SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2602.06883 · ICML · 2026

Vision Transformer Finetuning Benefits from Non-Smooth Components

Ievgen Redko, Laetitia Chapel, Ambroise Odonnat, Romain Tavenard

arXiv · PDF · Open in the Atlas

Code that ran

8 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, 0 violated it and 0 have no contract verdict; 6 failures of our instrument; 2 no recorded run), measured 2026-11-22; 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/2602.06883.json · badge: /api/ran/2602.06883.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
ambroiseodt/vit-plasticity canonical (classifier label)
classifier basis: census
8 ran · 2 no recorded run
FunctionStatusWhere it lives
build_with_type_check Ran
ran; our contract draft was wrong, not the code
ambroiseodt/vit-plasticity/src/vitef/utils.py
code served (permissive licence) · get_code("c9981055b9019584")
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 SimpleConfig. 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 copy
import logging
from dataclasses import dataclass, fields
from typing import Any, Literal, Union

logging.basicConfig(level=logging.DEBUG)

# Define some test dataclasses
@dataclass
class SimpleConfig:
    name: str
    value: int

@dataclass
class NestedConfig:
    simple: SimpleConfig
    flag: bool

@dataclass
class ListConfig:
    items: list[str]

@dataclass
class DictConfig:
    mapping: dict[str, int]

@dataclass
class UnionConfig:
    data: Union[str, int]

@dataclass
class LiteralConfig:
    mode: Literal["train", "eval", "test"]

@dataclass
class ComplexConfig:
    name: str
    nested: NestedConfig
    items: list[int]
    settings: dict[str, str]

# Test case 1: Simple dataclass
test_data_simple = {
    "name": "test",
    "value": 42
}

# Test case 2: Nested dataclass
test_data_nested = {
    "simple": {
        "name": "inner",
        "value": 10
    },
    "flag": True
}

# Test case 3: List of strings
test_data_list = ["apple", "banana", "cherry"]

# Test case 4: Dict with string keys and int values
test_data_dict = {
    "alpha": 1,
    "beta": 2,
    "gamma": 3
}

# Test case 5: Union type - use string variant
test_data_union_str = "hello"

# Test case 6: Union type - use int variant
test_data_union_int = 99

# Test case 7: Literal type
test_data_literal = "train"

# Test case 8: Complex nested structure
test_data_complex = {
    "name": "complex",
    "nested": {
        "simple": {
            "name": "deep",
            "value": 5
        },
        "flag": False
    },
    "items": [1, 2, 3, 4],
    "settings": {
        "key1": "val1",
        "key2": "val2"
    }
}

ARGS = {
    "object_type": SimpleConfig,
    "data": test_data_simple,
    "inplace": True
}
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 SimpleConfig — a SimpleConfig output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a SimpleConfig 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_jsonl_keys Ran
ran; our contract draft was wrong, not the code
ambroiseodt/vit-plasticity/src/vitef/utils.py
code served (permissive licence) · get_code("2b47253f170c2a89")
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 json
import tempfile
import os

# Create a temporary JSONL file for testing
temp_dir = tempfile.mkdtemp()
jsonl_path = os.path.join(temp_dir, "test.jsonl")

# Write some test data to the JSONL file
test_data = [
    {"name": "Alice", "age": 30, "city": "New York"},
    {"name": "Bob", "email": "bob@example.com", "country": "USA"},
    {"name": "Charlie", "age": 25, "email": "charlie@example.com"},
]

with open(jsonl_path, "w") as f:
    for item in test_data:
        f.write(json.dumps(item) + "\n")

ARGS = {
    "path": jsonl_path,
    "readall": True,
}
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
get_pca_basis Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
ambroiseodt/vit-plasticity/apps/plots/loss_landscape.py
code served (permissive licence) · get_code("d151a2ea39059f9c")
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
import torch.nn as nn
import torch.nn.functional as F
import numpy as np

torch.manual_seed(0)

# Define a simple model that takes inputs and produces outputs for cross_entropy loss
class SimpleModel(nn.Module):
    def __init__(self, input_dim, num_classes):
        super().__init__()
        self.linear1 = nn.Linear(input_dim, 8)
        self.linear2 = nn.Linear(8, num_classes)
    
    def forward(self, x):
        x = F.relu(self.linear1(x))
        x = self.linear2(x)
        return x

# Create model
model = SimpleModel(input_dim=4, num_classes=3)

# Create small input and target tensors
# batch_size=2, input_dim=4, num_classes=3
inputs = torch.randn(2, 4)
targets = torch.randint(0, 3, (2,))

# Get the parameters to optimize
params = list(model.parameters())

# Define ARGS dictionary
ARGS = {
    'params': params,
    'model': model,
    'inputs': inputs,
    'targets': targets,
    'n_steps': 3,
    'lr': 0.01,
    'device': 'cpu'
}
Setup block drafted by
a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API
Imports
numpy, sklearn, 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
interpolate_frames Ran
ran; our contract draft was wrong, not the code
ambroiseodt/vit-plasticity/apps/plots/loss_landscape.py
code served (permissive licence) · get_code("e0deed7ee32dc0c5")
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 4, 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 numpy as np

# Create small test frames as numpy arrays
# Each frame is a small RGB image (4x3x3 means 4 height, 3 width, 3 channels)
frame1 = np.random.randint(0, 256, size=(4, 3, 3), dtype=np.uint8)
frame2 = np.random.randint(0, 256, size=(4, 3, 3), dtype=np.uint8)
frame3 = np.random.randint(0, 256, size=(4, 3, 3), dtype=np.uint8)

frames = [frame1, frame2, frame3]

ARGS = {
    "frames": frames,
    "n_steps": 5
}
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 list, ndim 4, 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
load_jsonl_to_numpy Ran
ran; our contract draft was wrong, not the code
ambroiseodt/vit-plasticity/src/vitef/utils.py
code served (permissive licence) · get_code("d854539c5e14110e")
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 json
import tempfile
import os
import numpy as np

# Create a temporary JSONL file with sample data
temp_dir = tempfile.mkdtemp()
jsonl_path = os.path.join(temp_dir, "test_data.jsonl")

# Write sample JSONL data
with open(jsonl_path, 'w') as f:
    json.dump({"name": "Alice", "age": 30, "score": 95.5}, f)
    f.write('\n')
    json.dump({"name": "Bob", "age": 25, "score": 87.3}, f)
    f.write('\n')
    json.dump({"name": "Charlie", "age": 35, "score": 92.1}, f)
    f.write('\n')

ARGS = {
    "path": jsonl_path,
    "keys": ["name", "age", "score"]
}
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
lr_constant Ran
ran, honoured the contract we drafted
ambroiseodt/vit-plasticity/src/vitef/optim.py
code served (permissive licence) · get_code("4c20b73b071752f2")
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

ARGS = {
    "step": 0
}
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 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
lr_linear Ran
ran, honoured the contract we drafted
ambroiseodt/vit-plasticity/src/vitef/optim.py
code served (permissive licence) · get_code("712ef77b64ba510b")
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

# Define ARGS dict with all required parameters for lr_linear
ARGS = {
    "step": 500,
    "warmup": 100,
    "min_factor": 0.1,
    "n_steps": 1000,
}
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
set_seed Ran
ran; our contract draft was wrong, not the code
ambroiseodt/vit-plasticity/src/vitef/config.py
code served (permissive licence) · get_code("6a3bc5a7f790da13")
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 Generator. 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 numpy as np
import torch

torch.manual_seed(0)

ARGS = {
    "seed": 42,
}
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 Generator — a Generator output has no array shape of its own; none is stated — the instrument coerces any output to an array; for a Generator 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
build_optimizer Not yet run
unverified (harvested, no recorded run)
ambroiseodt/vit-plasticity/src/vitef/optim.py
code served (permissive licence) · get_code("0c8eb7277387806b")
How it ran

no invocation recorded.

get_raw_model Not yet run
unverified (harvested, no recorded run)
ambroiseodt/vit-plasticity/src/vitef/distributed.py
code served (permissive licence) · get_code("c4ee7ffac8194eb3")
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, 0 are also present in other papers whose code we hold (0 with at least one paper dated earlier, 0 the same month, 0 later) and 10 are only here among the 52,977 papers whose code Syntology holds (as of 2026-09-26).

Only here means: among the 52,977 papers whose code Syntology holds (51,887 with lifted samples, 1,090 through their assemblies only), as of 2026-09-26, no other paper's lifted samples or current assembly bricks contain an identical function body after normalisation.

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

Repositories linked to this paper

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

Abstract

The smoothness of the transformer architecture has been extensively studied in the context of generalization, training stability, and adversarial robustness. However, its role in transfer learning remains poorly understood. In this paper, we analyze the ability of vision transformer components to adapt their outputs to changes in inputs, or, in other words, their plasticity. Defined as an average rate of change, it captures the sensitivity to input perturbation; in particular, a high plasticity implies a low smoothness. Our theoretical analysis and extensive experiments -over 1, 000 finetuning runs on large-scale vision transformers -showcase that this perspective provides principled guidance in choosing the components to prioritize during adaptation. A key takeaway for practitioners is that the high plasticity of the attention modules and feedforward layers consistently leads to better finetuning performance. Our findings depart from the prevailing assumption that smoothness is desirable, offering a novel perspective on transformers' functional properties.

For agents

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

get_harvested_code_for_paper("2602.06883")
get_code_for_paper("2602.06883")
have("2602.06883")

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

curl https://syntology.ai/api/ran/2602.06883.json

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

[![Syntology run record](https://syntology.ai/api/ran/2602.06883.svg)](https://syntology.ai/paper/2602.06883)

Connect an agent — have() is free.