SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2503.14564 · CVPR · 2025

Effortless Active Labeling for Long-Term Test-Time Adaptation

Changxing Ding, Guowei Wang

arXiv · PDF · Open in the Atlas

Code that ran

10 of 13 samples (functions and classes) lifted from this paper's linked repositories executed on a synthesized input in Syntology's jail (6 of them constructed an object rather than computing a result; 0 honoured the contract we drafted, 0 violated it and 8 have no contract verdict; 2 failures of our instrument; 3 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/2503.14564.json · badge: /api/ran/2503.14564.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
flash1803/EATTA community
Papers with Code archive
10 ran · 3 no recorded run
FunctionStatusWhere it lives
Buffer Ran
ran on a synthesized input · constructed an object; no result was computed
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("c805e78993db72b6")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, Buffer: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import numpy as np

torch.manual_seed(0)

# Create small sample data
# Each item is a (sample, label) tuple
sample1 = torch.randn(4)
label1 = 0
sample2 = torch.randn(4)
label2 = 1

data = [
    (sample1, label1),
    (sample2, label2),
]

ARGS = {
    "data": data
}
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 Buffer — the run constructed an instance of its own class and did not call it; no shape 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
Entropy Ran
ran on a synthesized input
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("9e77fe01c3502ba4")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) and returned Tensor, 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 torch.nn as nn

class Entropy(nn.Module):

    def __init__(self):
        super(Entropy, self).__init__()

    def __call__(self, logits):
        return -(logits.softmax(1) * logits.log_softmax(1)).sum(1)


INIT_ARGS = {}

ARGS = {
    'logits': torch.randn(2, 4)
}

if __name__ == '__main__':
    entropy = Entropy(**INIT_ARGS)
    result = entropy(**ARGS)
    print(result)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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 — no output shape is stated: observe_signature.py observed this class with its constructor sizes canonicalised where it could, which the node does not record, so its shape is not stated as this setup block's
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
ImageNetXMaskingLayer Ran
ran on a synthesized input · constructed an object; no result was computed
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("9032562b014e775f")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ImageNetXMaskingLayer: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# The mask parameter is used in forward() as: x[:, self.mask]
# This means mask should be a valid index/slice for the second dimension of a tensor
# mask should be a tensor or list of indices to select columns from x

# Create a mask that selects certain columns from the second dimension
# If x has shape (batch, num_features), mask should index into num_features
mask = torch.tensor([0, 2])  # Select columns 0 and 2

ARGS = {
    "mask": mask
}
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 ImageNetXMaskingLayer — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ImageNetXMaskingLayer 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
ImageNetXWrapper Ran
ran on a synthesized input · constructed an object; no result was computed
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("2e59dd01c8d5a4af")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, ImageNetXWrapper: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create a minimal model that has normalize method and forward that returns logits
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(4, 10)
    
    def forward(self, x):
        return self.linear(x)

class ModelWithNormalize(nn.Module):
    def __init__(self):
        super().__init__()
        self.model = MinimalModel()
    
    def normalize(self, x):
        # Simple normalization: subtract mean and divide by std
        return (x - x.mean()) / (x.std() + 1e-5)
    
    def forward(self, x):
        return self.model(x)

# Create the model instance
model = ModelWithNormalize()

# Create a mask for selecting subset of output classes
# Model outputs 10 logits, mask selects 6 of them
mask = torch.tensor([0, 1, 2, 4, 6, 8])

ARGS = {
    "model": model,
    "mask": mask,
}
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 ImageNetXWrapper — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a ImageNetXWrapper 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
Registry Ran
ran on a synthesized input · constructed an object; no result was computed
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("9dcddb1632fb93b5")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, Registry: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

# Create a minimal Registry instance for testing
registry = nn.Module()

# The Registry.__init__ only takes a 'name' parameter
ARGS = {
    "name": "TEST_REGISTRY"
}
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 Registry — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a Registry 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
SampleBuffer Ran
ran on a synthesized input · constructed an object; no result was computed
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("a1931dd46fa423bd")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, SampleBuffer: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import random

torch.manual_seed(0)

# SampleBuffer only takes one parameter: capacity
# The add() and sample() methods work with samples and labels passed as arguments
# samples and labels are expected to be iterables that can be zipped together

ARGS = {
    "capacity": 300
}
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 SampleBuffer — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a SampleBuffer 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
TransformerWrapper Ran
ran on a synthesized input · constructed an object; no result was computed
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("701b69ab32deabd2")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, Claude Haiku 4.5, prompt python_fixture_v1, through the batch API) and returned an instance of its own class, TransformerWrapper: the run constructed the object and did not call it, so no result was computed. The image it ran in was not recorded. If your copy returns the same type on these inputs, it matched what we observed; nothing here checks that. It does NOT mean the output is correct: nothing has compared it to the paper.

import torch
import torch.nn as nn

torch.manual_seed(0)

# Create a minimal mock model that satisfies the forward method's requirements
class MockTransformerModel(nn.Module):
    def __init__(self):
        super().__init__()
        # class_token: shape [1, 1, hidden_dim]
        self.class_token = nn.Parameter(torch.randn(1, 1, 4))
        
        # encoder: processes concatenated tokens and returns same shape
        self.encoder = nn.Sequential(
            nn.Linear(4, 4),
            nn.ReLU()
        )
    
    def _process_input(self, x):
        # Takes input and returns [batch, seq_len, hidden_dim]
        # Input shape: [batch, feature_dim]
        # Output shape: [batch, num_patches, hidden_dim]
        batch_size = x.shape[0]
        # Expand to [batch, seq_len, hidden_dim]
        return x.unsqueeze(1)  # [batch, 1, hidden_dim]


# Create a minimal normalize function
class MockNormalizer:
    def __call__(self, x):
        return x / (x.norm(dim=-1, keepdim=True) + 1e-6)


# Create the mock model instance
mock_model = MockTransformerModel()

# Create a wrapper-like object with normalize attribute
class ModelWithNormalize:
    def __init__(self, model):
        self.model = model
        self.normalize = MockNormalizer()
        self.class_token = model.class_token
        self.encoder = model.encoder
        self._process_input = model._process_input


model_obj = ModelWithNormalize(mock_model)

# Create input tensor: [batch_size=2, feature_dim=4]
x = torch.randn(2, 4)

ARGS = {
    "model": model_obj
}
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 TransformerWrapper — the run constructed an instance of its own class and did not call it; no shape is stated — the instrument coerces any output to an array; for a TransformerWrapper 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
select_sample Ran
ran on a synthesized input
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("cb96e12d0f3ba177")
How it ran

Ran in Syntology's sandbox on this synthesized input (drafted by a language model, class construct-then-call prompt, model not recorded) 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

class select_sample(nn.Module):

    def __init__(self, oracle_num, device):
        self.last_cls = []
        self.oracle_num = oracle_num
        self.device = device

    def __call__(self, y_prime, sorted_indices, c_diff, cls_num_count, cls_diff):
        self.mask = torch.zeros(sorted_indices.size(0), dtype=torch.bool).to(self.device)
        if len(self.last_cls) == 0:
            can = y_prime[sorted_indices][:self.oracle_num]
            diff = c_diff[sorted_indices][:self.oracle_num]
            for idx, i in enumerate(can):
                cls_num_count[i] += 1
                cls_diff[i] = diff[idx]
                self.mask[idx] = True
            sorted_idx = sorted_indices[self.mask]
            for i in y_prime[sorted_idx]:
                self.last_cls.append(i.item())
        else:
            can = y_prime[sorted_indices][:self.oracle_num]
            diff = c_diff[sorted_indices][:self.oracle_num]
            for idx, i in enumerate(can):
                if i.item() not in self.last_cls:
                    cls_num_count[i.item()] += 1
                    cls_diff[i.item()] = diff[idx]
                    self.mask[idx] = True
                elif diff[idx] < cls_diff[i]:
                    can2 = y_prime[sorted_indices][self.oracle_num:]
                    diff2 = c_diff[sorted_indices][self.oracle_num:]
                    for idx_, j in enumerate(can2):
                        if j not in self.last_cls:
                            cls_num_count[j] += 1
                            cls_diff[j] = diff2[idx_]
                            self.mask[self.oracle_num + idx_] = True
                            break
                        else:
                            continue
                else:
                    cls_num_count[i] += 1
                    cls_diff[i] = diff[idx]
                    self.mask[idx] = True
            sorted_idx = sorted_indices[self.mask]
            if len(sorted_idx) >= 1:
                for i in y_prime[sorted_idx]:
                    self.last_cls.append(i.item())
            else:
                sorted_idx = sorted_indices[:self.oracle_num]
            if len(self.last_cls) > 2:
                self.last_cls = self.last_cls[len(self.last_cls) - 1:]
        return (sorted_idx, cls_num_count, cls_diff)

INIT_ARGS = {
    'oracle_num': 2,
    'device': 'cpu'
}

ARGS = {
    'y_prime': torch.tensor([0, 1, 0, 1, 2], dtype=torch.long),
    'sorted_indices': torch.tensor([0, 1, 2, 3, 4], dtype=torch.long),
    'c_diff': torch.tensor([0.5, 0.3, 0.8, 0.2, 0.6], dtype=torch.float32),
    'cls_num_count': torch.tensor([0, 0, 0], dtype=torch.long),
    'cls_diff': torch.tensor([0.0, 0.0, 0.0], dtype=torch.float32)
}

if __name__ == '__main__':
    model = select_sample(**INIT_ARGS)
    result = model(**ARGS)
    print("Result:", result)
Setup block drafted by
a language model, class construct-then-call prompt, model not recorded
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: torch and numpy seeded to 0 by the harness immediately before the class was constructed from INIT_ARGS (fixture_check.py)
split_up_model Ran
ran; our contract draft was wrong, not the code
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("a1f37fac1b9d0c32")
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
from copy import deepcopy

torch.manual_seed(0)

# Create a minimal model that satisfies the resnet case
class MinimalModel(nn.Module):
    def __init__(self):
        super().__init__()
        self.normalize = ImageNormalizer(
            mean=(0.485, 0.456, 0.406),
            std=(0.229, 0.224, 0.225)
        )
        self.model = nn.Module()
        self.model.fc = nn.Linear(64, 10)
        
        # Build a simple sequential model to act as backbone
        self.model.layer1 = nn.Conv2d(3, 16, kernel_size=3, padding=1)
        self.model.layer2 = nn.Conv2d(16, 32, kernel_size=3, padding=1)
        self.model.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        
    def forward(self, x):
        x = self.normalize(x)
        x = self.model.layer1(x)
        x = self.model.layer2(x)
        x = self.model.avgpool(x)
        x = x.flatten(1)
        x = self.model.fc(x)
        return x

class ImageNormalizer(nn.Module):
    def __init__(self, mean, std):
        super().__init__()
        self.register_buffer('mean', torch.as_tensor(mean).view(1, 3, 1, 1))
        self.register_buffer('std', torch.as_tensor(std).view(1, 3, 1, 1))

    def forward(self, input):
        return (input - self.mean) / self.std

# Build the model
model = MinimalModel()

# Create model with children structure that the function expects
def build_children_structure():
    m = nn.Module()
    m.normalize = ImageNormalizer(
        mean=(0.485, 0.456, 0.406),
        std=(0.229, 0.224, 0.225)
    )
    m.model = nn.Module()
    m.model.layer1 = nn.Conv2d(3, 16, kernel_size=3, padding=1)
    m.model.layer2 = nn.Conv2d(16, 32, kernel_size=3, padding=1)
    m.model.avgpool = nn.AdaptiveAvgPool2d((1, 1))
    m.model.fc = nn.Linear(32, 10)
    return m

model_for_split = build_children_structure()

ARGS = {
    'model': model_for_split,
    'arch_name': 'resnet50',
    'dataset_name': 'cifar10'
}
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
update_w1_w2 Ran
ran; the contract check could not run (RAISES, DEP_MISSING or TIMEOUT)
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("4b36ed781ee058a7")
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)

# Create small tensors for w1 and w2
w1 = torch.randn(2, 4)
w2 = torch.randn(3, 4)

# Create EMA tensors - initialized to 0 as integers (will be checked with == 0)
w1_ema = 0
w2_ema = 0

# Set momentum parameter
momentum = 0.8

ARGS = {
    "w1": w1,
    "w2": w2,
    "w1_ema": w1_ema,
    "w2_ema": w2_ema,
    "momentum": momentum,
}
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
EATTA Not yet run
unverified (harvested, no recorded run)
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("27e1b5f5542240da")
How it ran

no invocation recorded.

ResNetDomainNet126 Not yet run
unverified (harvested, no recorded run)
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("1a7ed187fe60d4cc")
How it ran

no invocation recorded.

TTAMethod Not yet run
unverified (harvested, no recorded run)
flash1803/EATTA/classification/methods/eatta.py
code served (permissive licence) · get_code("ae22d48dab7f8e79")
How it ran

no invocation recorded.

Also present in other papers' code

Of 13 samples lifted from this paper's linked repositories, 13 carry a function identity and 0 do not. Of the 13 distinct functions with an identity, 5 are also present in other papers whose code we hold (2 with at least one paper dated earlier, 0 the same month, 3 later) and 8 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
Entropy
brick of this paper's assembly: helper
2 other papers
2 later · 2 only as an assembly brick
2505.16441 (2025-05)
Ranked Entropy Minimization for Continual Test-Time Adaptation
2603.03796 (2026-03)
When and Where to Reset Matters for Long-Term Test-Time Adaptation
ImageNetXMaskingLayer
brick of this paper's assembly: helper
1 other paper
1 dated earlier · 1 only as an assembly brick
2311.18193 (2023-11)
Persistent Test-time Adaptation in Recurring Testing Scenarios
Registry
brick of this paper's assembly: helper
1 other paper
1 later · 1 only as an assembly brick
2603.03796 (2026-03)
When and Where to Reset Matters for Long-Term Test-Time Adaptation
ResNetDomainNet126
brick of this paper's assembly: helper
1 other paper
1 later · 1 only as an assembly brick
2603.03796 (2026-03)
When and Where to Reset Matters for Long-Term Test-Time Adaptation
TransformerWrapper
brick of this paper's assembly: helper
1 other paper
1 dated earlier · 1 only as an assembly brick
2311.18193 (2023-11)
Persistent Test-time Adaptation in Recurring Testing Scenarios

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

Long-term test-time adaptation (TTA) is a challenging task due to error accumulation. Recent approaches tackle this issue by actively labeling a small proportion of samples in each batch, yet the annotation burden quickly grows as the batch number increases. In this paper, we investigate how to achieve effortless active labeling so that a maximum of one sample is selected for annotation in each batch. First, we annotate the most valuable sample in each batch based on the single-step optimization perspective in the TTA context. In this scenario, the samples that border between the source-and target-domain data distributions are considered the most feasible for the model to learn in one iteration. Then, we introduce an efficient strategy to identify these samples using feature perturbation. Second, we discover that the gradient magnitudes produced by the annotated and unannotated samples have significant variations. Therefore, we propose balancing their impact on model optimization using two dynamic weights. Extensive experiments on the popular ImageNet-C, -R, -K, -A and PACS databases demonstrate that our approach consistently outperforms state-of-the-art methods with significantly lower annotation costs. Code is available at: https://github.com/flash1803/EATTA.

For agents

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

get_harvested_code_for_paper("2503.14564")
get_code_for_paper("2503.14564")
have("2503.14564")

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

curl https://syntology.ai/api/ran/2503.14564.json

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

[![Syntology run record](https://syntology.ai/api/ran/2503.14564.svg)](https://syntology.ai/paper/2503.14564)

Connect an agent — have() is free.