SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2408.12902 · 2024

IAA: Inner-Adaptor Architecture Empowers Frozen Large Language Model with Multimodal Capabilities

arXiv · PDF · Open in the Atlas

Code that ran

9 of 10 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; 2 honoured the contract we drafted, 1 violated it and 0 have no contract verdict; 6 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/2408.12902.json · badge: /api/ran/2408.12902.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
360cvgroup/inner-adaptor-architecture canonical (classifier label)
classifier basis: census
9 ran · 1 no recorded run
FunctionStatusWhere it lives
compute_iou Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
360cvgroup/inner-adaptor-architecture/iaa/eval/compute_precision.py
code served (permissive licence) · get_code("d7c36fd89f5fcad4")
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

# Set random seed for reproducibility
torch.manual_seed(0)

# bbox1 and bbox2 are tuples/lists representing (x0, y0, x1, y1)
# These represent bounding boxes with top-left corner (x0, y0) and bottom-right corner (x1, y1)

bbox1 = (10, 20, 50, 80)
bbox2 = (30, 40, 70, 100)

ARGS = {
    "bbox1": bbox1,
    "bbox2": bbox2
}
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
expand2square Ran
ran; our contract draft was wrong, not the code
360cvgroup/inner-adaptor-architecture/iaa/mm_utils.py
code served (permissive licence) · get_code("592b3c1a88f93d7c")
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 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 PIL import Image

# Create a small test image (non-square)
pil_img = Image.new('RGB', (8, 6), color='red')

# Define background color for padding
background_color = (255, 255, 255)

ARGS = {
    'pil_img': pil_img,
    'background_color': background_color,
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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
get_chunk Ran
ran; our contract draft was wrong, not the code
360cvgroup/inner-adaptor-architecture/iaa/eval/model_vqa_loader_llama3.py
code served (permissive licence) · get_code("42a46570620cd9fa")
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 list, 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 math

def split_list(lst, n):
    """Split a list into n (roughly) equal-sized chunks"""
    chunk_size = math.ceil(len(lst) / n)
    return [lst[i:i + chunk_size] for i in range(0, len(lst), chunk_size)]

def get_chunk(lst, n, k):
    chunks = split_list(lst, n)
    return chunks[k]

# Setup code
lst = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
n = 3
k = 1

ARGS = {
    "lst": lst,
    "n": n,
    "k": k
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 int — 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
get_gt Ran
ran; our contract draft was wrong, not the code
360cvgroup/inner-adaptor-architecture/MME/convert_answer_to_mme.py
code served (permissive licence) · get_code("3a3b1cb3a821cbbe")
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 tempfile
import shutil

# Create a temporary directory structure matching what get_gt expects
temp_dir = tempfile.mkdtemp()

# Create category subdirectories with the expected structure
category1 = os.path.join(temp_dir, "category1")
os.makedirs(category1)

images_dir = os.path.join(category1, "images")
os.makedirs(images_dir)

qa_dir = os.path.join(category1, "questions_answers_YN")
os.makedirs(qa_dir)

# Create a sample QA file with tab-separated question and answer pairs
qa_file = os.path.join(qa_dir, "qa_sample.txt")
with open(qa_file, 'w') as f:
    f.write("Is this a cat?\tyes\n")
    f.write("Is this a dog?\tno\n")

# Create another category without the images subdirectory structure
category2 = os.path.join(temp_dir, "category2")
os.makedirs(category2)

qa_file2 = os.path.join(category2, "qa_sample2.txt")
with open(qa_file2, 'w') as f:
    f.write("Does it fly?\tyes\n")

# Create a non-directory file to test the isdir check
non_dir_file = os.path.join(temp_dir, "not_a_directory.txt")
with open(non_dir_file, 'w') as f:
    f.write("This is not a directory\n")

ARGS = {
    "data_path": temp_dir
}
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
load_image_from_base64 Ran
ran, honoured the contract we drafted
360cvgroup/inner-adaptor-architecture/iaa/mm_utils.py
code served (permissive licence) · get_code("c3ee9d07c900dd55")
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 PngImageFile. 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
from PIL import Image
from io import BytesIO
import numpy as np

# Create a small valid image in memory
img_array = np.uint8(np.random.rand(10, 10, 3) * 255)
img = Image.fromarray(img_array, mode='RGB')

# Encode to base64
buffer = BytesIO()
img.save(buffer, format='PNG')
buffer.seek(0)
image_base64 = base64.b64encode(buffer.read()).decode('utf-8')

ARGS = {
    'image': image_base64
}
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 PngImageFile — a PngImageFile output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a PngImageFile 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
ori_bbox Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
360cvgroup/inner-adaptor-architecture/iaa/eval/compute_precision.py
code served (permissive licence) · get_code("b0718fd6553c1a39")
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 torch
import numpy as np

# Analyze the function:
# - nor_bbox: expects indexing [0], [1], [2], [3] - so a list or array with 4 elements
# - img_size: expects indexing [0], [1] - so a tuple/list with 2 elements (width, height)
# - All operations are numerical: multiplication, subtraction, integer division
# - Returns a list of 4 values

# Create test arguments
nor_bbox = [0.1, 0.2, 0.8, 0.9]  # normalized bbox coordinates (x1, y1, x2, y2)
img_size = (640, 480)  # (width, height) - width > height case

ARGS = {
    "nor_bbox": nor_bbox,
    "img_size": img_size
}
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 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
process_images Ran
ran, honoured the contract we drafted
360cvgroup/inner-adaptor-architecture/iaa/mm_utils.py
code served (permissive licence) · get_code("344dff4791fd1381")
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 Tensor, ndim 4, dtype family float, shape (2, 3, 4, 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.

from PIL import Image
import torch
import numpy as np

torch.manual_seed(0)

# Create a minimal image processor mock
class MockImageProcessor:
    def __init__(self):
        self.image_mean = [0.5, 0.5, 0.5]  # normalized values
    
    def preprocess(self, image, return_tensors=None):
        # Convert PIL image to tensor
        img_array = np.array(image).astype(np.float32) / 255.0
        # Convert to (C, H, W) format
        if len(img_array.shape) == 3:
            img_tensor = torch.from_numpy(img_array).permute(2, 0, 1)
        else:
            img_tensor = torch.from_numpy(img_array).unsqueeze(0)
        return {'pixel_values': img_tensor.unsqueeze(0)}
    
    def __call__(self, images, return_tensors=None):
        # Handle batch of images
        processed = []
        for img in images:
            img_array = np.array(img).astype(np.float32) / 255.0
            if len(img_array.shape) == 3:
                img_tensor = torch.from_numpy(img_array).permute(2, 0, 1)
            else:
                img_tensor = torch.from_numpy(img_array).unsqueeze(0)
            processed.append(img_tensor)
        stacked = torch.stack(processed, dim=0)
        return {'pixel_values': stacked}

# Create a minimal model config mock
class MockModelConfig:
    def __init__(self, image_aspect_ratio=None):
        self.image_aspect_ratio = image_aspect_ratio

# Create sample PIL images with different sizes
# First image: 4x3 (width x height)
img1 = Image.new('RGB', (4, 3), color=(100, 150, 200))
# Second image: 3x4 (width x height) - different aspect ratio
img2 = Image.new('RGB', (3, 4), color=(50, 100, 150))

images = [img1, img2]
image_processor = MockImageProcessor()
model_cfg = MockModelConfig(image_aspect_ratio='pad')

ARGS = {
    'images': images,
    'image_processor': image_processor,
    'model_cfg': model_cfg
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
PIL, numpy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type Tensor, ndim 4, dtype family float, shape (2, 3, 4, 4) — shape (2, 3, 4, 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
split_list Ran
ran; our contract draft was wrong, not the code
360cvgroup/inner-adaptor-architecture/iaa/eval/model_vqa_loader_llama3.py
code served (permissive licence) · get_code("076c252c52cbb161")
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 list. 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 math

def split_list(lst, n):
    """Split a list into n (roughly) equal-sized chunks"""
    chunk_size = math.ceil(len(lst) / n)
    return [lst[i:i + chunk_size] for i in range(0, len(lst), chunk_size)]

# Setup code
lst = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
n = 3

ARGS = {
    "lst": lst,
    "n": n
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1
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 — a list 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
violates_moderation Ran
ran, violated the contract we drafted
360cvgroup/inner-adaptor-architecture/iaa/utils.py
pointer only (licence: Apache-2.0); the stored copy get_code serves, from Mi3-Lab/doScenes-VLM-Planning, is pointer only (licence: AGPL-3.0) · get_code("f9939a84b9a65279")
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 Mi3-Lab/doScenes-VLM-Planning, and AGPL-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.

pretty_print_semaphore Not yet run
unverified (harvested, no recorded run)
360cvgroup/inner-adaptor-architecture/iaa/utils.py
code served (permissive licence) · get_code("37899f22fb191b37")
How it ran

no invocation recorded.

Also present in other papers' code

Of 10 samples lifted from this paper's linked repositories, 10 carry a function identity and 0 do not. Of the 10 distinct functions with an identity, 7 are also present in other papers whose code we hold (7 with at least one paper dated earlier, 6 the same month, 7 later) and 3 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
split_list 197 other papers
87 dated earlier, 4 the same month, 99 later, 6 same year, order not determinable, 1 undated
2112.05682 (2021-12)
Self-attention Does Not Need $O(n^2)$ Memory
2306.17107 (2023-06)
LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding
2307.10490 (2023-07)
Abusing Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs
and 194 more not named here
get_chunk 189 other papers
84 dated earlier, 3 the same month, 96 later, 6 same year, order not determinable
2112.05682 (2021-12)
Self-attention Does Not Need $O(n^2)$ Memory
2306.17107 (2023-06)
LLaVAR: Enhanced Visual Instruction Tuning for Text-Rich Image Understanding
2307.10490 (2023-07)
Abusing Images and Sounds for Indirect Instruction Injection in Multi-Modal LLMs
and 186 more not named here
load_image_from_base64 85 other papers
49 dated earlier, 1 the same month, 31 later, 4 same year, order not determinable
2112.05682 (2021-12)
Self-attention Does Not Need $O(n^2)$ Memory
2308.10253 (2023-08)
StableLLaVA: Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data
2309.07120 (2023-09)
Sight Beyond Text: Multi-Modal Training Enhances LLMs in Truthfulness and Ethics
and 82 more not named here
expand2square 68 other papers
39 dated earlier, 1 the same month, 25 later, 3 same year, order not determinable · 1 only as an assembly brick
2112.05682 (2021-12)
Self-attention Does Not Need $O(n^2)$ Memory
2204.03541 (2022-04)
End-to-End Zero-Shot HOI Detection via Vision and Language Knowledge Distillation
2308.10253 (2023-08)
StableLLaVA: Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data
and 65 more not named here
violates_moderation 56 other papers
34 dated earlier, 1 the same month, 20 later, 1 same year, order not determinable
2106.06909 (2021-06)
GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio
2308.03279 (2023-08)
UniversalNER: TARGETED DISTILLATION FROM LARGE LANGUAGE MODELS FOR OPEN NAMED ENTITY RECOGNITION
2308.13437 (2023-08)
Position-Enhanced Visual Instruction Tuning for Multimodal Large Language Models
and 53 more not named here
pretty_print_semaphore 54 other papers
31 dated earlier, 1 the same month, 21 later, 1 same year, order not determinable
2106.06909 (2021-06)
GigaSpeech: An Evolving, Multi-domain ASR Corpus with 10,000 Hours of Transcribed Audio
2308.13437 (2023-08)
Position-Enhanced Visual Instruction Tuning for Multimodal Large Language Models
2309.07120 (2023-09)
Sight Beyond Text: Multi-Modal Training Enhances LLMs in Truthfulness and Ethics
and 51 more not named here
process_images 44 other papers
30 dated earlier, 11 later, 3 same year, order not determinable · 1 only as an assembly brick
2112.05682 (2021-12)
Self-attention Does Not Need $O(n^2)$ Memory
2308.10253 (2023-08)
StableLLaVA: Enhanced Visual Instruction Tuning with Synthesized Image-Dialogue Data
2310.01779 (2023-10)
HallE-Control: Controlling Object Hallucination in Large Multimodal Models
and 41 more not named here

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

In the field of multimodal large language models (MLLMs), common methods typically involve unfreezing the language model during training to foster profound visual understanding. However, the fine-tuning of such models with vision-language data often leads to a diminution of their natural language processing (NLP) capabilities. To avoid this performance degradation, a straightforward solution is to freeze the language model while developing multimodal competencies. Unfortunately, previous works have not attained satisfactory outcomes. Building on the strategy of freezing the language model, we conduct thorough structural exploration and introduce the Inner-Adaptor Architecture (IAA). Specifically, the architecture incorporates multiple multimodal adaptors at varying depths within the large language model to facilitate direct interaction with the inherently text-oriented transformer layers, thereby enabling the frozen language model to acquire multimodal capabilities. Unlike previous approaches of freezing language models that require large-scale aligned data, our proposed architecture is able to achieve superior performance on small-scale datasets. We conduct extensive experiments to improve the general multimodal capabilities and visual grounding abilities of the MLLM. Our approach remarkably outperforms previous state-of-the-art methods across various vision-language benchmarks without sacrificing performance on NLP tasks. Code and models are available at https://github.com/360CVGroup/Inner-Adaptor-Architecture.

For agents

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

get_harvested_code_for_paper("2408.12902")
get_code_for_paper("2408.12902")
have("2408.12902")

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

curl https://syntology.ai/api/ran/2408.12902.json

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

[![Syntology run record](https://syntology.ai/api/ran/2408.12902.svg)](https://syntology.ai/paper/2408.12902)

Connect an agent — have() is free.