SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2303.17491 · 2023

Language Models can Solve Computer Tasks

arXiv · PDF · Open in the Atlas

Code that ran

11 of 11 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 11 have no contract verdict; 0 failures of our instrument; 0 no recorded run), measured 2026-11-22; repository links from the Papers with Code archive attributed under CC BY-SA, no affiliation; 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/2303.17491.json · badge: /api/ran/2303.17491.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
posgnu/rci-agent canonical (classifier label)
classifier basis not recorded on the link
11 ran · 0 no recorded run
FunctionStatusWhere it lives
extract_click_checkboxes Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/fields.py
code served (permissive licence) · get_code("7f932a7eac5a840a")
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 Fields. 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 collections
import re

class Fields(object):
    """Wrapper around a dict to make it immutable.

    Args:
        d (dict): the wrapped dict.
    """

    def __init__(self, d):
        self._d = collections.OrderedDict(sorted(d.items()))
        if not self._d:
            self._d['dummy'] = 'dummy'

    def __getitem__(self, key):
        return self._d[key]

    def __len__(self):
        return len(self._d)

    @property
    def keys(self):
        return self._d.keys()

    @property
    def values(self):
        return self._d.values()

    def __repr__(self):
        return '\n'.join(('{}: {}'.format(k, repr(v)) for k, v in self._d.items()))
    __str__ = __repr__

def extract_click_checkboxes(utterance):
    targets = re.match('Select (.*) and click Submit\\.', utterance).group(1)
    if targets == 'nothing':
        targets = []
    else:
        targets = re.split(', ?', targets)
    fields = dict(zip(['target {}'.format(i) for i in range(len(targets))], targets))
    fields['button'] = 'submit'
    return Fields(fields)

# Setup code to construct valid arguments
# The function takes an utterance string that matches the pattern "Select ... and click Submit."
# It can be "Select nothing and click Submit." or "Select option1 and click Submit." or multiple options

ARGS = {
    'utterance': 'Select checkbox1, checkbox2 and click Submit.'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Fields — a Fields output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Fields 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
extract_click_checkboxes_soft Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/fields.py
code served (permissive licence) · get_code("09db1eb1eb3d66a6")
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 Fields. 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 collections
import re

class Fields(object):
    """Wrapper around a dict to make it immutable.

    Args:
        d (dict): the wrapped dict.
    """

    def __init__(self, d):
        self._d = collections.OrderedDict(sorted(d.items()))
        if not self._d:
            self._d['dummy'] = 'dummy'

    def __getitem__(self, key):
        return self._d[key]

    def __len__(self):
        return len(self._d)

    @property
    def keys(self):
        return self._d.keys()

    @property
    def values(self):
        return self._d.values()

    def __repr__(self):
        return '\n'.join(('{}: {}'.format(k, repr(v)) for k, v in self._d.items()))
    __str__ = __repr__

def extract_click_checkboxes_soft(utterance):
    targets = re.match('Select words similar to (.*) and click Submit\\.', utterance).group(1)
    targets = re.split(', ?', targets)
    fields = dict(zip(['target {}'.format(i) for i in range(len(targets))], targets))
    fields['button'] = 'submit'
    return Fields(fields)

# Setup valid utterance string that matches the expected pattern
utterance = 'Select words similar to apple, banana, cherry and click Submit.'

ARGS = {
    'utterance': utterance
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
none outside the standard library — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Fields — a Fields output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Fields 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
find_sublist Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/utils.py
code served (permissive licence) · get_code("7fd716eeb200413a")
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

torch.manual_seed(0)

# Analyzing the function:
# - find_sublist takes two parameters: l and sublist
# - Both are lists of objects
# - The function checks if elements are equal (==) and supports slicing
# - It compares l[i:i + len(sublist)] with sublist
# - This works with any comparable objects (ints, strings, etc.)

# Create test data: a list and a sublist to find
l = [1, 2, 3, 4, 5, 3, 4, 6]
sublist = [3, 4]

ARGS = {
    "l": l,
    "sublist": sublist
}
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 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
get_click_checkboxes_hard Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/reward.py
code served (permissive licence) · get_code("fd646d427848e7ee")
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 torch

# Define ARGS with a metadata dict that satisfies the function's requirements
ARGS = {
    'metadata': {
        'done': True,
        'raw_reward': 1.0
    }
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — 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
get_field_extractor Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/fields.py
code served (permissive licence) · get_code("5b2c9e8b14418640")
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 function. 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

FIELD_EXTRACTORS = {}

def get_field_extractor(task_name):
    try:
        return FIELD_EXTRACTORS[task_name]
    except KeyError:
        def extractor(utterance):
            raise ValueError('{} does not have a field extractor.'.format(task_name))
        return extractor

ARGS = {
    'task_name': 'nonexistent_task'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type function — a function output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a function 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
get_original_reward Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/reward.py
code served (permissive licence) · get_code("14e9be149aace800")
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 torch
import numpy as np

def setup_metadata():
    """Create a metadata dict with env_reward key"""
    metadata = {
        'env_reward': 42.5,
        'episode': 0,
        'timestep': 100
    }
    return metadata

ARGS = {
    'metadata': setup_metadata()
}
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 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
get_raw_reward Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/reward.py
code served (permissive licence) · get_code("ed9cdb15ddb55957")
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 torch
import numpy as np

torch.manual_seed(0)

# The function get_raw_reward takes a metadata dict and accesses metadata['raw_reward']
# It converts the value to float and returns it
# So we need to provide a dict with a 'raw_reward' key containing a numeric value

metadata = {
    'raw_reward': 1.0
}

ARGS = {
    'metadata': metadata
}
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 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
get_screenshot Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/screenshot.py
code served (permissive licence) · get_code("3973a30170a6113a")
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 Image. 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.

from unittest.mock import Mock
from io import BytesIO
from PIL import Image

# Create a mock Chrome WebDriver
mock_driver = Mock()

# Create a small test PNG image (1x1 pixel, RGB)
test_image = Image.new('RGB', (1, 1), color='red')
png_bytes = BytesIO()
test_image.save(png_bytes, format='PNG')
png_data = png_bytes.getvalue()

# Configure the mock driver to return PNG data
mock_driver.get_screenshot_as_png.return_value = png_data

ARGS = {
    'driver': mock_driver,
    'width': 160,
    'height': 210
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Image — a Image output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Image output the recorded ndim 3 and dtype family int 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
pil_to_numpy_array Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/screenshot.py
code served (permissive licence) · get_code("962335f24b95d8c0")
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 ndarray, ndim 3, dtype family float, shape (3, 4, 3) 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.

from PIL import Image
import numpy as np

# Create a small PIL image (4x3 RGB image)
pil_image = Image.new('RGB', (4, 3), color=(255, 128, 64))

ARGS = {
    'pil_image': pil_image
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type ndarray, ndim 3, dtype family float, shape (3, 4, 3) — shape (3, 4, 3) 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
strip_punctuation Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/utils.py
code served (permissive licence) · get_code("28f90015263219d5")
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 regex as re

def strip_punctuation(uni):
    """Strips punctuation from a unicode string. Returns the new unicode.

    Args:
        uni (unicode)

    Returns:
        unicode
    """
    return re.sub('\\p{P}+', ' ', uni)

ARGS = {
    "uni": "Hello, World! How are you?"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
regex — 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
strip_whitespace Ran
ran on a synthesized input
posgnu/rci-agent/computergym/computergym/miniwob/miniwob_interface/utils.py
code served (permissive licence) · get_code("9793a008041e04f6")
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 regex as re

def strip_whitespace(uni):
    """Strips all whitespace from a unicode string.

    Args:
        uni (unicode)

    Returns:
        unicode
    """
    return re.sub('\\s+', '', uni)

# Setup code
ARGS = {
    "uni": "hello   world\t\nfoo  \r\n  bar"
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
regex — 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

Also present in other papers' code

Of 11 samples lifted from this paper's linked repositories, 11 carry a function identity and 0 do not. Of the 11 distinct functions with an identity, 11 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 11 later) and 0 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.

FunctionAlso present inEarliest-datedOthers named
extract_click_checkboxes 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
extract_click_checkboxes_soft 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
find_sublist 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
get_click_checkboxes_hard 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
get_field_extractor 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
get_original_reward 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
get_raw_reward 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
get_screenshot 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
pil_to_numpy_array 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
strip_punctuation 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models
strip_whitespace 1 other paper
1 later
2305.16653 (2023-05)
AdaPlanner: Adaptive Planning from Feedback with Language Models

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

Agents capable of carrying out general tasks on a computer can improve efficiency and productivity by automating repetitive tasks and assisting in complex problem-solving. Ideally, such agents should be able to solve new computer tasks presented to them through natural language commands. However, previous approaches to this problem require large amounts of expert demonstrations and task-specific reward functions, both of which are impractical for new tasks. In this work, we show that a pre-trained large language model (LLM) agent can execute computer tasks guided by natural language using a simple prompting scheme where the agent Recursively Criticizes and Improves its output (RCI). The RCI approach significantly outperforms existing LLM methods for automating computer tasks and surpasses supervised learning (SL) and reinforcement learning (RL) approaches on the MiniWoB++ benchmark. We compare multiple LLMs and find that RCI with the InstructGPT-3+RLHF LLM is state-of-the-art on MiniWoB++, using only a handful of demonstrations per task rather than tens of thousands, and without a task-specific reward function. Furthermore, we demonstrate RCI prompting's effectiveness in enhancing LLMs' reasoning abilities on a suite of natural language reasoning tasks, outperforming chain of thought (CoT) prompting with external feedback. We find that RCI combined with CoT performs better than either separately. Our code can be found here: https://github.com/posgnu/rci-agent.

For agents

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

get_harvested_code_for_paper("2303.17491")
get_code_for_paper("2303.17491")
have("2303.17491")

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

curl https://syntology.ai/api/ran/2303.17491.json

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

[![Syntology run record](https://syntology.ai/api/ran/2303.17491.svg)](https://syntology.ai/paper/2303.17491)

Connect an agent — have() is free.