SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2404.06474 · 2024

Autonomous Evaluation and Refinement of Digital Agents

arXiv · PDF · Open in the Atlas

Code that ran

9 of 9 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; 1 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 7 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/2404.06474.json · badge: /api/ran/2404.06474.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
berkeley-nlp/agent-eval-refine canonical (classifier label)
classifier basis: census
9 ran · 0 no recorded run
FunctionStatusWhere it lives
built_user_prompt_v1 Ran
ran; our contract draft was wrong, not the code
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/eval/annotator.py
code served (permissive licence) · get_code("dd383df4089af65c")
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

torch.manual_seed(0)

# Analyzing the function: built_user_prompt_v1
# Parameters: intent, current_state, action, next_state
# The function uses these parameters in an f-string, so they need to be convertible to strings
# No methods or attributes are called on them - they're just formatted into a string

ARGS = {
    "intent": "Navigate to the kitchen",
    "current_state": "Robot is in the living room. Door to kitchen is open. No obstacles detected.",
    "action": "move_forward(2.0)",
    "next_state": "Robot is now in the kitchen. Cabinet doors are visible. Refrigerator is on the right side."
}
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
calculate_performance Ran
ran, honoured the contract we drafted
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/eval/metrics.py
code served (permissive licence) · get_code("f980bae9046c60f8")
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, ndim 1, dtype family int; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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 pandas as pd
import numpy as np

# Create a simple dataframe where each column contains a tuple of (ground_truth, prediction)
# Both values should be boolean-like for the function logic to work
data = {
    'col1': (True, True),      # TP case
    'col2': (False, False),    # TN case
    'col3': (False, True),     # FP case
    'col4': (True, False),     # FN case
}

dataframe = pd.DataFrame(data)

ARGS = {
    'dataframe': dataframe
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, pandas — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family int — a tuple output has no array shape of its own; none is stated
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
extract_eval_results Ran
ran; our contract draft was wrong, not the code
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/domains/webarena.py
code served (permissive licence) · get_code("3b9d781f8cd9b1a1")
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 re

def extract_eval_results(merged_log: str):
    """Extract the evaluation results from the merged log file."""
    results = {}
    for line in merged_log.splitlines():
        if '[Result]' in line:
            result_status = 'PASS' if '(PASS)' in line else 'FAIL'
            match = re.search('/(\\d+)\\.json', line)
            if match:
                index = match.group(1)
                results[index] = result_status == 'PASS'
    return results


merged_log = """
[Result] Test 1: /0.json (PASS)
[Result] Test 2: /1.json (FAIL)
[Result] Test 3: /2.json (PASS)
Some other log line
[Result] Another test: /123.json (PASS)
"""

ARGS = {
    "merged_log": merged_log
}
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 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
extract_intent_from_text Ran
ran; our contract draft was wrong, not the code
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/domains/webarena.py
code served (permissive licence) · get_code("d31dda6a94c3338c")
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 torch.nn as nn

# Construct valid arguments for extract_intent_from_text
html_content = """
<div>
    <p>Some content here</p>
    <span>intent: Find the nearest restaurant</span>
    <p>More content</p>
</div>
"""

ARGS = {
    'html_content': html_content
}
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 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
get_metrics_from_result_json Ran
ran; our contract draft was wrong, not the code
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/eval/metrics.py
code served (permissive licence) · get_code("1e4d66c87eab208b")
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, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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 json
import tempfile
import os

# Create a temporary JSON file with test data
temp_dir = tempfile.mkdtemp()
result_json_path = os.path.join(temp_dir, 'result.json')

test_data = {
    '1': {'gt': True, 'rm': True},      # TP
    '2': {'gt': False, 'rm': False},    # TN
    '3': {'gt': False, 'rm': True},     # FP
    '4': {'gt': True, 'rm': False},     # FN
    '24': {'gt': True, 'rm': False},    # FN (but gt will be corrected to False by LABEL_CORRECTION)
    '5': {'gt': True, 'rm': True},      # TP
    '6': {'gt': False, 'rm': False},    # TN
}

with open(result_json_path, 'w') as f:
    json.dump(test_data, f)

ARGS = {
    'result_json_path': result_json_path
}
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 tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
is_last_two Ran
ran, violated the contract we drafted
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/captioner/prepare_caps_sub.py
code served (permissive licence) · get_code("fc819cea894082b1")
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 bool, 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

torch.manual_seed(0)

# Analyzing the function:
# - file_path: a string like "traj_name_0.png" or "trajectory_5.jpg"
# - total_images_per_traj: a dict mapping trajectory names to total image counts
# The function extracts trajectory name and index from file_path,
# then checks if traj_idx >= total_images_per_traj[traj_name] - 2

file_path = "demo_trajectory_5.png"
total_images_per_traj = {
    "demo_trajectory": 10
}

ARGS = {
    "file_path": file_path,
    "total_images_per_traj": total_images_per_traj
}
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 bool, 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
predict Ran
ran; our contract draft was wrong, not the code
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/captioner/prepare_caps_sub.py
code served (permissive licence) · get_code("c4f6c62dbdbdbb23")
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
from unittest.mock import Mock

torch.manual_seed(0)

# Create a mock client that has a predict method
client = Mock()

# Mock the predict method to return a simple result
def mock_predict(img_path, api_name=None):
    return {"prediction": "test_result", "confidence": 0.95}

client.predict = mock_predict

# Create a simple image path (string)
img_path = "/path/to/test/image.jpg"

ARGS = {
    "img_path": img_path,
    "client": client
}
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 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
query_anyscale_api Ran
ran; our contract draft was wrong, not the code
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/clients.py
code served (permissive licence) · get_code("a591edde183921f9")
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, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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 requests
import json

# Mock the requests.post to avoid actual network calls
class MockResponse:
    def __init__(self):
        self.data = {
            'choices': [
                {
                    'message': {
                        'content': '  This is a test response from the API'
                    }
                }
            ]
        }
    
    def json(self):
        return self.data

def mock_post(url, headers=None, data=None):
    return MockResponse()

# Patch requests.post
requests.post = mock_post

# Define the arguments
ARGS = {
    'messages': [
        {'role': 'user', 'content': 'Hello, how are you?'},
        {'role': 'assistant', 'content': 'I am doing well, thank you for asking.'}
    ],
    'model': 'meta-llama/Llama-2-7b-chat-hf',
    'temperature': 0.7
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
requests — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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
query_openai_api Ran
ran; our contract draft was wrong, not the code
berkeley-nlp/agent-eval-refine/agent_eval/agent_eval/clients.py
code served (permissive licence) · get_code("e42d0ea7ee303a30")
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, ndim 1, dtype family float; no output shape is stated. The image it ran in was not recorded. If your copy returns the same type and ndim 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 requests
import json

def query_openai_api(messages, model, temperature=0, api_key=None):
    try:
        url = 'https://api.openai.com/v1/chat/completions'
        headers = {'Content-Type': 'application/json', 'Authorization': f'Bearer {api_key}'}
        data = {'model': model, 'messages': messages, 'temperature': temperature, 'max_tokens': 4096}
        response = requests.post(url, headers=headers, data=json.dumps(data)).json()
        from pprint import pprint
        return (response['choices'][0]['message']['content'].lstrip(), response)
    except Exception as e:
        print(f'An error occurred: {e}')
        return (f'API_ERROR: {e}', None)


ARGS = {
    'messages': [
        {'role': 'user', 'content': 'Hello, what is 2 + 2?'},
        {'role': 'assistant', 'content': 'The answer is 4.'}
    ],
    'model': 'gpt-3.5-turbo',
    'temperature': 0.7,
    'api_key': 'sk-test-key-12345'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
requests — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 1, dtype family float — a tuple output has no array shape of its own; none is stated
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

Also present in other papers' code

Of 9 samples lifted from this paper's linked repositories, 9 carry a function identity and 0 do not. Of the 9 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 9 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 show that domain-general automatic evaluators can significantly improve the performance of agents for web navigation and device control. We experiment with multiple evaluation models that trade off between inference cost, modularity of design, and accuracy. We validate the performance of these models in several popular benchmarks for digital agents, finding between 74.4 and 92.9% agreement with oracle evaluation metrics. Finally, we use these evaluators to improve the performance of existing agents via fine-tuning and inference-time guidance. Without any additional supervision, we improve state-of-the-art performance by 29% on the popular benchmark WebArena, and achieve around 75% relative improvement in device control settings.

For agents

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

get_harvested_code_for_paper("2404.06474")
get_code_for_paper("2404.06474")
have("2404.06474")

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

curl https://syntology.ai/api/ran/2404.06474.json

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

[![Syntology run record](https://syntology.ai/api/ran/2404.06474.svg)](https://syntology.ai/paper/2404.06474)

Connect an agent — have() is free.