SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2501.03012 · ICCV · 2025

Analyzing Fine-tuning Representation Shift for Multimodal LLMs Steering

Matthieu Cord, Jayneel Parekh, Mustafa Shukor, Pegah Khayatan, Arnaud Dapogny

arXiv · PDF · Open in the Atlas

Code that ran

11 of 21 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; 3 honoured the contract we drafted, 2 violated it and 0 have no contract verdict; 6 failures of our instrument; 10 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/2501.03012.json · badge: /api/ran/2501.03012.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
mshukor/xl-vlms marked official in Papers with Code
Papers with Code archive
11 ran · 10 no recorded run
FunctionStatusWhere it lives
common_local_indices Ran
ran; our contract draft was wrong, not the code
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("f23b6db9cd4f073c")
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)

# The function expects two metadata dictionaries
# It accesses 'token_of_interest_mask' key and iterates over dict values
# We need to provide metadata dicts with consistent structure

# Case 1: With token_of_interest_mask provided
metadata_original_feature = {
    'token_of_interest_mask': torch.tensor([1.0, 1.0, 0.0, 1.0]),
    'hidden_state': torch.randn(4, 8),
}

meta_data_destination_feature = {
    'token_of_interest_mask': torch.tensor([1.0, 0.0, 1.0, 1.0]),
    'hidden_state': torch.randn(4, 8),
}

ARGS = {
    'metadata_original_feature': metadata_original_feature,
    'meta_data_destination_feature': meta_data_destination_feature,
}
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
compute_mean_per_concept Ran
ran, honoured the contract we drafted
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("b6646a299e4bee79")
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 2, 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
from scipy import stats

torch.manual_seed(0)

# Create sample features for each cluster
# Each element should be a tensor with shape (num_samples, feature_dim)
features_cluster_1 = torch.randn(5, 4)  # 5 samples, 4 features
features_cluster_2 = torch.randn(3, 4)  # 3 samples, 4 features
features_cluster_3 = torch.randn(7, 4)  # 7 samples, 4 features

features_associated_to_clusters = [features_cluster_1, features_cluster_2, features_cluster_3]

ARGS = {
    "features_associated_to_clusters": features_associated_to_clusters,
    "mean_method": 1
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type list, ndim 2, 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
compute_reference_clusters Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("6f4dfb20225ba3fa")
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)

# Construct arguments for compute_reference_clusters

# concepts: tensor of shape (num_concepts, feature_dim)
# Let's use 3 concepts with 4-dimensional features
concepts = torch.randn(3, 4)

# concept_activations: tensor of shape (num_samples, num_concepts)
# Let's use 2 samples and 3 concepts
concept_activations = torch.randn(2, 3)

# features: tensor of shape (num_samples, feature_dim)
# 2 samples with 4-dimensional features
features = torch.randn(2, 4)

# mask_samples: a collection of sample indices that are valid
# We'll use a set or list of indices from the samples (0, 1)
mask_samples = {0, 1}

# similarity: string parameter for the similarity metric
similarity = 'cosine'

ARGS = {
    'concepts': concepts,
    'concept_activations': concept_activations,
    'features': features,
    'mask_samples': mask_samples,
    'similarity': similarity,
}
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
concept_image_grounding Ran
ran; our contract draft was wrong, not the code
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("f6e5ae21703e74d4")
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 2, dtype family int, shape (4, 5) 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)

# The function takes:
# - activations: a torch.Tensor that will be sorted along dim=0
# - num_images_per_concept: an int (default 5)
#
# The function performs torch.argsort on dim=0, takes top num_images_per_concept rows,
# then transposes. So activations should have shape where dim=0 has at least
# num_images_per_concept elements.

# Create a small activations tensor
# Shape: (num_images, num_concepts) = (6, 4) so we can take top 5 from dim=0
activations = torch.randn(6, 4)

num_images_per_concept = 5

ARGS = {
    "activations": activations,
    "num_images_per_concept": num_images_per_concept,
}
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 Tensor, ndim 2, dtype family int, shape (4, 5) — shape (4, 5) 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
cosine_similarity Ran
ran; our contract draft was wrong, not the code
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("bcb8d066adb36735")
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 2, dtype family float, shape (2, 3) as a later run recorded it. The image it ran in was not recorded. If your copy returns the same type, ndim and shape on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch

torch.manual_seed(0)

# Create two small batches of vectors
# batch1: shape (2, 4) - 2 vectors of dimension 4
# batch2: shape (3, 4) - 3 vectors of dimension 4
batch1 = torch.randn(2, 4)
batch2 = torch.randn(3, 4)

# eps is a small float for numerical stability
eps = 1e-08

ARGS = {
    "batch1": batch1,
    "batch2": batch2,
    "eps": eps
}
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 Tensor, ndim 2, dtype family float, shape (2, 3) — shape (2, 3) as a later run recorded it
dtype family
int means an integer dtype; float means any other dtype the instrument saw, bool and complex included (fixture_check.py records only these two families)
Environment
image not recorded; re-verify against syntology-harvest-sandbox:v5-deps; not recorded for this sample: the harness installs a cuda-to-cpu shim (cuda_shim.py) before every sample and no run records whether a sample needed it; seed: none by the harness; any seed is in the setup block
get_stopwords Ran
ran; our contract draft was wrong, not the code
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("dae6013ca8724453")
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 tempfile
import os

# Create a temporary file with stopwords content
temp_dir = tempfile.mkdtemp()
gist_file_path = os.path.join(temp_dir, 'gist_stopwords.txt')

# Write sample stopwords to the file
with open(gist_file_path, 'w') as f:
    f.write('the,and,or,a,an,in,on,at,to,for,of,with,by,from,is,are,was,were')

ARGS = {
    'gist_file_path': gist_file_path
}
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
is_number Ran
ran, violated the contract we drafted
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("3b5feddb8481ce67")
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

ARGS = {
    "string": "42.5"
}
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
l2_distance Ran
ran; our contract draft was wrong, not the code
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("00990ef5c064dcc2")
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 1, dtype family float, shape (2,) 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)

# Create two small tensors for L2 distance computation
# batch1 and batch2 should be tensors that can be subtracted
# The function computes distance along the last dimension (-1)
batch1 = torch.randn(2, 4)  # shape: (2, 4)
batch2 = torch.randn(2, 4)  # shape: (2, 4)

ARGS = {
    "batch1": batch1,
    "batch2": batch2,
}
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 Tensor, ndim 1, dtype family float, shape (2,) — shape (2,) 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
metric_recovery Ran
ran, honoured the contract we drafted
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("1cd7320229c7e6cf")
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 2, 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 torch
import numpy as np

torch.manual_seed(0)

# Create small concept tensors for the two datasets
# Shape: (num_concepts, embedding_dim) = (20, 4)
concepts_original = torch.randn(20, 4)
concepts_destination = torch.randn(20, 4)

# Create text_grounding data - each concept has a set of associated texts
# For each concept, store a list of text indices
text_grounding_original = {i: [0, 1, 2] for i in range(20)}
text_grounding_destination = {i: [1, 2, 3] for i in range(20)}

# Create shifted grounding dict with similar structure
shift_grounding_dict = {
    'text_grounding': {i: [2, 3, 4] for i in range(20)}
}

# Create analysis_data dicts with required structure
analysis_data_original = {
    'concepts': concepts_original,
    'text_grounding': text_grounding_original
}

analysis_data_destination = {
    'concepts': concepts_destination,
    'text_grounding': text_grounding_destination
}

ARGS = {
    'num_concepts': 20,
    'metric_key': 'text_grounding',
    'shift_grounding_dict': shift_grounding_dict,
    'analysis_data_original': analysis_data_original,
    'analysis_data_destination': analysis_data_destination
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, scipy, torch — import names as written, not package names to install; a name the repository itself provides is not a package
Returned
type tuple, ndim 2, 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
process_shift_vectors Ran
ran, honoured the contract we drafted
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("8875caa93604d496")
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 2, 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

torch.manual_seed(0)

# per_samples_shifts is a List[torch.Tensor]
# Each tensor represents shift vectors for a concept
# The function processes each tensor with shape (num_shifts, embedding_dim)

# Create a few small tensors with distinct shapes for each concept
concept1_shifts = torch.randn(2, 4)  # 2 shift vectors, 4-dimensional embeddings
concept2_shifts = torch.randn(3, 4)  # 3 shift vectors, 4-dimensional embeddings

per_samples_shifts = [concept1_shifts, concept2_shifts]

ARGS = {
    "per_samples_shifts": per_samples_shifts
}
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, ndim 2, 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
valid_word Ran
ran, violated the contract we drafted
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("3f6b7b65f18a9232")
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
import numpy as np
from typing import List

def is_number(string: str) -> bool:
    try:
        float(string)
        return True
    except ValueError:
        return False

def valid_word(word, eng_corpus: List[str], stopwords: List[str]) -> bool:
    word = word.lower().strip()
    return word in eng_corpus and len(word) > 2 and (word not in stopwords) or is_number(word)

# Setup arguments
torch.manual_seed(0)

# Create a small English corpus
eng_corpus = [
    "hello",
    "world",
    "python",
    "code",
    "testing",
    "function",
    "example"
]

# Create a small stopwords list
stopwords = [
    "the",
    "a",
    "an",
    "and",
    "or",
    "is"
]

# Test word - a valid word in corpus, not in stopwords, longer than 2 chars
word = "python"

ARGS = {
    "word": word,
    "eng_corpus": eng_corpus,
    "stopwords": stopwords
}
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 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
analyse_clusters Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("b16fa7a71c74041f")
How it ran

no invocation recorded.

comp_vis Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("30cf485a5461b7d4")
How it ran

no invocation recorded.

compute_analysis_features Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("b067a9f69b4df21f")
How it ran

no invocation recorded.

compute_shift_concepts Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("99a205cfaa653c1b")
How it ran

no invocation recorded.

compute_shift_grounding_words Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("795bd0c515620e97")
How it ran

no invocation recorded.

concept_text_grounding Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("d7db242219335f6e")
How it ran

no invocation recorded.

decompose_activations Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("41c21b78be088fb8")
How it ran

no invocation recorded.

get_multimodal_grounding Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("e404019518020e3d")
How it ran

no invocation recorded.

save_analysis_to_file Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("9111c2c506ecd059")
How it ran

no invocation recorded.

visualize_concepts Not yet run
unverified (harvested, no recorded run)
mshukor/xl-vlms/src/analysis/cluster_analysis.py
code served (permissive licence) · get_code("28cc2345d5518533")
How it ran

no invocation recorded.

Also present in other papers' code

Of 21 samples lifted from this paper's linked repositories, 21 carry a function identity and 0 do not. Of the 21 distinct functions with an identity, 1 are also present in other papers whose code we hold (1 with at least one paper dated earlier, 0 the same month, 0 later) and 20 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
decompose_activations
brick of this paper's assembly: helper
1 other paper
1 dated earlier
2406.08074 (2024-06)
A Concept-Based Explainability Framework for Large Multimodal Models

Presence, not direction: an identical function body in two papers' repositories does not say who wrote it, which came from which, or that either paper's method rests on it. Matching is exact after normalisation, so a renamed or edited function is not matched. Dates are when papers appeared, not when code was written. Only .py files the harvester lifted are covered; notebooks are not read. A new-style arXiv id YYMM.NNNN dates a paper to the month 20YY-MM; otherwise Paper.year dates it to the year; otherwise it is undated. Dates are when the paper appeared, not when the code was written. Read 2026-11-22.

Repositories linked to this paper

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

Abstract

Multimodal LLMs (MLLMs) have reached remarkable levels of proficiency in understanding multimodal inputs. However, understanding and interpreting the behavior of such complex models is a challenging task, not to mention the dynamic shifts that may occur during fine-tuning, or due to covariate shift between datasets. In this work, we apply concept-level analysis towards MLLM understanding. More specifically, we propose to map hidden states to interpretable visual and textual concepts. This enables us to more efficiently compare certain semantic dynamics, such as the shift from an original and fine-tuned model, revealing concept alteration and potential biases that may occur during fine-tuning. We also demonstrate the use of shift vectors to capture these concepts changes. These shift vectors allow us to recover fine-tuned concepts by applying simple, computationally inexpensive additive concept shifts in the original model. Finally, our findings also have direct applications for MLLM steering, which can be used for model debiasing as well as enforcing safety in MLLM output. All in all, we propose a novel, training-free, ready-to-use framework for MLLM behavior interpretability and control. Our implementation is publicly available.

For agents

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

get_harvested_code_for_paper("2501.03012")
get_code_for_paper("2501.03012")
have("2501.03012")

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

curl https://syntology.ai/api/ran/2501.03012.json

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

[![Syntology run record](https://syntology.ai/api/ran/2501.03012.svg)](https://syntology.ai/paper/2501.03012)

Connect an agent — have() is free.