BabyTheorist

BabyTheorist is a benchmark for observational theory learning: acquiring reusable theories from pure observations and applying them to new situations. It generates continual streams of before–after visual observations on 15×15 color grids from hidden compositional programs. During training, learners see only the resulting transitions, with no program labels, task identities, or support sets.

• Paper: BabyTheorist: A Benchmark for Learning to Theorize the World from Observation Alone (NeurIPS 2026, Evaluations and Datasets Track)
• Code (generator, environment, evaluation): https://github.com/ahn-ml/babytheorist
• Archived copy with persistent DOI: https://doi.org/10.7910/DVN/H9UGTL (Harvard Dataverse). The archives/ folder here holds the same tar.gz files; verify them with SHA256SUMS.txt.

Dataset structure

Four benchmark suites, each with eight curriculum levels. Each level has a training split and an evaluation split.

Suite (folder)	Name in the paper	Objects per grid	Selectors	Primitives
single_a	Single-A	one	by_index	8 base primitives (rot90, flip_h, flip_v, next_color, translate_{up,down,left,right}), fixed across levels
single_b	Single-B	one	by_index	8 base primitives, plus hollow (from level 2), fill_interior (level 4) and add_border (level 6)
multi_a	Multi-A	several	by_index	11 primitives (base plus swap_colors, align_row, align_col), fixed across levels
multi_b	Multi-B	several	ALL, by_index, by_color, is_hollow, topmost, bottommost, leftmost, rightmost	11 primitives, plus hollow (level 2), fill_interior (level 4) and add_border (level 6)

single_a/  lv1_train.h5  lv1_train.info.json  lv1_eval.h5  lv1_eval.info.json  ...  lv8_eval.info.json
single_b/  multi_a/  multi_b/        (same layout, 32 files each)
sample/{single_a,single_b,multi_a,multi_b}/   reduced inspection copy: 128 train / 32 eval tasks per level
archives/  BabyTheorist-{single_a,single_b,multi_a,multi_b,sample}-v0.5.5.tar.gz, SHA256SUMS.txt

Splits per level

• lvL_train.h5: 250,000 single transitions (one before–after pair each). These form the observation stream for level L.
• lvL_eval.h5: 10,000 tasks. Each task has 2–5 support pairs generated by one hidden program and one query pair. Single suites use 2 support pairs.

HDF5 fields

All grids are int8 color indices 0–9 on a 15×15 canvas. The pair axis P is 1 for train. For eval it is padded to max_support + 1: entries 0..n_support-1 are support pairs, entry n_support is the query, and later entries are zero padding.

Field	Shape / type	Meaning
inputs	(N, P, 15, 15) int8	grids before the transformation
outputs	(N, P, 15, 15) int8	grids after the transformation
grid_sizes	(N, 2) int32	grid height and width (15, 15 in v0.5.5)
n_shapes	(N, P) int32	number of objects in each input grid
depths	(N,) int32	depth (number of steps) of the hidden program
transformation_suites	(N,) JSON bytes	the hidden program: list of {selector, primitive, target} steps. Provided for analysis only; learners should not see it during training
n_support	(N,) int32	eval only: number of support pairs
valid_query_outputs	(N,) JSON bytes	eval only: every query output accepted as correct

Each .info.json sidecar records the generation configuration of its file: level, task count, primitive and selector vocabulary, the primitive introduced at that level, depth distribution parameters and empirical counts, frequent-subprogram pool settings, filter rule versions, and the generator commit.

Generation

• Programs are compositions of grid primitives (for example rot90, flip_h, next_color, translate_up, hollow) applied to objects chosen by selectors.
• Curriculum. Level L centers the program depth prior at L+1, so programs grow longer across levels. In the B suites, levels 2, 4 and 6 also introduce a new primitive. The prior is a truncated piecewise-geometric distribution; its parameters are recorded in info.json.
• Frequent subprograms. The generator keeps hidden per-depth pools of frequent subprograms. A depth-d program is either taken from the depth-d pool or built by chaining shorter pool entries, so the same subprograms recur across observations of different complexity. In the B suites, pools containing the newly introduced primitive join the inherited library.
• Filtering. A rule-based filter (filter.rules.v3) screens out degenerate programs whose observed transformation admits a shorter explanation, for example redundant or mutually cancelling steps. This keeps program depth tied to transformation complexity. The paper reports the filter's reliability against an independent breadth-first-search oracle.
• Evaluation tasks are generated with the unique_support policy recorded in info.json. valid_query_outputs lists every query output accepted as correct.
• Settings. Generator version v0.5.5, commit ae071ad, frequent-subprogram pool seed 42. The full procedure, including the continual train–test protocol with level skipping and the 80% pass threshold, is in the paper and the code repository.

Usage

from huggingface_hub import hf_hub_download
import h5py, json

path = hf_hub_download("ahn-ml/babytheorist", "single_a/lv1_eval.h5", repo_type="dataset")
with h5py.File(path) as f:
    i = 0
    k = int(f["n_support"][i])
    support_in, support_out = f["inputs"][i, :k], f["outputs"][i, :k]
    query_in = f["inputs"][i, k]
    program = json.loads(f["transformation_suites"][i])

To download a whole suite: hf download ahn-ml/babytheorist --repo-type dataset --include "single_a/*" --local-dir babytheorist. For the official curriculum loop (PARCgym), see the code repository.

License

The dataset is released under CC0 1.0. The generator and evaluation code are released under the MIT License in the code repository.

Citation

@misc{baek2026babytheoristbenchmarklearningtheorize,
      title={BabyTheorist: A Benchmark for Learning to Theorize the World from Observation Alone}, 
      author={Doojin Baek and Junyeob Baek and Mingyu Jo and Hosung Lee and Taegu Kang and Gyubin Lee and Sungjin Ahn},
      year={2026},
      eprint={2611.08177},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2611.08177}, 
}

Doojin Baek and Junyeob Baek contributed equally. Contact: Sungjin Ahn (sungjin.ahn@kaist.ac.kr).