SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2104.08066 · EMNLP · 2021

Effect of Visual Extensions on Natural Language Understanding in Vision-and-Language Models

Taichi Iki, Akiko Aizawa

arXiv · PDF · Open in the Atlas

Code that ran

8 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; 0 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 7 failures of our instrument; 1 no recorded run), measured 2026-11-21; 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/2104.08066.json · badge: /api/ran/2104.08066.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
alab-nii/eval_vl_glue canonical (classifier label)
classifier basis: census
8 ran · 1 no recorded run
FunctionStatusWhere it lives
find_pruneable_heads_and_indices Ran
ran; our contract draft was wrong, not the code
alab-nii/eval_vl_glue/eval_vl_glue/transformers_volta/modeling_utils.py
code served (permissive licence) · get_code("6d254a1b39c650ff")
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

torch.manual_seed(0)

# Small example values for the function parameters
heads_to_prune = [1, 3]  # List of head indices to prune
n_heads = 8  # Total number of heads in the model
head_size = 4  # Size of each head
already_pruned = {0, 2}  # Set of heads already pruned

ARGS = {
    "heads": heads_to_prune,
    "n_heads": n_heads,
    "head_size": head_size,
    "already_pruned_heads": already_pruned,
}
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 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
gelu_fast Ran
ran; our contract draft was wrong, not the code
alab-nii/eval_vl_glue/eval_vl_glue/transformers_volta/activations.py
code served (permissive licence) · get_code("a4475703ff58ecf9")
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) and returned Tensor, ndim 2, dtype family float, shape (2, 4) 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)

# gelu_fast takes a single parameter 'x' which is used in arithmetic operations
# It should be a tensor that supports multiplication, addition, and tanh operations
x = torch.randn(2, 4)

ARGS = {
    "x": x
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
Imports
torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 2, dtype family float, shape (2, 4) — shape (2, 4) 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
gelu_new Ran
ran; our contract draft was wrong, not the code
alab-nii/eval_vl_glue/eval_vl_glue/transformers_volta/activations.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from astro-jon/LCWL, is pointer only (licence: NONE) · get_code("45bb87451230d5e8")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from astro-jon/LCWL, 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.

is_remote_url Ran
ran, violated the contract we drafted
alab-nii/eval_vl_glue/eval_vl_glue/transformers_volta/file_utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from vis-nlp/ChartQA, is pointer only (licence: GPL-3.0) · get_code("899d1e2736236351")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from vis-nlp/ChartQA, and GPL-3.0, as GitHub detects it on that repository, is not in the permissive set Syntology serves inline. 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.

load_tsv Ran
ran; our contract draft was wrong, not the code
alab-nii/eval_vl_glue/eval_vl_glue/volta_image_feature.py
code served (permissive licence) · get_code("598a2bf62b8c0a24")
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 base64
import numpy as np
import tempfile
import os

# Create a temporary TSV file with valid data
temp_dir = tempfile.mkdtemp()
tsv_path = os.path.join(temp_dir, "test.tsv")

# Create sample data for TSV file
def create_sample_tsv(path):
    lines = []
    
    # Sample row 1
    img_id = "img_001"
    img_h = 480
    img_w = 640
    
    # Create numpy arrays
    objects_id = np.array([1, 2, 3], dtype=np.int64)
    objects_conf = np.array([0.9, 0.8, 0.7], dtype=np.float32)
    attrs_id = np.array([10, 11], dtype=np.int64)
    attrs_conf = np.array([0.85, 0.75], dtype=np.float32)
    num_boxes = 3
    boxes = np.array([[10.5, 20.5, 100.5, 150.5], [200.5, 300.5, 400.5, 450.5], [50.5, 60.5, 150.5, 200.5]], dtype=np.float32)
    features = np.array([[0.1, 0.2, 0.3, 0.4], [0.5, 0.6, 0.7, 0.8], [0.9, 0.1, 0.2, 0.3]], dtype=np.float32)
    
    # Encode numpy arrays to base64
    objects_id_b64 = base64.b64encode(objects_id.tobytes()).decode('ascii')
    objects_conf_b64 = base64.b64encode(objects_conf.tobytes()).decode('ascii')
    attrs_id_b64 = base64.b64encode(attrs_id.tobytes()).decode('ascii')
    attrs_conf_b64 = base64.b64encode(attrs_conf.tobytes()).decode('ascii')
    boxes_b64 = base64.b64encode(boxes.tobytes()).decode('ascii')
    features_b64 = base64.b64encode(features.tobytes()).decode('ascii')
    
    # Create TSV line
    row = [
        img_id,
        str(img_h),
        str(img_w),
        objects_id_b64,
        objects_conf_b64,
        attrs_id_b64,
        attrs_conf_b64,
        str(num_boxes),
        boxes_b64,
        features_b64
    ]
    lines.append('\t'.join(row))
    
    # Sample row 2
    img_id2 = "img_002"
    img_h2 = 360
    img_w2 = 480
    objects_id2 = np.array([4, 5], dtype=np.int64)
    objects_conf2 = np.array([0.95, 0.82], dtype=np.float32)
    attrs_id2 = np.array([12], dtype=np.int64)
    attrs_conf2 = np.array([0.88], dtype=np.float32)
    num_boxes2 = 2
    boxes2 = np.array([[30.5, 40.5, 120.5, 160.5], [250.5, 280.5, 420.5, 470.5]], dtype=np.float32)
    features2 = np.array([[0.2, 0.3, 0.4, 0.5], [0.6, 0.7, 0.8, 0.9]], dtype=np.float32)
    
    objects_id2_b64 = base64.b64encode(objects_id2.tobytes()).decode('ascii')
    objects_conf2_b64 = base64.b64encode(objects_conf2.tobytes()).decode('ascii')
    attrs_id2_b64 = base64.b64encode(attrs_id2.tobytes()).decode('ascii')
    attrs_conf2_b64 = base64.b64encode(attrs_conf2.tobytes()).decode('ascii')
    boxes2_b64 = base64.b64encode(boxes2.tobytes()).decode('ascii')
    features2_b64 = base64.b64encode(features2.tobytes()).decode('ascii')
    
    row2 = [
        img_id2,
        str(img_h2),
        str(img_w2),
        objects_id2_b64,
        objects_conf2_b64,
        attrs_id2_b64,
        attrs_conf2_b64,
        str(num_boxes2),
        boxes2_b64,
        features2_b64
    ]
    lines.append('\t'.join(row2))
    
    with open(path, 'w') as f:
        f.write('\n'.join(lines))

create_sample_tsv(tsv_path)

ARGS = {
    "tsv_path": tsv_path
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy — 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
mish Ran
ran; our contract draft was wrong, not the code
alab-nii/eval_vl_glue/eval_vl_glue/transformers_volta/activations.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from thu-coai/Emotional-Support-Conversation, is pointer only (licence: NOASSERTION) · get_code("f889d53c68539013")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from thu-coai/Emotional-Support-Conversation, and GitHub could not classify the licence of that repository (NOASSERTION). 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.

replace_return_docstrings Ran
ran; our contract draft was wrong, not the code
alab-nii/eval_vl_glue/eval_vl_glue/transformers_volta/file_utils.py
code served (permissive licence) · get_code("2cd9135e83a62a04")
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
from typing import Optional

# Create a minimal output type class with docstring
class MockOutput:
    """
    Base class for model outputs.
    
    Args:
        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head.
    """
    def __init__(self, loss=None, logits=None):
        self.loss = loss
        self.logits = logits


class TFMockOutput:
    """
    Base class for TensorFlow model outputs.
    
    Args:
        loss (`tf.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
            Language modeling loss (for next-token prediction).
        logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
            Prediction scores of the language modeling head.
    """
    def __init__(self, loss=None, logits=None):
        self.loss = loss
        self.logits = logits


# Create a test function with Returns: placeholder
def sample_forward_function(x):
    """
    Forward pass of the model.
    
    Args:
        x: Input tensor
    
    Returns:
        Model output
    """
    return x


# Create config class name string
config_class_name = "BertConfig"

# Create output type that has proper __doc__, __module__, and __name__
MockOutput.__module__ = "transformers.models.bert.modeling_bert"
MockOutput.__name__ = "BaseModelOutput"

ARGS = {
    "output_type": MockOutput,
    "config_class": config_class_name,
}
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
torch_only_method Ran
ran; our contract draft was wrong, not the code
alab-nii/eval_vl_glue/eval_vl_glue/transformers_volta/file_utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from miccunifi/ScanTalk, is pointer only (licence: NOASSERTION) · get_code("56ffa4b32330ceab")
How it ran

The body of this sample is withheld: this paper's link allows the body inline (Apache-2.0), but get_code serves the stored copy, first harvested from miccunifi/ScanTalk, and GitHub could not classify the licence of that repository (NOASSERTION). 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.

top_k_top_p_filtering Not yet run
unverified (harvested, no recorded run)
alab-nii/eval_vl_glue/eval_vl_glue/transformers_volta/generation_utils.py
code served (permissive licence) · get_code("9c750e2920328942")
How it ran

no invocation recorded.

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, 8 are also present in other papers whose code we hold (5 with at least one paper dated earlier, 2 the same month, 8 later) and 1 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
gelu_new 24 other papers
1 dated earlier, 1 the same month, 22 later
2004.10964 (2020-04)
Don't Stop Pretraining: Adapt Language Models to Domains and Tasks
2104.08400 (2021-04)
Structure-Aware Abstractive Conversation Summarization via Discourse and Action Graphs
2106.00420 (2021-06)
Dialogue-oriented Pre-training
and 21 more not named here
gelu_fast 22 other papers
1 dated earlier, 1 the same month, 20 later
2010.03957 (2020-10)
Transformers for Modeling Physical Systems
2104.08400 (2021-04)
Structure-Aware Abstractive Conversation Summarization via Discourse and Action Graphs
2106.00420 (2021-06)
Dialogue-oriented Pre-training
and 19 more not named here
is_remote_url 14 other papers
3 dated earlier, 11 later
1905.00641 (2019-05)
RetinaFace: Single-stage Dense Face Localisation in the Wild
1910.13461 (2019-10)
BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
2004.10964 (2020-04)
Don't Stop Pretraining: Adapt Language Models to Domains and Tasks
and 11 more not named here
find_pruneable_heads_and_indices 11 other papers
1 dated earlier, 9 later, 1 same year, order not determinable
2004.10964 (2020-04)
Don't Stop Pretraining: Adapt Language Models to Domains and Tasks
2021.emnlp-main.154 (2021 (year only))
Reconstruction Attack on Instance Encoding for Language Understanding
2108.10904 (2021-08)
SimVLM: Simple Visual Language Model Pretraining with Weak Supervision
and 8 more not named here
mish 4 other papers
4 later
2107.00910 (2021-07)
Learned Token Pruning for Transformers
2109.03808 (2021-09)
Smelting Gold and Silver for Improved Multilingual AMR-to-Text Generation
2405.15179 (2024-05)
VB-LoRA: Extreme Parameter Efficient Fine-Tuning with Vector Banks
and 1 more not named here
replace_return_docstrings 4 other papers
1 dated earlier, 3 later
1910.13461 (2019-10)
BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
2210.00312 (2022-10)
Multimodal Analogical Reasoning over Knowledge Graphs
2305.01938 (2023-05)
Doc2SoarGraph: Discrete Reasoning over Visually-Rich Table-Text Documents via Semantic-Oriented Hierarchical Graphs
and 1 more not named here
top_k_top_p_filtering 4 other papers
4 later
2305.00909 (2023-05)
Outline, Then Details: Syntactically Guided Coarse-To-Fine Code Generation
2310.05573 (2023-10)
ODEFormer: Symbolic Regression of Dynamical Systems with Transformers
2025.emnlp-main.1418 (2025 (year only))
Synth-SBDH: A Synthetic Dataset of Social and Behavioral Determinants of Health for Clinical Text
and 1 more not named here
torch_only_method 3 other papers
3 later
2305.01938 (2023-05)
Doc2SoarGraph: Discrete Reasoning over Visually-Rich Table-Text Documents via Semantic-Oriented Hierarchical Graphs
2311.06761 (2023-11)
Learning Knowledge-Enhanced Contextual Language Representations for Domain Natural Language Understanding
2403.10942 (2024-03)
ScanTalk: 3D Talking Heads from Unregistered Scans

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 2026-11-21.

Repositories linked to this paper

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

Abstract

A method for creating a vision-and-language (V&L) model is to extend a language model through structural modifications and V&L pretraining. Such an extension aims to make a V&L model inherit the capability of natural language understanding (NLU) from the original language model. To see how well this is achieved, we propose to evaluate V&L models using an NLU benchmark (GLUE). We compare five V&L models, including singlestream and dual-stream models, trained with the same pre-training. Dual-stream models, with their higher modality independence achieved by approximately doubling the number of parameters, are expected to preserve the NLU capability better. Our main finding is that the dual-stream scores are not much different than the single-stream scores, contrary to expectation. Further analysis shows that pre-training causes the performance drop in NLU tasks with few exceptions. These results suggest that adopting a single-stream structure and devising the pre-training could be an effective method for improving the maintenance of language knowledge in V&L extensions.

For agents

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

get_harvested_code_for_paper("2104.08066")
get_code_for_paper("2104.08066")
have("2104.08066")

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

curl https://syntology.ai/api/ran/2104.08066.json

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

[![Syntology run record](https://syntology.ai/api/ran/2104.08066.svg)](https://syntology.ai/paper/2104.08066)

Connect an agent — have() is free.