SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2305.10355 · EMNLP · 2023

Evaluating Object Hallucination in Large Vision-Language Models

Ji-Rong Wen, Kun Zhou, Wayne Zhao, Yifan Du, Yifan Li, Jinpeng Wang

arXiv · PDF · Open in the Atlas

Code that ran

11 of 12 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (2 of them constructed an object rather than computing a result; 1 honoured the contract we drafted, 0 violated it and 7 have no contract verdict; 3 failures of our instrument; 1 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/2305.10355.json · badge: /api/ran/2305.10355.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
rucaibox/pope canonical (classifier label)
classifier basis not recorded on the link
6 ran · 0 no recorded run
mengchuang123/vasparse-github community
Papers with Code archive
2 ran · 1 no recorded run
ziwei-zheng/nullu community
Papers with Code archive
2 ran · 0 no recorded run
yellow-binary-tree/pfram community
Papers with Code archive
1 ran · 0 no recorded run
FunctionStatusWhere it lives
BaseDataset Ran
ran on a synthesized input · constructed an object; no result was computed
ziwei-zheng/nullu/dataset/POPE.py
code served (permissive licence) · get_code("7efc8d99218ee581")
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 an instance of its own class, BaseDataset: the run constructed the object and did not call it, so no result was computed. 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 numpy as np

torch.manual_seed(0)

ARGS = {
}
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 BaseDataset — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a BaseDataset 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
POPEDataset Ran
ran on a synthesized input · constructed an object; no result was computed
ziwei-zheng/nullu/dataset/POPE.py
code served (permissive licence) · get_code("74d3f7e9e4fd9414")
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 an instance of its own class, POPEDataset: the run constructed the object and did not call it, so no result was computed. 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 os
import json
import random
import tempfile
import shutil

class BaseDataset:

    def __init__(self):
        pass

    def get_data(self):
        return []

class POPEDataset(BaseDataset):

    def __init__(self, split='val', data_root='/data/coco/', sampling='random', num_samples=500):
        super(POPEDataset, self).__init__()
        self.testfiles = ['./data/pope/coco_val/coco_val_pope_random.json', './data/pope/coco_val/coco_val_pope_popular.json', './data/pope/coco_val/coco_val_pope_adversarial.json']
        self.img_root = os.path.join(data_root, f'{split}2014')
        self.sampling = sampling
        self.num_samples = num_samples

    def get_data(self):
        val_data = {}
        for testfile in self.testfiles:
            with open(testfile, 'r') as f:
                inputs = [json.loads(line) for line in f]
            groups = [inputs[i:i + 6] for i in range(0, len(inputs), 6)]
            if self.num_samples:
                if self.num_samples > len(groups):
                    print(f'num_samples {self.num_samples} exceeds the number of images ({len(groups)})')
                    sampled_groups = groups
                elif self.sampling == 'first':
                    sampled_groups = groups[:self.num_samples]
                elif self.sampling == 'random':
                    sampled_groups = random.sample(groups, self.num_samples)
                else:
                    raise ValueError(f'Unsupported sampling strategy: {self.sampling}')
            else:
                sampled_groups = groups
            strategy = os.path.basename(testfile).split('_')[-1].replace('.json', '')
            val_data[strategy] = [{'image_path': os.path.join(self.img_root, ins['image']), 'question': ins['text'], 'label': ins['label']} for group in sampled_groups for ins in group]
        return val_data

# Create temporary directory and test files
tmpdir = tempfile.mkdtemp()
data_dir = os.path.join(tmpdir, 'data', 'pope', 'coco_val')
os.makedirs(data_dir, exist_ok=True)

# Create mock POPE test files with sample data
test_files = [
    os.path.join(data_dir, 'coco_val_pope_random.json'),
    os.path.join(data_dir, 'coco_val_pope_popular.json'),
    os.path.join(data_dir, 'coco_val_pope_adversarial.json')
]

for testfile in test_files:
    with open(testfile, 'w') as f:
        for i in range(12):
            sample = {
                'image': f'COCO_val2014_{i:012d}.jpg',
                'text': f'Is there a {["cat", "dog", "bird"][i % 3]} in the image?',
                'label': i % 2
            }
            f.write(json.dumps(sample) + '\n')

# Create mock image root
img_root = os.path.join(tmpdir, 'val2014')
os.makedirs(img_root, exist_ok=True)

ARGS = {
    'split': 'val',
    'data_root': tmpdir,
    'sampling': 'random',
    'num_samples': 2
}
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 POPEDataset — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a POPEDataset 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
compute_co_occurrence Ran
ran; our contract draft was wrong, not the code
mengchuang123/vasparse-github/eval/opope_eval.py
pointer only (licence: NONE) · get_code("d7741ef6128f74cd")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, mengchuang123/vasparse-github when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

create_question Ran
ran; our contract draft was wrong, not the code
mengchuang123/vasparse-github/eval/opope_eval.py
pointer only (licence: NONE) · get_code("9f48c9c176418803")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, mengchuang123/vasparse-github when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

create_question Ran
ran on a synthesized input
rucaibox/pope/utils.py
code served (permissive licence) · get_code("2715c0290945600d")
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)

# Analyze the function usage:
# - question_id: used as a dict value (any hashable/serializable value)
# - image: used as a dict value (can be any object)
# - Object: indexed with [0], compared to vowels, used in .format() - needs to be a string
# - label: used as a dict value (any value)
# - template: called with .replace() and .format() - needs to be a string

question_id = 42
image = torch.randn(3, 224, 224)  # Small image tensor
Object = "dog"  # String starting with consonant
label = 1  # Integer label
template = "This is {} in the image"

ARGS = {
    "question_id": question_id,
    "image": image,
    "Object": Object,
    "label": label,
    "template": template
}
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
evaluate Ran
ran, honoured the contract we drafted
yellow-binary-tree/pfram/POPE/evaluate.py
pointer only (licence: NONE) · get_code("dee83049b41f8c7d")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, yellow-binary-tree/pfram when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

generate_ground_truth_objects Ran
ran on a synthesized input
rucaibox/pope/utils.py
code served (permissive licence) · get_code("152eebdc8a308016")
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 os
import json
import tempfile

# Create a temporary directory for the output file
temp_dir = tempfile.mkdtemp()

# Construct segment_results - a list of dicts with 'objects' key
# Each object is a string that will be counted
segment_results = [
    {
        'objects': ['cat', 'dog', 'cat']
    },
    {
        'objects': ['dog', 'bird', 'cat']
    },
    {
        'objects': ['bird', 'bird']
    }
]

save_path = temp_dir
dataset = 'test_dataset'

ARGS = {
    'segment_results': segment_results,
    'save_path': save_path,
    'dataset': dataset
}
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
get_image Ran
ran on a synthesized input
rucaibox/pope/utils.py
code served (permissive licence) · get_code("d812ee50bd3ae6d8")
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 list, 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 os
import random
import tempfile

# Create a temporary directory with some dummy image files
temp_dir = tempfile.mkdtemp()

# Create a few dummy image files
for i in range(5):
    open(os.path.join(temp_dir, f"image_{i}.png"), 'w').close()

# Set seed for reproducibility
random.seed(42)

ARGS = {
    "img_path": temp_dir,
    "seg_num": 3
}
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 list, ndim 1, dtype family float — a list 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
load_opt_command Ran
ran on a synthesized input
rucaibox/pope/seem/utils/arguments.py
code served (permissive licence) · get_code("7025c6d903d31ec1")
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 tempfile
import os
import yaml

# Create a temporary config file
temp_dir = tempfile.mkdtemp()
config_file = os.path.join(temp_dir, "test_config.yaml")

config_data = {
    "model": {
        "name": "test_model",
        "hidden_size": 256,
        "num_layers": 2
    },
    "training": {
        "learning_rate": 0.001,
        "batch_size": 32,
        "epochs": 10
    },
    "data": {
        "path": "/path/to/data",
        "split": 0.8
    }
}

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

# Setup arguments for load_opt_command
args = [
    'train',
    '--conf_files', config_file,
    '--config_overrides', '{"model.hidden_size": 512, "training.epochs": 20}',
    '--overrides', 'training.batch_size', '64', 'training.learning_rate', '0.0001'
]

ARGS = {
    "args": args
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
yaml — 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
load_opt_from_config_files Ran
ran; our contract draft was wrong, not the code
rucaibox/pope/seem/utils/arguments.py
code served (permissive licence) · get_code("d3ab49d45b152e64")
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 yaml
import os

# Create a temporary YAML config file
config_content = {
    'model': {
        'name': 'resnet50',
        'layers': 4,
        'hidden_dim': 256
    },
    'training': {
        'epochs': 100,
        'batch_size': 32,
        'learning_rate': 0.001
    },
    'data.path': '/tmp/data',
    'data.split': 0.8
}

# Write config to a temporary file
temp_dir = tempfile.mkdtemp()
conf_file = os.path.join(temp_dir, 'config.yaml')
with open(conf_file, 'w', encoding='utf-8') as f:
    yaml.dump(config_content, f)

ARGS = {
    'conf_file': conf_file
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
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
register_norm_module Ran
ran on a synthesized input
rucaibox/pope/seem/utils/model.py
pointer only (licence: MIT); the stored copy get_code serves, from UARK-AICV/OpenFusion, is pointer only (licence: NONE) · get_code("e4f6733fa1ab1999")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (MIT), but get_code serves the stored copy, first harvested from UARK-AICV/OpenFusion, and GitHub detected no licence in that repository when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. A run is recorded.

pope Not yet run
unverified (harvested, no recorded run)
mengchuang123/vasparse-github/eval/opope_eval.py
pointer only (licence: NONE) · get_code("c8ccdee5027e2c81")
How it ran

The body of this sample is withheld: GitHub detected no licence in this paper's repository, mengchuang123/vasparse-github when it was checked. The setup block that called it is withheld with it, because it is written against that body. source_path and origin say where the code lives, to read under the repository's own terms. No invocation recorded.

Also present in other papers' code

Of 12 samples lifted from this paper's linked repositories, 12 carry a function identity and 0 do not. Of the 12 distinct functions with an identity, 6 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 6 later) and 6 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
register_norm_module 4 other papers
4 later
2306.09347 (2023-06)
Segment Any Point Cloud Sequences by Distilling Vision Foundation Models
2310.03923 (2023-10)
Open-Fusion: Real-time Open-Vocabulary 3D Mapping and Queryable Scene Representation
2311.00571 (2023-11)
LLaVA-Interactive: An All-in-One Demo for Image Chat, Segmentation, Generation and Editing
and 1 more not named here
load_opt_from_config_files 3 other papers
3 later
2312.09158 (2023-12)
General Object Foundation Model for Images and Videos at Scale
2407.16696 (2024-07)
PartGLEE: A Foundation Model for Recognizing and Parsing Any Objects
2025.emnlp-main.534 (2025 (year only))
On the Role of Model Prior in Real-World Inductive Reasoning
create_question 1 other paper
1 later
2025.emnlp-main.534 (2025 (year only))
On the Role of Model Prior in Real-World Inductive Reasoning
generate_ground_truth_objects 1 other paper
1 later
2025.emnlp-main.534 (2025 (year only))
On the Role of Model Prior in Real-World Inductive Reasoning
get_image 1 other paper
1 later
2025.emnlp-main.534 (2025 (year only))
On the Role of Model Prior in Real-World Inductive Reasoning
load_opt_command 1 other paper
1 later
2025.emnlp-main.534 (2025 (year only))
On the Role of Model Prior in Real-World Inductive Reasoning

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

Inspired by the superior language abilities of large language models (LLM), large visionlanguage models (LVLM) have been recently proposed by integrating powerful LLMs for improving the performance on complex multimodal tasks. Despite the promising progress on LVLMs, we find that they suffer from object hallucinations, i.e., they tend to generate objects inconsistent with the target images in the descriptions. To investigate it, this work presents the first systematic study on object hallucination of LVLMs. We conduct the evaluation experiments on several representative LVLMs, and show that they mostly suffer from severe object hallucination issues. We further discuss that the visual instructions may influence the hallucination, and find that: objects that frequently appear in the visual instructions or co-occur with the image objects are obviously prone to be hallucinated by LVLMs. Besides, we further design a polling-based query method called POPE for better evaluation of object hallucination. Experiment results show that our POPE can evaluate object hallucination in a more stable and flexible way.

For agents

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

get_harvested_code_for_paper("2305.10355")
get_code_for_paper("2305.10355")
have("2305.10355")

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

curl https://syntology.ai/api/ran/2305.10355.json

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

[![Syntology run record](https://syntology.ai/api/ran/2305.10355.svg)](https://syntology.ai/paper/2305.10355)

Connect an agent — have() is free.