truth_rules", split="train")

Citation

@inproceedings{beger2026distinguishing,
  title     = {Distinguishing Performance From Competence in Evaluations of Humanlike Abstract Reasoning},
  author    = {Beger, Claas and Yi, Ryan and Fu, Shuhao and Denton, Kaleda and Moskvichev, Arseny and Tsai, Sarah W. and Rajamanickam, Sivasankaran and Mitchell, Melanie},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

Please also cite ConceptARC:

@article{moskvichev2023conceptarc,
  title   = {The {ConceptARC} Benchmark: Evaluating Understanding and Generalization in the {ARC} Domain},
  author  = {Moskvichev, Arseny and Odouard, Victor Vikram and Mitchell, Melanie},
  journal = {Transactions on Machine Learning Research},
  year    = {2023}
}