The Flexibility Trap — ICML 2026 reproduction
Collection
Scaled reproduction of arXiv:2601.15165 with H200 Jobs, raw outputs, released checkpoints, Trackio logbook, and poster. • 6 items • Updated
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
smoke: bool
model: string
n_problems: int64
n_samples: int64
sample_batch_size: int64
gen_length: int64
ao_block_length: int64
temperature: double
seed: int64
output: string
trackio_space: string
results_repo: string
run_name: string
paper_arxiv: string
paper_openreview: string
hardware: string
token_id: int64
entropy_ao_or_ar: double
relative_delay: double
token_text: string
generation_position: int64
connector: string
problem_index: int64
entropy_drop: double
sample_id: int64
absolute_position: int64
local_step: int64
global_step: int64
batch_row: int64
method: string
block: int64
entropy_ar_counterfactual: double
to
{'method': Value('string'), 'batch_row': Value('int64'), 'generation_position': Value('int64'), 'absolute_position': Value('int64'), 'block': Value('int64'), 'local_step': Value('int64'), 'global_step': Value('int64'), 'relative_delay': Value('float64'), 'token_id': Value('int64'), 'token_text': Value('string'), 'connector': Value('string'), 'entropy_ao_or_ar': Value('float64'), 'problem_index': Value('int64'), 'sample_id': Value('int64'), 'entropy_ar_counterfactual': Value('float64'), 'entropy_drop': 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
smoke: bool
model: string
n_problems: int64
n_samples: int64
sample_batch_size: int64
gen_length: int64
ao_block_length: int64
temperature: double
seed: int64
output: string
trackio_space: string
results_repo: string
run_name: string
paper_arxiv: string
paper_openreview: string
hardware: string
token_id: int64
entropy_ao_or_ar: double
relative_delay: double
token_text: string
generation_position: int64
connector: string
problem_index: int64
entropy_drop: double
sample_id: int64
absolute_position: int64
local_step: int64
global_step: int64
batch_row: int64
method: string
block: int64
entropy_ar_counterfactual: double
to
{'method': Value('string'), 'batch_row': Value('int64'), 'generation_position': Value('int64'), 'absolute_position': Value('int64'), 'block': Value('int64'), 'local_step': Value('int64'), 'global_step': Value('int64'), 'relative_delay': Value('float64'), 'token_id': Value('int64'), 'token_text': Value('string'), 'connector': Value('string'), 'entropy_ao_or_ar': Value('float64'), 'problem_index': Value('int64'), 'sample_id': Value('int64'), 'entropy_ar_counterfactual': Value('float64'), 'entropy_drop': 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.
method string | batch_row int64 | generation_position int64 | absolute_position int64 | block int64 | local_step int64 | global_step int64 | relative_delay float64 | token_id int64 | token_text string | connector string | entropy_ao_or_ar float64 | problem_index int64 | sample_id int64 | entropy_ar_counterfactual float64 | entropy_drop float64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
AO | 0 | 3 | 79 | 0 | 0 | 0 | 0 | 16 | 1 | null | 3.417109 | 0 | 0 | null | null |
AO | 0 | 4 | 80 | 0 | 1 | 1 | 0.032258 | 21 | 6 | null | 0.115252 | 0 | 0 | null | null |
AO | 0 | 2 | 78 | 0 | 2 | 2 | 0.064516 | 220 | null | 0.312683 | 0 | 0 | null | null | |
AO | 0 | 1 | 77 | 0 | 3 | 3 | 0.096774 | 723 | has | null | 1.154946 | 0 | 0 | null | null |
AO | 0 | 0 | 76 | 0 | 4 | 4 | 0.129032 | 117,603 | Janet | null | 0.764283 | 0 | 0 | null | null |
AO | 0 | 9 | 85 | 0 | 5 | 5 | 0.16129 | 220 | null | 2.922451 | 0 | 0 | null | null | |
AO | 0 | 6 | 82 | 0 | 6 | 6 | 0.193548 | 220 | null | 0.230733 | 0 | 0 | null | null | |
AO | 0 | 7 | 83 | 0 | 7 | 7 | 0.225806 | 18 | 3 | null | 0.070071 | 0 | 0 | null | null |
AO | 0 | 10 | 86 | 0 | 8 | 8 | 0.258065 | 19 | 4 | null | 0.096001 | 0 | 0 | null | null |
AO | 0 | 11 | 87 | 0 | 9 | 9 | 0.290323 | 373 | = | null | 0.189819 | 0 | 0 | null | null |
AO | 0 | 8 | 84 | 0 | 10 | 10 | 0.322581 | 537 | - | null | 0.133719 | 0 | 0 | null | null |
AO | 0 | 5 | 81 | 0 | 11 | 11 | 0.354839 | 537 | - | null | 0.055193 | 0 | 0 | null | null |
AO | 0 | 13 | 89 | 0 | 12 | 12 | 0.387097 | 24 | 9 | null | 0.360706 | 0 | 0 | null | null |
AO | 0 | 12 | 88 | 0 | 13 | 13 | 0.419355 | 220 | null | 0.007688 | 0 | 0 | null | null | |
AO | 0 | 14 | 90 | 0 | 14 | 14 | 0.451613 | 16,832 | eggs | null | 0.589377 | 0 | 0 | null | null |
AO | 0 | 15 | 91 | 0 | 15 | 15 | 0.483871 | 2,914 | left | null | 1.459708 | 0 | 0 | null | null |
AO | 0 | 19 | 95 | 0 | 16 | 16 | 0.516129 | 198 | null | 3.293597 | 0 | 0 | null | null | |
AO | 0 | 18 | 94 | 0 | 17 | 17 | 0.548387 | 13 | . | null | 0.085217 | 0 | 0 | null | null |
AO | 0 | 21 | 97 | 0 | 18 | 18 | 0.580645 | 31,310 | sells | null | 1.536795 | 0 | 0 | null | null |
AO | 0 | 20 | 96 | 0 | 19 | 19 | 0.612903 | 6,572 | She | null | 0.073024 | 0 | 0 | null | null |
AO | 0 | 22 | 98 | 0 | 20 | 20 | 0.645161 | 1,671 | each | null | 1.031217 | 0 | 0 | null | null |
AO | 0 | 26 | 102 | 0 | 21 | 21 | 0.677419 | 17 | 2 | null | 0.566362 | 0 | 0 | null | null |
AO | 0 | 23 | 99 | 0 | 22 | 22 | 0.709677 | 10,540 | egg | null | 0.015025 | 0 | 0 | null | null |
AO | 0 | 25 | 101 | 0 | 23 | 23 | 0.741935 | 558 | $ | null | 0.013468 | 0 | 0 | null | null |
AO | 0 | 24 | 100 | 0 | 24 | 24 | 0.774194 | 352 | for | null | 0.013628 | 0 | 0 | null | null |
AO | 0 | 27 | 103 | 0 | 25 | 25 | 0.806452 | 11 | , | null | 0.212453 | 0 | 0 | null | null |
AO | 0 | 28 | 104 | 0 | 26 | 26 | 0.83871 | 740 | so | so | 0.020286 | 0 | 0 | 0.019283 | -0.001003 |
AO | 0 | 29 | 105 | 0 | 27 | 27 | 0.870968 | 1,285 | she | null | 0.254571 | 0 | 0 | null | null |
AO | 0 | 31 | 107 | 0 | 28 | 28 | 0.903226 | 220 | null | 0.362186 | 0 | 0 | null | null | |
AO | 0 | 30 | 106 | 0 | 29 | 29 | 0.935484 | 3,524 | makes | null | 0.441426 | 0 | 0 | null | null |
AO | 0 | 17 | 93 | 0 | 30 | 30 | 0.967742 | 1,813 | day | null | 0.889108 | 0 | 0 | null | null |
AO | 0 | 16 | 92 | 0 | 31 | 31 | 1 | 1,380 | every | null | 0.846782 | 0 | 0 | null | null |
AO | 0 | 32 | 108 | 1 | 0 | 32 | 0 | 24 | 9 | null | 0.049385 | 0 | 0 | null | null |
AO | 0 | 35 | 111 | 1 | 1 | 33 | 0.032258 | 17 | 2 | null | 0.198033 | 0 | 0 | null | null |
AO | 0 | 36 | 112 | 1 | 2 | 34 | 0.064516 | 373 | = | null | 0.027001 | 0 | 0 | null | null |
AO | 0 | 38 | 114 | 1 | 3 | 35 | 0.096774 | 16 | 1 | null | 0.029315 | 0 | 0 | null | null |
AO | 0 | 39 | 115 | 1 | 4 | 36 | 0.129032 | 23 | 8 | null | 0.003087 | 0 | 0 | null | null |
AO | 0 | 37 | 113 | 1 | 5 | 37 | 0.16129 | 558 | $ | null | 0.011339 | 0 | 0 | null | null |
AO | 0 | 34 | 110 | 1 | 6 | 38 | 0.193548 | 558 | $ | null | 0.115635 | 0 | 0 | null | null |
AO | 0 | 41 | 117 | 1 | 7 | 39 | 0.225806 | 1,813 | day | null | 0.241772 | 0 | 0 | null | null |
AO | 0 | 40 | 116 | 1 | 8 | 40 | 0.258065 | 1,380 | every | null | 0.253577 | 0 | 0 | null | null |
AO | 0 | 46 | 122 | 1 | 9 | 41 | 0.290323 | 2,201 | market | null | 1.843135 | 0 | 0 | null | null |
AO | 0 | 43 | 119 | 1 | 10 | 42 | 0.322581 | 268 | the | null | 0.00004 | 0 | 0 | null | null |
AO | 0 | 45 | 121 | 1 | 11 | 43 | 0.354839 | 6 | ' | null | 0.003725 | 0 | 0 | null | null |
AO | 0 | 44 | 120 | 1 | 12 | 44 | 0.387097 | 19,149 | farmers | null | 0.005252 | 0 | 0 | null | null |
AO | 0 | 47 | 123 | 1 | 13 | 45 | 0.419355 | 13 | . | null | 0.004888 | 0 | 0 | null | null |
AO | 0 | 42 | 118 | 1 | 14 | 46 | 0.451613 | 482 | at | null | 0.007497 | 0 | 0 | null | null |
AO | 0 | 48 | 124 | 1 | 15 | 47 | 0.483871 | 198 | null | 0.025544 | 0 | 0 | null | null | |
AO | 0 | 63 | 139 | 1 | 16 | 48 | 0.516129 | 126,081 | <|endoftext|> | null | 0.194676 | 0 | 0 | null | null |
AO | 0 | 61 | 137 | 1 | 17 | 49 | 0.548387 | 126,081 | <|endoftext|> | null | 0.621129 | 0 | 0 | null | null |
AO | 0 | 62 | 138 | 1 | 18 | 50 | 0.580645 | 126,081 | <|endoftext|> | null | 0 | 0 | 0 | null | null |
AO | 0 | 60 | 136 | 1 | 19 | 51 | 0.612903 | 126,081 | <|endoftext|> | null | 0.146872 | 0 | 0 | null | null |
AO | 0 | 59 | 135 | 1 | 20 | 52 | 0.645161 | 126,081 | <|endoftext|> | null | 0.176683 | 0 | 0 | null | null |
AO | 0 | 58 | 134 | 1 | 21 | 53 | 0.677419 | 126,081 | <|endoftext|> | null | 0.2503 | 0 | 0 | null | null |
AO | 0 | 57 | 133 | 1 | 22 | 54 | 0.709677 | 126,081 | <|endoftext|> | null | 0.453297 | 0 | 0 | null | null |
AO | 0 | 52 | 128 | 1 | 23 | 55 | 0.741935 | 220 | null | 1.384185 | 0 | 0 | null | null | |
AO | 0 | 53 | 129 | 1 | 24 | 56 | 0.774194 | 16 | 1 | null | 0.000162 | 0 | 0 | null | null |
AO | 0 | 54 | 130 | 1 | 25 | 57 | 0.806452 | 23 | 8 | null | 0.00004 | 0 | 0 | null | null |
AO | 0 | 56 | 132 | 1 | 26 | 58 | 0.83871 | 126,081 | <|endoftext|> | null | 0.000502 | 0 | 0 | null | null |
AO | 0 | 55 | 131 | 1 | 27 | 59 | 0.870968 | 126,348 | <|eot_id|> | null | 0.00009 | 0 | 0 | null | null |
AO | 0 | 33 | 109 | 1 | 28 | 60 | 0.903226 | 500 | * | null | 0.594111 | 0 | 0 | null | null |
AO | 0 | 51 | 127 | 1 | 29 | 61 | 0.935484 | 25 | : | null | 1.304117 | 0 | 0 | null | null |
AO | 0 | 49 | 125 | 1 | 30 | 62 | 0.967742 | 1,870 | Con | null | 1.52094 | 0 | 0 | null | null |
AO | 0 | 50 | 126 | 1 | 31 | 63 | 1 | 89,772 | clusively | null | 0.013202 | 0 | 0 | null | null |
AO | 1 | 2 | 78 | 0 | 0 | 0 | 0 | 220 | null | 3.628681 | 0 | 1 | null | null | |
AO | 1 | 4 | 80 | 0 | 1 | 1 | 0.032258 | 21 | 6 | null | 1.253314 | 0 | 1 | null | null |
AO | 1 | 3 | 79 | 0 | 2 | 2 | 0.064516 | 16 | 1 | null | 0.004021 | 0 | 1 | null | null |
AO | 1 | 1 | 77 | 0 | 3 | 3 | 0.096774 | 723 | has | null | 1.154946 | 0 | 1 | null | null |
AO | 1 | 0 | 76 | 0 | 4 | 4 | 0.129032 | 117,603 | Janet | null | 0.764283 | 0 | 1 | null | null |
AO | 1 | 5 | 81 | 0 | 5 | 5 | 0.16129 | 16,832 | eggs | null | 1.409281 | 0 | 1 | null | null |
AO | 1 | 7 | 83 | 0 | 6 | 6 | 0.193548 | 220 | null | 2.011551 | 0 | 1 | null | null | |
AO | 1 | 8 | 84 | 0 | 7 | 7 | 0.225806 | 18 | 3 | null | 0.122297 | 0 | 1 | null | null |
AO | 1 | 9 | 85 | 0 | 8 | 8 | 0.258065 | 16,832 | eggs | null | 0.553022 | 0 | 1 | null | null |
AO | 1 | 11 | 87 | 0 | 9 | 9 | 0.290323 | 220 | null | 0.866019 | 0 | 1 | null | null | |
AO | 1 | 12 | 88 | 0 | 10 | 10 | 0.322581 | 19 | 4 | null | 0.062892 | 0 | 1 | null | null |
AO | 1 | 13 | 89 | 0 | 11 | 11 | 0.354839 | 16,832 | eggs | null | 0.082748 | 0 | 1 | null | null |
AO | 1 | 16 | 92 | 0 | 12 | 12 | 0.387097 | 24 | 9 | null | 0.263846 | 0 | 1 | null | null |
AO | 1 | 15 | 91 | 0 | 13 | 13 | 0.419355 | 220 | null | 0.003593 | 0 | 1 | null | null | |
AO | 1 | 14 | 90 | 0 | 14 | 14 | 0.451613 | 373 | = | null | 0.027889 | 0 | 1 | null | null |
AO | 1 | 17 | 93 | 0 | 15 | 15 | 0.483871 | 16,832 | eggs | null | 0.252727 | 0 | 1 | null | null |
AO | 1 | 10 | 86 | 0 | 16 | 16 | 0.516129 | 537 | - | null | 0.427565 | 0 | 1 | null | null |
AO | 1 | 6 | 82 | 0 | 17 | 17 | 0.548387 | 537 | - | null | 0.037956 | 0 | 1 | null | null |
AO | 1 | 18 | 94 | 0 | 18 | 18 | 0.580645 | 2,914 | left | null | 1.297682 | 0 | 1 | null | null |
AO | 1 | 20 | 96 | 0 | 19 | 19 | 0.612903 | 5,918 | sell | null | 1.453286 | 0 | 1 | null | null |
AO | 1 | 19 | 95 | 0 | 20 | 20 | 0.645161 | 297 | to | null | 0.001611 | 0 | 1 | null | null |
AO | 1 | 21 | 97 | 0 | 21 | 21 | 0.677419 | 13 | . | null | 0.93198 | 0 | 1 | null | null |
AO | 1 | 22 | 98 | 0 | 22 | 22 | 0.709677 | 198 | null | 0.038648 | 0 | 1 | null | null | |
AO | 1 | 23 | 99 | 0 | 23 | 23 | 0.741935 | 6,572 | She | null | 1.330847 | 0 | 1 | null | null |
AO | 1 | 24 | 100 | 0 | 24 | 24 | 0.774194 | 31,310 | sells | null | 0.081476 | 0 | 1 | null | null |
AO | 1 | 29 | 105 | 0 | 25 | 25 | 0.806452 | 17 | 2 | null | 1.075108 | 0 | 1 | null | null |
AO | 1 | 28 | 104 | 0 | 26 | 26 | 0.83871 | 558 | $ | null | 0.012228 | 0 | 1 | null | null |
AO | 1 | 27 | 103 | 0 | 27 | 27 | 0.870968 | 352 | for | null | 0.046932 | 0 | 1 | null | null |
AO | 1 | 25 | 101 | 0 | 28 | 28 | 0.903226 | 1,671 | each | null | 0.115352 | 0 | 1 | null | null |
AO | 1 | 26 | 102 | 0 | 29 | 29 | 0.935484 | 10,540 | egg | null | 0.002797 | 0 | 1 | null | null |
AO | 1 | 30 | 106 | 0 | 30 | 30 | 0.967742 | 11 | , | null | 0.124038 | 0 | 1 | null | null |
AO | 1 | 31 | 107 | 0 | 31 | 31 | 1 | 740 | so | so | 0.005542 | 0 | 1 | 0.005542 | 0 |
AO | 1 | 32 | 108 | 1 | 0 | 32 | 0 | 1,285 | she | null | 0.121299 | 0 | 1 | null | null |
AO | 1 | 35 | 111 | 1 | 1 | 33 | 0.032258 | 24 | 9 | null | 0.072586 | 0 | 1 | null | null |
AO | 1 | 34 | 110 | 1 | 2 | 34 | 0.064516 | 220 | null | 0.067469 | 0 | 1 | null | null | |
AO | 1 | 39 | 115 | 1 | 3 | 35 | 0.096774 | 17 | 2 | null | 0.107838 | 0 | 1 | null | null |
End of preview.
The Flexibility Trap — reproduction results
Raw and summarized outputs for a scaled reproduction of arXiv:2601.15165 / OpenReview kpgURPRMGf.
Result folders
claims-1-3/claims-1-3-h200-p8-s4-g64-r7-local-download: 8 evenly spaced GSM8K problems, 4 stochastic samples per AO/AR order, 64 generated tokens, 41 paired logical-fork events.claims-4-5/claims-4-5-h200-p12-b4-g256-r7-local-download: 12 evenly spaced GSM8K problems, base and released JustGRPO checkpoint, 1/2/4/8 confidence-ranked tokens per call, 256 generated tokens.
Every folder contains run_config.json, summary.json, raw JSONL, CSV metrics, and standalone Plotly HTML.
Validated headline results
- Claim 1: not reproduced at this scale; AO Pass@1/2/4 = 0.750/0.875/1.000, AR = 0.781/0.875/1.000.
- Claim 2: compatible but non-discriminative; all 8 problems were solved by both orders.
- Claim 3: reproduced in mean; AR-counterfactual minus AO entropy = +0.229 nats, bootstrap 95% CI [0.054, 0.402], n=41.
- Claim 4: released checkpoint 11/12 = 91.7% at strict one-token decoding; 89.1% lies inside the Wilson interval. Training was not repeated.
- Claim 5: proxy reproduced; JustGRPO-minus-base gain rose from 0.0 points at 1 token/call to +58.3 points at 8 tokens/call.
Provenance
- Claims 1–3 H200 job: https://huggingface.co/jobs/jgallego9/6a57c4b2b1669a49bf075a9c
- Claims 4–5 H200 job: https://huggingface.co/jobs/jgallego9/6a57c4bc85d9643ce16d5314
- Base checkpoint:
GSAI-ML/LLaDA-8B-Instruct - Tuned checkpoint:
nzl-thu/LLaDA-Instruct-JustGRPO-GSM8K - Official-code audit commit:
1a2fddb5c6655597e63081c0af5ebb718a849f39 - Independent validation: 64/64 and 96/96 expected raw rows; both summaries recomputed exactly.
Download
hf download jgallego9/flexibility-trap-repro-results --repo-type dataset --local-dir flexibility-trap-results
See the published Trackio logbook for claim-by-claim interpretation, limitations, costs, poster, and artifact bundle.
- Downloads last month
- 10