SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2106.03634 · ACL Findings · 2021

PROST : Physical Reasoning about Objects through Space and Time

Alessandro Roncone, Stéphane Aroca-Ouellette, Cory Paik, Katharina Kann

arXiv · PDF · Open in the Atlas

Code that ran

10 of 13 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (0 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 10 have no contract verdict; 0 failures of our instrument; 3 no recorded run), measured 2026-11-22; bodies inline only under a permissive licence as GitHub detects it. Not a reproduction of the paper's results.

"Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results. This record, dated, as JSON: /api/ran/2106.03634.json · badge: /api/ran/2106.03634.svg

Each repository's role is a label: a classifier's output recorded on the link, or, where no role is recorded, what the Papers with Code archive marked. The line under it names its basis. Neither says who wrote the code.

RepositoryRoleRan
nala-cub/prost canonical (classifier label)
classifier basis: census
10 ran · 3 no recorded run
FunctionStatusWhere it lives
get_model_type Ran
ran on a synthesized input
nala-cub/prost/src/baselines/results.py
code served (permissive licence) · get_code("3da1a72ed0d0e883")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

def get_model_type(x):
    if x == 'openai-gpt' or x.startswith('allenai-unifiedqa'):
        return x.split('-')[1]
    elif x.startswith('albert'):
        return ' '.join([x.split('-')[0], x.split('-')[-1]])
    else:
        return x.split('-')[0]

# Test cases to understand the function behavior
test_cases = [
    'openai-gpt',
    'allenai-unifiedqa-v2',
    'albert-base-v2',
    'bert-base-uncased',
]

# The function takes a single string parameter 'x' and processes it
# based on different model naming conventions

ARGS = {
    'x': 'bert-base-uncased'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_and_check_config_yml Ran
ran on a synthesized input
nala-cub/prost/src/prost/prost.py
code served (permissive licence) · get_code("404e639140047862")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import os
import yaml

# Create a temporary YAML file with valid config structure
temp_dir = tempfile.mkdtemp()
config_path = os.path.join(temp_dir, "config.yml")

config_data = {
    'configs': {
        'task1': {'param1': 'value1'},
        'task2': {'param2': 'value2'}
    }
}

with open(config_path, 'w') as f:
    yaml.dump(config_data, f)

ARGS = {
    'path': config_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets, yaml — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_preds Ran
ran on a synthesized input
nala-cub/prost/src/baselines/results.py
code served (permissive licence) · get_code("a410f0abdd3fb511")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
from pathlib import Path
import pandas as pd
import os

# Create a temporary directory structure with mock prediction files
tmpdir = tempfile.mkdtemp()
cachedir = Path(tmpdir)
preds_dir = cachedir / 'preds'
preds_dir.mkdir(parents=True, exist_ok=True)

# Create mock prediction CSV files
pred_data_1 = pd.DataFrame({
    'example_idx': [0, 1],
    'prediction': [0.5, 0.7]
})
pred_data_1.to_csv(preds_dir / 'run1_model_a_preds.csv', index=False)

pred_data_2 = pd.DataFrame({
    'example_idx': [0, 1],
    'prediction': [0.6, 0.8]
})
pred_data_2.to_csv(preds_dir / 'run2_model_b_preds.csv', index=False)

# Create a mock dataset directory structure
import sys
from unittest.mock import Mock, patch

# Mock the datasets.load_dataset to return a mock dataset
mock_ds = Mock()
mock_ds_data = pd.DataFrame({
    'example_idx': [0, 1],
    'text': ['sample text 1', 'sample text 2'],
    'label': [0, 1]
})
mock_ds.__getitem__ = Mock(return_value=mock_ds_data)
mock_ds.set_format = Mock()

# Patch datasets.load_dataset
datasets_patcher = patch('datasets.load_dataset', return_value=mock_ds)
datasets_patcher.start()

ARGS = {
    'cachedir': cachedir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
datasets, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
load_raw_preds Ran
ran on a synthesized input
nala-cub/prost/src/baselines/results.py
code served (permissive licence) · get_code("8b6c4bddb433b1c8")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned DataFrame. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import tempfile
import shutil
from pathlib import Path
import pandas as pd

# Create a temporary directory structure with mock prediction files
temp_dir = tempfile.mkdtemp()
cachedir = Path(temp_dir)
preds_dir = cachedir / 'preds'
preds_dir.mkdir(parents=True, exist_ok=True)

# Create sample CSV files for models that don't end with 'v1'
df1 = pd.DataFrame({
    'prediction': [0.1, 0.2, 0.3],
    'actual': [0, 1, 0]
})
df1.to_csv(preds_dir / 'model1_bert_preds.csv', index=False)

df2 = pd.DataFrame({
    'prediction': [0.4, 0.5],
    'actual': [1, 1]
})
df2.to_csv(preds_dir / 'model2_gpt_preds.csv', index=False)

# Create a CSV file for a model that ends with 'v1' (should be skipped)
df3 = pd.DataFrame({
    'prediction': [0.6, 0.7],
    'actual': [0, 0]
})
df3.to_csv(preds_dir / 'model3_roberta_v1_preds.csv', index=False)

ARGS = {
    'cachedir': cachedir
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type DataFrame — a DataFrame output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a DataFrame output the recorded ndim 2 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
make_enum_vars Ran
ran on a synthesized input
nala-cub/prost/src/prost/prost.py
code served (permissive licence) · get_code("8adad0c7719ff831")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned tuple. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from enum import IntEnum

def to_ek(s: str) -> str:
    return s.replace(' ', '').upper()

def make_enum_vars(k, vals):
    return (k, IntEnum(k, {to_ek(v): i for i, v in enumerate(vals)}))

# Setup code
ARGS = {
    'k': 'MyEnum',
    'vals': ['foo bar', 'baz qux', 'hello world']
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple — a tuple output has no array shape of its own; none is stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
normalize_answer Ran
ran on a synthesized input
nala-cub/prost/src/baselines/unifiedqa.py
code served (permissive licence) · get_code("77c174355b5aa8d2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned str. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re
import string
import toolz.curried as T

# The function takes a single parameter 's' which is a string
# The function normalizes text by:
# 1. Converting to lowercase
# 2. Removing punctuation
# 3. Fixing buggy characters
# 4. Removing articles (a, an, the)
# 5. Fixing whitespace

ARGS = {
    "s": "The quick brown Fox! Has some buggy^characters<and>articles."
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
toolz — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type str — a str output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a str output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
prep_example_prost_gcp Ran
ran on a synthesized input
nala-cub/prost/src/baselines/unifiedqa.py
code served (permissive licence) · get_code("e1cb1fa52f9ae046")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Construct a valid example dict for PROST
example = {
    'ex_question': 'What is the capital of France?',
    'A': 'Paris',
    'B': 'London',
    'C': 'Berlin',
    'D': 'Madrid',
    'label': 0,  # Points to option A (Paris)
    'context': 'France is a country in Western Europe.'
}

ARGS = {
    'example': example
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
score_unifiedqa Ran
ran on a synthesized input
nala-cub/prost/src/baselines/unifiedqa.py
code served (permissive licence) · get_code("effefd745752a218")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned float, ndim 0, dtype family float, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import re
import string
import toolz.curried as T

def normalize_answer(s: str) -> str:
    """Normalize UnifiedQA Generated Text [1]. """
    import re
    import string

    def remove_articles(text):
        return re.sub('\\b(a|an|the)\\b', ' ', text)

    def white_space_fix(text):
        return ' '.join(text.split())

    def fix_buggy_characters(str):
        return re.sub('[{}^\\\\`⁇<]', ' ', str)

    def remove_punc(text):
        exclude = set(string.punctuation)
        return ''.join((ch for ch in text if ch not in exclude))
    return T.pipe(s, str.lower, remove_punc, fix_buggy_characters, remove_articles, white_space_fix)

def score_unifiedqa(prediction: str, groundtruth: str) -> float:
    if prediction == groundtruth:
        return 3.0
    prediction = normalize_answer(prediction)
    groundtruth = normalize_answer(groundtruth)
    if prediction == groundtruth:
        return 2.0
    if ' ' in prediction or ' ' in groundtruth:
        prediction_split = prediction.split(' ')
        groundtruth_split = groundtruth.split(' ')
        overlap = list(set(prediction_split) & set(groundtruth_split))
        return len(overlap) / max(len(prediction_split), len(groundtruth_split))
    else:
        return 0.0

ARGS = {
    "prediction": "The quick brown fox jumps over the lazy dog",
    "groundtruth": "A quick brown fox jumps over a lazy dog"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
toolz — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type float, ndim 0, dtype family float, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
test_step Ran
ran on a synthesized input
nala-cub/prost/src/baselines/zs_runner.py
code served (permissive licence) · get_code("2d47fac6ddc5af35")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned dict. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, reduce, repeat

torch.manual_seed(0)

# Define a minimal model that supports both generative and non-generative modes
class MinimalModel(nn.Module):
    def __init__(self, vocab_size=256, hidden_size=4, model_type='bert'):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, hidden_size)
        self.lm_head = nn.Linear(hidden_size, vocab_size)
        self.config = type('Config', (), {'model_type': model_type})()
        self.vocab_size = vocab_size
        self.hidden_size = hidden_size
    
    def forward(self, input_ids, attention_mask=None):
        x = self.embedding(input_ids)
        logits = self.lm_head(x)
        return type('Output', (), {'logits': logits})()
    
    def generate(self, input_ids):
        batch_size = input_ids.shape[0]
        return torch.randint(0, self.vocab_size, (batch_size, 1))

# Test case 1: Causal model (generative=False, is_causal=True)
batch_size = 2
seq_len = 5
vocab_size = 256
hidden_size = 4

model_causal = MinimalModel(vocab_size=vocab_size, hidden_size=hidden_size, model_type='gpt')

input_ids_causal = torch.randint(0, vocab_size, (batch_size, seq_len))
attention_mask_causal = torch.ones((batch_size, seq_len))

model_input_causal = {
    'input_ids': input_ids_causal,
    'attention_mask': attention_mask_causal
}

output_causal = {}

# Test case 2: Non-causal model with T5 (is_causal=False, model_type='t5')
model_t5 = MinimalModel(vocab_size=vocab_size, hidden_size=hidden_size, model_type='t5')

# T5 mode expects logits with shape (batch, 3, vocab_size)
class T5MinimalModel(nn.Module):
    def __init__(self, vocab_size=256, hidden_size=4):
        super().__init__()
        self.embedding = nn.Embedding(vocab_size, hidden_size)
        self.lm_head = nn.Linear(hidden_size, vocab_size)
        self.config = type('Config', (), {'model_type': 't5'})()
    
    def forward(self, input_ids, attention_mask=None):
        x = self.embedding(input_ids)
        # Always output 3 tokens for T5
        x = x.unsqueeze(1).expand(-1, 3, -1, -1)
        x = rearrange(x, 'b t s h -> (b t s) h')
        logits = self.lm_head(x)
        logits = rearrange(logits, '(b t s) v -> b t v', b=input_ids.shape[0], t=3)
        return type('Output', (), {'logits': logits})()

model_t5 = T5MinimalModel(vocab_size=vocab_size, hidden_size=hidden_size)

input_ids_t5 = torch.randint(0, vocab_size, (batch_size, 1))
attention_mask_t5 = torch.ones((batch_size, 1))

model_input_t5 = {
    'input_ids': input_ids_t5,
    'attention_mask': attention_mask_t5
}

num_options = 3
option_input_ids_t5 = torch.randint(0, vocab_size, (batch_size, num_options))

output_t5 = {
    'option_input_ids': option_input_ids_t5,
    'example_idx': torch.arange(batch_size),
    'target_tok_idx': torch.zeros(batch_size, dtype=torch.long)
}

# Test case 3: Non-causal model with non-T5 (is_causal=False, model_type='bert')
model_bert = MinimalModel(vocab_size=vocab_size, hidden_size=hidden_size, model_type='bert')

seq_len_bert = 4
input_ids_bert = torch.randint(0, vocab_size, (batch_size, seq_len_bert))
attention_mask_bert = torch.ones((batch_size, seq_len_bert))

model_input_bert = {
    'input_ids': input_ids_bert,
    'attention_mask': attention_mask_bert
}

option_input_ids_bert = torch.randint(0, vocab_size, (batch_size, num_options))

output_bert = {
    'option_input_ids': option_input_ids_bert,
    'example_idx': torch.arange(batch_size),
    'target_tok_idx': torch.randint(0, seq_len_bert, (batch_size,))
}

# Build ARGS for the curried function - using causal case as primary
ARGS = {
    'model': model_causal,
    'is_generative': False,
    'is_causal': True,
    'output': output_causal,
    'model_input': model_input_causal
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, einops, toolz, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type dict — a dict output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a dict output the recorded ndim 0 and dtype family float describe that coercion, not the output, and are not stated
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
try_get_index Ran
ran on a synthesized input
nala-cub/prost/src/baselines/input_pipeline.py
code served (permissive licence) · get_code("1ebbf7526d810f87")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned int, ndim 0, dtype family int, shape () as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np
from absl import logging

torch.manual_seed(0)

# The function try_get_index takes:
# - tokens: a sequence that supports .index() method (list of strings)
# - stoken: a string token to search for
# - s: start index (default 0)
# - e: end index (default -1)

ARGS = {
    "tokens": ["hello", "world", "foo", "bar", "baz"],
    "stoken": "world",
    "s": 0,
    "e": -1,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
absl, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type int, ndim 0, dtype family int, shape () — shape () as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
build_examples_from_config Not yet run
unverified (harvested, no recorded run)
nala-cub/prost/src/prost/prost.py
code served (permissive licence) · get_code("9ad5fc60ea43e8f1")
How it ran

no invocation recorded.

create_examples_t5 Not yet run
unverified (harvested, no recorded run)
nala-cub/prost/src/baselines/input_pipeline.py
code served (permissive licence) · get_code("774ce0489997b06c")
How it ran

no invocation recorded.

hf_auto_configure Not yet run
unverified (harvested, no recorded run)
nala-cub/prost/src/baselines/configs.py
code served (permissive licence) · get_code("b8695e1d11efba89")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 13 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

We present a new probing dataset named PROST: Physical Reasoning about Objects Through Space and Time. This dataset contains 18,736 multiple-choice questions made from 14 manually curated templates, covering 10 physical reasoning concepts. All questions are designed to probe both causal and masked language models in a zero-shot setting. We conduct an extensive analysis which demonstrates that state-of-the-art pretrained models are inadequate at physical reasoning: they are influenced by the order in which answer options are presented to them, they struggle when the superlative in a question is inverted (e.g., most ↔ least), and increasing the amount of pretraining data and parameters only yields minimal improvements. These results provide support for the hypothesis that current pretrained models' ability to reason about physical interactions is inherently limited by a lack of real world experience. By highlighting these limitations, we hope to motivate the development of models with a human-like understanding of the physical world. * *Email has no accent, but includes the hyphen A person drops a glass, a pillow, a coin, and a pen from a balcony. The [MASK] is most likely to break. Attribute Objects mass leaf, coin, egg, apple, brick, microwave height book, microwave, table, car, house, mountain circ book, microwave, table, car, house, mountain Affordances Affordance Objects stack books, blocks, boxes, coins, plates no stack balls, bottles, eggs, flowers, lamps roll apple, ball, bottle, egg, can no roll book, box, block, mirror, microwave grasp balls, blocks, books, bottles, flowers no grasp flour, rice, salt, snow, sugar break bottle, egg, glass, mirror, plate no break ball, coin, pen, pillow, shirt

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2106.03634")
get_code_for_paper("2106.03634")
have("2106.03634")

The run record, dated, one paper per request, free:

curl https://syntology.ai/api/ran/2106.03634.json

A badge for a README (the split and the date, never a ratio):

[![Syntology run record](https://syntology.ai/api/ran/2106.03634.svg)](https://syntology.ai/paper/2106.03634)

Connect an agent — have() is free.