Learning to Theorize the World from Observation
Paper • 2605.03413 • Published • 2
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
steps: int64
id: struct<transfer_correct: double, equals_support_answer: double, equals_query_input: double, chance: (... 63 chars omitted)
child 0, transfer_correct: double
child 1, equals_support_answer: double
child 2, equals_query_input: double
child 3, chance: double
child 4, n: int64
child 5, distinct_predictions: int64
child 6, self_ex: double
comp_ood: struct<transfer_correct: double, equals_support_answer: double, equals_query_input: double, chance: (... 63 chars omitted)
child 0, transfer_correct: double
child 1, equals_support_answer: double
child 2, equals_query_input: double
child 3, chance: double
child 4, n: int64
child 5, distinct_predictions: int64
child 6, self_ex: double
length_ood: struct<transfer_correct: double, equals_support_answer: double, equals_query_input: double, chance: (... 63 chars omitted)
child 0, transfer_correct: double
child 1, equals_support_answer: double
child 2, equals_query_input: double
child 3, chance: double
child 4, n: int64
child 5, distinct_predictions: int64
child 6, self_ex: double
vae_seed0: list<item: struct<step: int64, rec: double, kl: double>>
child 0, item: struct<step: int64, rec: double, kl: double>
child 0, step: int64
child 1, rec: double
child 2, kl: double
gridworld|a=0.33|wrap|base|seed0: list<item: struct<step: int64, rec: double, vq: double, gnd: double, mean_len: double>>
child 0, item: struct<step: int64, rec: double, vq: double, gnd: double, mean_len: double>
child 0, step: int64
child 1, rec: double
child 2, vq: double
child 3, gnd: double
child 4, mean_len: double
to
{'vae_seed0': List({'step': Value('int64'), 'rec': Value('float64'), 'kl': Value('float64')}), 'gridworld|a=0.33|wrap|base|seed0': List({'step': Value('int64'), 'rec': Value('float64'), 'vq': Value('float64'), 'gnd': Value('float64'), 'mean_len': Value('float64')})}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1816, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
steps: int64
id: struct<transfer_correct: double, equals_support_answer: double, equals_query_input: double, chance: (... 63 chars omitted)
child 0, transfer_correct: double
child 1, equals_support_answer: double
child 2, equals_query_input: double
child 3, chance: double
child 4, n: int64
child 5, distinct_predictions: int64
child 6, self_ex: double
comp_ood: struct<transfer_correct: double, equals_support_answer: double, equals_query_input: double, chance: (... 63 chars omitted)
child 0, transfer_correct: double
child 1, equals_support_answer: double
child 2, equals_query_input: double
child 3, chance: double
child 4, n: int64
child 5, distinct_predictions: int64
child 6, self_ex: double
length_ood: struct<transfer_correct: double, equals_support_answer: double, equals_query_input: double, chance: (... 63 chars omitted)
child 0, transfer_correct: double
child 1, equals_support_answer: double
child 2, equals_query_input: double
child 3, chance: double
child 4, n: int64
child 5, distinct_predictions: int64
child 6, self_ex: double
vae_seed0: list<item: struct<step: int64, rec: double, kl: double>>
child 0, item: struct<step: int64, rec: double, kl: double>
child 0, step: int64
child 1, rec: double
child 2, kl: double
gridworld|a=0.33|wrap|base|seed0: list<item: struct<step: int64, rec: double, vq: double, gnd: double, mean_len: double>>
child 0, item: struct<step: int64, rec: double, vq: double, gnd: double, mean_len: double>
child 0, step: int64
child 1, rec: double
child 2, vq: double
child 3, gnd: double
child 4, mean_len: double
to
{'vae_seed0': List({'step': Value('int64'), 'rec': Value('float64'), 'kl': Value('float64')}), 'gridworld|a=0.33|wrap|base|seed0': List({'step': Value('int64'), 'rec': Value('float64'), 'vq': Value('float64'), 'gnd': Value('float64'), 'mean_len': Value('float64')})}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1683, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1869, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
vae_seed0 list | gridworld|a=0.33|wrap|base|seed0 list |
|---|---|
[
{
"step": 0,
"rec": 4.605195045471191,
"kl": 0.03322009742259979
},
{
"step": 600,
"rec": 0.00005003531987313181,
"kl": 76.4076156616211
},
{
"step": 1200,
"rec": 0.000015073124814080074,
"kl": 84.6890869140625
},
{
"step": 1800,
"rec": 0.000001014515532915538... | [
{
"step": 0,
"rec": 156.0826416015625,
"vq": 1.7894285917282104,
"gnd": 3.9360008308175765e-14,
"mean_len": 4
},
{
"step": 2000,
"rec": 1.4900479316711426,
"vq": 1.5738171339035034,
"gnd": 1.3279951810836792,
"mean_len": 3.5234375
},
{
"step": 4000,
"rec": 0.2... |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Independent reproduction — Learning to Theorize the World from Observation
Reproduction of NEO (Neural Theorizer) and the OTIB benchmark from
arXiv:2605.03413v2 (ICML 2026,
OpenReview wsA8LgHU5U),
for the ICML 2026 Agent Reproduction Challenge.
Everything here is written from the paper text alone — no author code was available or consulted.
Layout
protocol.md pre-registered plan, derivations, success/falsification criteria
src/otib.py OTIB benchmark: GridWorld + Arithmetic Factorization
src/models.py NEO, Disc-Mono, Cont-Mono, Cont-Mono-Opt, shared CNN VAE
src/run.py consolidated driver: pretrain -> train -> evaluate -> JSON
src/figures.py builds every figure from stored results
tests/ unit tests, incl. mechanism audits for claims C1-C3
configs/ exact configs used for the local smoke run and the HF Job
outputs/ raw + aggregate results (JSON)
figures/ figures and their underlying data
logs/ job stdout
One-command rerun
pip install torch numpy plotly
python src/run.py --domain gridworld --alpha 0.33 --ablations \
--steps 6000 --vae-steps 1500 --n-eval 2000 --seeds 0 1 2 \
--out outputs/gridworld.json
python src/figures.py
--domain arithmetic --alpha 0.33 reproduces the Arithmetic Factorization
results. Every number in the logbook comes from outputs/*.json; the figures
are regenerated from those files by src/figures.py.
Tests
python -m pytest tests/ -q
Key implementation notes
- Boundary semantics are an assumption. The paper does not say what a
primitive does at a wall; we use toroidal wrap (the only choice under which
the benchmark is well posed) and run clamping as a control. See
protocol.md§1.2. - The 24-program short-program space is derived, not copied — see
protocol.md§1.1 andtests/test_otib.py::test_derived_short_program_space_matches_paper. - Reconstruction uses categorical cross-entropy, not pixel-wise BCE: observations are one-hot, and CE is the likelihood whose arg-max is the paper's exact-match metric.
- Conditioning hygiene was required: the pretrained VAE latent has scale ‖s‖≈19, which makes an unnormalised FiLM layer diverge. We LayerNorm the FiLM input/conditioning streams, normalise the policy output before quantisation, and identity-initialise the execution network. These are stability fixes, not method changes, and are documented in the logbook.
- Downloads last month
- 37